Daily Digest — 2026-06-10
345 items · 5 research labs, 338 arxiv papers, 2 industry media
MarkTechPost: all feed URLs failed (last tried: https://www.marktechpost.com/feed/)AI News: all feed URLs failed (last tried: https://artificialintelligence-news.com/feed/)
🏛️ Research Labs (5)
How engineers at Nextdoor use Codex to build without limits
OpenAI Codex enables Nextdoor engineers to shift from iterative prompting to outcome engineering, accelerating development by allowing single engineers to own end-to-end product experiences. The method involves using Codex for debugging, feature development, and strategic planning across embedded Rust databases and Kubernetes environments. Results include 3x faster feature deployment (e.g., Opportunity Alerts with map integration) and organizational bottlenecks shifting from engineering execution to product strategy decisions, with GPT-5.5 upgrades improving persistence in root-cause analysis.
outcome engineeringkubernetes podsembedded rustroot-cause analysisfeedback loop
Confidential submission of draft S-1 to the SEC
OpenAI has submitted a confidential draft S-1 registration statement to the SEC, signaling potential intent to pursue an initial public offering (IPO). The submission, made under Rule 135 of the Securities Act of 1933, does not constitute an immediate offer to sell securities but provides flexibility in timing the IPO. OpenAI acknowledges the complexity of transitioning from private to public status, emphasizing that certain operational goals may be more achievable as a private entity. The announcement preemptively addresses potential leaks of the confidential filing.
confidential s-1secrule 135securities actipo
Can Voice Agents Handle Bilingual Customers? Benchmarking Frontier ASR on Code-Switched Speech
The study introduces a novel benchmark for evaluating automatic speech recognition (ASR) systems on code-switched speech in enterprise settings, focusing on Spanish-English, French-English, Canadian French-English, and German-English pairs. Using a dataset of 918 HR and IT service management utterances synthesized via GPT-5 and ElevenLabs Multilingual V2, the authors assess seven ASR models (including ElevenLabs Scribe V2, Gemini 3 Flash, and AssemblyAI Universal 3-Pro) across Word Error Rate (WER), Semantic WER (SWER), and Answer Error Rate (AER). Key findings show top models incur minimal performance degradation from code-switching (ΔWER < 0.15), while logistic regression reveals utterance-level switches and Code-Mixing Index (CMI) as primary error predictors.
automatic speech recognitioncode-switchingword error ratesemantic word error rateanswer error rate
Introducing North Mini Code: Cohere’s First Model For Developers
Cohere introduces North Mini Code, a 30B-parameter Mixture-of-Experts model with 3B active parameters, optimized for agentic coding tasks. The model employs a decoder-only Transformer architecture with interleaved sliding-window and global attention, trained via two-stage supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR). It achieves 33.4 on Artificial Analysis’ Coding Index, outperforming comparable models, and demonstrates robustness across diverse coding harnesses, with 80.2% pass@10 on SWE-Bench Verified and 55.1% pass@10 on Terminal-Bench v2.
mixture-of-expertsagentic codingsupervised fine-tuningreinforcement learningtransformer decoder
How an Agent Built a 3D Paris Gallery by Chaining Two Hugging Face Spaces
A coding agent autonomously constructed a 3D Paris gallery by chaining two Hugging Face Spaces: ideogram-ai/ideogram4 for image generation and VAST-AI/TripoSplat for 3D Gaussian splat reconstruction. The agent leveraged agents.md files, which provide plain-text API schemas for Spaces, enabling seamless integration without client libraries. The pipeline involved generating isolated monument images, reconstructing 3D splats, compressing files, and deploying a Three.js viewer. The marginal cost of creating new galleries (e.g., Japan, Egypt) approached the cost of describing them, demonstrating the building-block economy. This approach highlights the potential for agents to assemble multimedia applications from modular, documented components.
gaussian splatagents.mdgradio apithree.jsbuilding-block economy
📜 arXiv Papers (338)
OmniGameArena: A Unified UE5 Benchmark for VLM Game Agents with Improvement Dynamics
OmniGameArena introduces a unified Unreal Engine 5 benchmark for evaluating vision-language model (VLM) agents across twelve games in Solo, PvP, and Coop modes, addressing limitations of existing benchmarks. The method includes a standardized action interface for heterogeneous agent classes (commercial VLMs, open-weight VLMs, specialized policies) and proposes Improvement Dynamics Curve (IDC), a reflection-based protocol where a tool-using LLM autonomously refines skill prompts across multiple rounds. Results show cold-start leaderboard scores for twelve VLMs and IDC performance metrics (score evolution, generalization) for four top agents, capturing both initial capability and learning dynamics.
vision-language modelsunreal engine 5improvement dynamics curveheterogeneous agentsgame benchmarks
An Agency-Transferring Model-Free Policy Enhancement Technique
The paper introduces an agency-transferring technique for enhancing reinforcement learning (RL) policies by embedding a functional baseline policy into the training process. The method arbitrates between the baseline and a trainable learning policy, gradually transferring control from the baseline to the learning policy, which ultimately operates independently. Theoretical analysis formalizes the baseline's functional property and provides lower bounds for the learning policy's goal-reaching probability. Empirical evaluations on continuous-control benchmarks demonstrate that the method achieves competitive returns while maintaining higher goal-reaching rates throughout training compared to alternative approaches.
reinforcement learningbaseline policycontinuous-controlgoal-reaching probabilityagency-transferring
PTL-Diffusion: Manifold-Aware Diffusion with Periodic Terminal Laws
PTL-Diffusion introduces a diffusion framework with periodic terminal laws to improve manifold-aware generation, replacing the standard single Gaussian terminal distribution. The method employs a periodically forced Ornstein-Uhlenbeck forward process, deriving closed-form marginals and reverse posteriors while maintaining compatibility with noise-prediction training. Experiments on torus/cylinder point-clouds and Olivetti faces demonstrate improved manifold matching over DDPM baselines, measured by phase-conditioned errors, covariance discrepancies, and nearest-neighbor distances.
diffusion modelsperiodic terminal lawsmanifold-aware generationornstein-uhlenbeck processphase-conditioned denoising
AHA-WAM:Asynchronous Horizon-Adaptive World-Action Modeling with Observation-Guided Context Routing
The paper introduces AHA-WAM, an Asynchronous Horizon-Adaptive World-Action Model for robot manipulation that decouples world prediction and action execution at different temporal resolutions. The method employs a dual Diffusion Transformer (DiT) architecture: a low-frequency video DiT for long-horizon world planning with rolling key-value memory, and a high-frequency action DiT for short action execution via layerwise joint attention. Key innovations include horizon-adaptive offset training and Observation-Guided Video-Context Routing (OVCR) to maintain real-time responsiveness. Experiments on RoboTwin and real-world tasks show state-of-the-art performance (92.80% success on RoboTwin, 78.3% on real-world tasks) with 24.17 Hz closed-loop control and 4.59x speedup over Fast-WAM.
world-action modeldiffusion transformerhorizon-adaptivekey-value memoryclosed-loop control
Evaluation Cards: An Interpretive Layer for AI Evaluation Reporting
The authors introduce EvalCards, an interpretive reporting layer for AI evaluation that unifies benchmark metadata, evaluation run data, and model metadata into a single record. The method involves deriving a reporting schema from 52 papers and 10 stakeholder interviews, implementing four interpretive signals (reproducibility, documentation completeness, provenance and risk, and score comparability), and deploying a monitoring tool across 5,816 models, 635 benchmarks, and 101,843 results. Results reveal systematic gaps in current reporting practices, demonstrating the tool's utility for both research and non-research audiences.
evalcardsbenchmark metadatainterpretive signalsreporting schemaevaluation lifecycle
Topological Neural Operators
The authors introduce Topological Neural Operators (TNOs), a framework extending neural operators to cell complexes by modeling data as features on cells of varying dimensions. TNOs employ Discrete Exterior Calculus for cross-dimensional coupling via gradient-, curl-, and divergence-type operators, decoupling information flow from learned transformations. Hierarchical TNOs (HTNOs) incorporate learned coarse complexes for long-range propagation. The framework generalizes existing neural operators and demonstrates improved accuracy on PDE benchmarks, including irregular-geometry flows, while isolating benefits of higher-rank and topological structure.
topological neural operatorscell complexesdiscrete exterior calculusoperator learningpde benchmarks
Bandits for Efficient Experimentation: Adapting to Control Group, Preferences, and Context Drifts
The paper introduces Dri-MED, a linear contextual bandit algorithm for personalized recommendations under non-stationary context distributions and heteroskedastic noise. The method extends the MED strategy to handle drifting contexts while ensuring each decision's mean reward exceeds a baseline policy π₀. Theoretical analysis shows Õ(κ/Δ̃ d² log(T)) instance-dependent regret with variance-aware term κ and Õ(d) constraint violations. Empirical results demonstrate superiority over drift-agnostic baselines.
contextual banditsnon-stationary environmentsheteroskedastic regressionconstraint-aware learningpersonalized recommendations
FASE: Fast Adaptive Semantic Entropy for Code Quality
We introduce Fast Adaptive Semantic Entropy (FASE), a novel metric for quantifying uncertainty in multi-agent code generation systems without ground-truth answers. FASE approximates functional correctness by constructing minimum spanning trees from structural and semantic dissimilarity graphs, eliminating costly LLM-driven equivalence checks. Evaluations on HumanEval and BigCodeBench show FASE outperforms state-of-the-art semantic entropy methods, achieving 25% higher Spearman correlation and 19% improved ROCAUC against Pass@1 metrics. FASE reduces computational overhead to 0.3% of traditional semantic entropy approaches, enabling practical uncertainty quantification in real-world workflows.
semantic entropyminimum spanning treemulti-agent code generationfunctional correctnessdissimilarity graphs
Who Earns the Safety? Intervention-Aware Quantum Predictive Control with Safety Attribution
The paper introduces Intervention-Aware Variational Quantum Differentiable Predictive Control (IA-VQC-DPC), a method that trains variational quantum circuit policies under a primal-dual intervention budget to reduce reliance on safety filters. The approach includes a safety-attribution protocol to decompose trajectory corrections into Control-Barrier-Function (CBF) and runtime-guard terms, enabling policy-level safety evaluation. Empirical results on building-control emulators (5 seeds, 60 episodes) show the quantum policy (400 parameters) significantly reduces pre-filter violations and safety-layer reliance (p < 10^-4) without energy regression, outperforming classical counterparts. Guard-off evaluation reveals learned energy heads require distribution-aware runtime guards for safety.
intervention-awarevariational quantum circuitcontrol-barrier-functiondifferentiable predictive controlsafety-attribution
SIGA: Self-Evolving Coding-Agent Adapters for Scientific Simulation
SIGA introduces a Simulator-Interface Grounding Adapter to enable off-the-shelf coding agents to operate scientific simulators by supplying vocabulary, structural constraints, validation rules, and termination conditions. The method employs retrieval, procedural memory, in-trajectory validation, and validation-enforced termination. Evaluated on GEOS, SIGA achieves a TreeSim score above 0.90 in five minutes, matching human experts in three hours, and improves held-out set performance from 0.720 to 0.789. Self-evolution further enhances SIGA by rewriting adapter contents from prior trajectories. Transfers to OpenFOAM and LAMMPS reveal that validation, memory, and retrieval mechanisms dominate based on interface characteristics.
simulator-interface groundingprocedural memoryin-trajectory validationself-evolutiontreesim
Data Synthesis and Parameter-Efficient Fine-Tuning for Low-Resource NMT: A Case Study on Q'eqchi' Mayan
This study introduces a data synthesis methodology for neural machine translation (NMT) in low-resource Indigenous languages, focusing on Q'eqchi' Mayan. Community-sourced dictionaries were transformed into a synthetic corpus, and Parameter-Efficient Fine-Tuning (PEFT) was applied using LoRA adapters on an mT5-base model. In-domain evaluation achieved a BLEU score of 42.02, demonstrating effective acquisition of complex agglutinative morphology and VOS word order. However, evaluation against an organic glossary revealed a structural-semantic gap (BLEU 0.59), indicating overfitting to synthetic templates and limited syntactic fluidity. An ablation study using Multi-Task Learning resulted in negative transfer, suggesting competition for parameter capacity in LoRA adapters. Synthetic bootstrapping proved effective as a structural primer but requires authentic data for semantic refinement via Curriculum Learning.
neural machine translationparameter-efficient fine-tuninglora adapterssynthetic corpuscurriculum learning
Preserving Plasticity in Continual Learning via Dynamical Isometry
The paper introduces dynamical isometry as a mechanism to preserve plasticity in continual learning, linking it to the Neural Tangent Kernel. It proposes AdamO, an Adam-style optimizer that decouples isometry regularization from gradient updates, and demonstrates its efficacy in reactivating dormant ReLU units. The method is validated across supervised and reinforcement-learning benchmarks, showing superior performance in mitigating plasticity loss compared to existing approaches.
dynamical isometryneural tangent kernelcontinual learningplasticity preservationadam optimizer
Difference-Aware Retrieval Policies for Imitation Learning
The paper introduces Difference-Aware Retrieval Policies for Imitation Learning (DARP), a semi-parametric retrieval-based approach that improves generalization in imitation learning by leveraging local neighborhood structure instead of direct state-action mappings. DARP predicts actions using k-nearest neighbors from expert demonstrations, their actions, and relative distance vectors between neighbor and query states, requiring no additional data or assumptions beyond standard behavior cloning. Experiments show 15-46% performance gains over behavior cloning across continuous control, robotic manipulation, and visual feature domains.
imitation learningbehavior cloningsemi-parametricretrieval-basedgeneralization
Collaborative Human-Agent Protocol (CHAP)
The paper introduces CHAP (Collaborative Human-Agent Protocol), a structured framework for accountable human-agent collaboration in production deployments. Addressing the lack of standards for shared workspaces, CHAP formalizes interactions through a Core (workspaces, participants, tasks, artefacts, evidence logs) and composable profiles for review, routing, signatures, and audit. The protocol captures human judgments as structured events with diffs, rationales, and content hashes, enabling non-repudiable decisions. Implementations include specification, reference code, conformance tests, and examples at https://github.com/BrightbeamAI/chap.
human-agent collaborationaccountability protocolstructured eventsevidence lognon-repudiable decisions
Multi-Turn Evaluation of Deep Research Agents Under Process-Level Feedback
The paper evaluates deep research agents (DRAs) through multi-turn feedback, contrasting self-reflection with process-level feedback using Research Gap Inference (RGI). RGI identifies gaps in research strategy by analyzing rubric criteria. Key findings: (i) self-reflection yields negligible improvement; (ii) process-level feedback boosts scores by 8-15 points with 35-40% incorporation; (iii) gains plateau as agents regress on 24% of criteria in subsequent turns. Current DRA architectures fail to sustain multi-turn improvement despite targeted guidance.
deep research agentsmulti-turn feedbackresearch gap inferencerubric criteriaprocess-level feedback
Hybrid Robustness Verification for Spatio-Temporal Neural Networks
This work introduces Spatio-Temporal Bound Propagation (STBP), a hybrid robustness verification framework for 3D CNNs processing video and volumetric inputs. STBP computes an exact closed-form characterization of the first convolutional layer and propagates certified bounds through subsequent layers using scalable approximations, leveraging realistic spatio-temporal constraints on adversarial perturbations. The method targets applications in action recognition (UCF-101), autonomous driving (Udacity), and medical imaging (MedMNIST). STBP achieves 1.7x higher certified robust accuracy under identical perturbation budgets compared to existing verification approaches. Additionally, the authors propose ST-Bench, a verification benchmark for autonomous driving and activity recognition to systematically evaluate verifiable robustness.
spatio-temporal constraintsrobustness verification3d cnnsbound propagationcertified accuracy
SearchSwarm: Towards Delegation Intelligence in Agentic LLMs for Long-Horizon Deep Research
The paper introduces SearchSwarm-30B-A3B, a language model with enhanced delegation intelligence for long-horizon tasks, addressing the challenge of finite context windows in agentic LLMs. The method employs a harness to generate supervised fine-tuning data by guiding task decomposition, subagent delegation, and result integration, internalizing these capabilities into the model. Evaluated on BrowseComp and BrowseComp-ZH, the model achieves 68.1 and 73.3 scores respectively, outperforming comparable-scale models.
delegation intelligencelong-horizon taskscontext windowsupervised fine-tuningagentic llms
Beyond Probabilistic Similarity: Structural, Temporal, and Causal Limitations of Retrieval-Augmented Generation in the Legal Domain
This paper critiques Retrieval-Augmented Generation (RAG) architectures in legal AI, identifying structural mismatches between probabilistic retrieval and the hierarchical, temporal, and institutional nature of legal knowledge. The authors analyze legal knowledge through a triad of properties—hierarchical structure, diachronic dynamism, and causal traceability—and diagnose three retrieval pathologies: mereological blindness, diachronic blindness, and causal opacity. They propose four architectural commitments—ontological primacy, event reification, bitemporal correctness, and deterministic interaction protocols—to address these limitations, focusing on legislative and constitutional retrieval while explicitly extending to interpretive time.
retrieval-augmented generationmereological blindnessdiachronic dynamismcausal traceabilitybitemporal correctness
Proxy Reward Internalization and Mechanistic Exploitation: A Learned Precursor to Reward Hacking and Its Generalization
The study introduces Proxy Reward Internalization and Mechanistic Exploitation (PRIME), a learned capability that assesses task correctness, predicts proxy acceptance, and reasons about exploitable proxy-gold gaps in reinforcement learning (RL). Using coding RL environments with exploitable pytest rewards, PRIME is measured via chain-of-thought monitoring, direct probes, and activation-level concept vectors. Results show PRIME emerges in stages before sustained reward hacking, with direct-probe scores forecasting hack onset and severity. PRIME adapts to evaluator changes, persists when gold reward suppresses hacking, and ablation of activation directions reduces hacking. PRIME also tracks out-of-domain misalignment, suggesting it as an early-warning signal for broader alignment risk.
proxy reward internalizationmechanistic exploitationchain-of-thought monitoringactivation-level concept vectorsproxy-gold gaps
Learning to Attack and Defend: Adaptive Red Teaming of Language Models via GRPO
The paper introduces AdvGRPO, a co-training framework for adaptive red teaming of language models that stabilizes GRPO (Generalized Reinforcement Learning with Policy Optimization) via dense multi-channel rewards and decoupled advantage normalization. The method employs a curriculum progressing from single-turn to closed-loop multi-turn attacks before bootstrapping co-training, where attacker and defender models are updated alternately. Results demonstrate that AdvGRPO produces highly effective, transferable attacks and co-trained defenders outperform baselines on safety benchmarks.
red teamingco-traininggrpomulti-channel rewardsadvantage normalization
Observability for Delegated Execution in Agentic AI Systems
The authors address the challenge of delegation-scoped execution observability in LLM-based agentic systems, where standard audit logs and execution traces fail to identify delegation assignments due to dynamic tool selection, execution variability, and sub-agent cooperation. They propose an agent-aware observability substrate comprising a lightweight gateway and a common information model that binds delegation context at execution time. This enables reliable cross-tool delegation-scoped reconstruction and direct forensic queries without heuristic time-window correlation, resolving the structural underdetermination of delegation-scoped attribution and access/share footprint reconstruction.
delegation-scopedobservabilityagentic systemsforensic queriesexecution traces
An 84-Format Numeric Catalog with Bit-Exact Conformance Vectors: A Vendor-Neutral Reference for FP8, BF16, MXFP4, and Microscaling Formats
The paper introduces a vendor-neutral reference catalog of 84 numeric formats spanning 13 families, addressing format proliferation in machine learning hardware (FP8, BF16, MXFP4, microscaling). It provides six bit-exact conformance packs with JSON documents, SHA-256 fingerprints, and anchor vectors for cross-validation. Packs are validated against ml_dtypes 0.5.4, with documented divergences treated as spec-permitted interpretation gaps. The catalog includes an IEEE P3109 v3.2.0 cross-walk and is publicly available under an open license.
numeric formatsbit-exact conformancefp8bf16microscaling
MeCo: One-Step MeanFlow-based Corrector for Multi-Channel Speech Separation
The paper introduces MeCo, a MeanFlow-based one-step generative corrector for multi-channel speech separation, addressing suboptimal listening quality in discriminative models. MeCo employs a conditional average velocity field to map discriminative estimates directly onto the clean speech manifold in one step, enhanced by Data-Space Optimization (DSO) integrating an xr-loss for generative objectives and Endpoint SI-SDR loss for signal fidelity. Experiments show MeCo achieves state-of-the-art performance with minimal computational overhead, improving both signal fidelity and listening quality in in-domain and out-of-domain scenarios.
meanflowspeech separationgenerative correctordata-space optimizationsi-sdr
(Auto)formalization is supposed to be easy: Trellis process semantics for spelling out rigorous proofs
Trellis introduces an autoformalization system using LLM agents in a constrained workflow to incrementally refine natural language proofs into Lean formalizations. The method enforces rigorous proof elaboration through process semantics inspired by mathematical practice, avoiding task-specific agent training. The system demonstrates feasibility by producing an end-to-end Lean formalization of a recent Ramsey theory result, targeting reliable autoformalization with generalist models and modest computational resources.
autoformalizationllm agentslean theorem provingprocess semanticsramsey theory
Correlation Is Not Enough: Embedding Human Metadata for Individual Causal Discovery
The study addresses the critical flaw in biomedical language models where high cosine similarity scores erroneously suggest relationships between unrelated cross-domain terms (e.g., cortisol and stock-market volatility). The authors introduce BODHI, a method that improves embedding discrimination through contrastive learning and hard negative mining from knowledge graphs. Results show a 0.828 BIOSSES correlation (up from 0.633) and 2.30x domain separation, with optimized inference latency of 10 ms (133x speedup) using OpenVINO on Intel Xeon. The work includes released benchmarks, training corpora, and optimization scripts.
biomedicalembeddingcontrastiveknowledge graphopenvino
Transition-Based Digital Twin Modelling for Alzheimer's Disease under Sparse Longitudinal Data
The study introduces a transition-based digital twin framework for personalized Alzheimer's disease (AD) progression modeling under sparse longitudinal data. The method integrates multimodal clinical data (cognitive assessments, MRI phenotypes) from ADNI, combining transition-based and sequence-based approaches to capture temporal dependencies. Results show transition modeling between adjacent visits outperforms sequence-based methods in predictive accuracy (72.3% vs 68.1% diagnosis classification), suggesting better data efficiency for sparse clinical settings, while sequence models remain useful for uncertainty-aware trajectory forecasting.
digital twinalzheimer's diseasetransition-based modelinglongitudinal datapredictive uncertainty
Visual Prompting Meets Feature Reconstruction-Based Anomaly Detection with Dual-Teacher Supervision
The paper introduces three innovations for anomaly detection in visual data: (1) a visual prompting pipeline with foreground-background masking, (2) unfreezing teacher models in student-teacher frameworks for domain adaptation, and (3) diffusion-based synthetic data augmentation. Using Masked Multiscale Reconstruction (MMR) as backbone, the method achieves a 3.5 percentage point improvement over prior state-of-the-art on the AeBAD dataset, addressing challenges like object scale variation and viewpoint changes that hinder existing approaches.
visual promptinganomaly detectiondiffusion augmentationstudent-teacher modelsmasked multiscale reconstruction
SpatialWorld: Benchmarking Interactive Spatial Reasoning of Multimodal Agents in Real-World Tasks
The paper introduces SpatialWorld, a unified benchmark for evaluating interactive spatial reasoning in multimodal agents across real-world tasks. The benchmark integrates eight simulation backends under a simulator-agnostic protocol, featuring 760 human-annotated tasks with vision-only partial observability and text-based action interfaces. Evaluation of 15 advanced agents reveals significant challenges, with GPT-5 achieving only a 17.4% task success rate and Qwen-3.5 reaching 14.1%, highlighting bottlenecks in active exploration and long-horizon planning.
spatial reasoningmultimodal agentsbenchmarkpartial observabilitytask success rate
Frequency-based Constrained Sampling for Interval Patterns
The paper introduces CFips, a constrained sampling method for interval patterns that integrates user-defined syntactic constraints directly into the sampling process. CFips decomposes constraints into elementary predicates on interval bounds while maintaining exact sampling guarantees, ensuring patterns are sampled proportionally to their frequency within the constrained space. Experimental results demonstrate that CFips efficiently completes mining tasks that would otherwise timeout under traditional methods.
interval patternsconstrained samplingfrequency-basedsyntactic constraintsexact sampling guarantees
From 0-to-1 to 1-to-N: Reproducible Engineering Evidence for MetaAI Recursive Self-Design
The paper introduces a framework for evaluating recursive self-design in AI systems, operationalized through four criteria: inspectable target system, meta-level modifier, feedback-directed selection, and recursive continuation. It analyzes existing systems like Darwin Goedel Machine (DGM), STOP, Goedel Agent, and ShinkaEvolve against these criteria, with DGM demonstrating performance improvements from 20% to 50% on SWE-bench Verified and 14.2% to 30.7% on Polyglot over 80 iterations. The study also presents MetaAI-Mini, a reproducible protocol for HumanEval-based evaluation, though it lacks completed model runs.
recursive self-designmeta-level modifierfeedback-directed selectiondarwin goedel machinehumaneval
End-to-End Context Compression at Scale
The paper introduces Latent Context Language Models (LCLMs), an encoder-decoder architecture for efficient KV cache compression in long-context language model inference. Through systematic architecture search and pre-training of 0.6B-encoder, 4B-decoder models on 350B+ tokens, LCLMs achieve compression ratios of 1:4 to 1:16 while improving the Pareto frontier for general-task performance, compression speed, and memory usage. Results demonstrate LCLMs' effectiveness as backbones for long-horizon agents, enabling adaptive context expansion from compressed representations.
kv cacheencoder-decodercontext compressionlatent embeddingslong-context inference
Muon Learns More Robust and Transferable Features than Adam
The paper demonstrates that Muon, a state-of-the-art optimizer, learns more robust and transferable features than Adam and SGD across transformers and CNNs. Through evaluations on corrupted images/texts and layer-wise probes, Muon exhibits larger logit margins and higher robustness. Transferability is shown via linear classifiers and fine-tuning, supported by greater hidden-state diversity (measured by effective rank). Theoretical analysis confirms Muon's superiority in margin maximization and feature diversity in multi-component classification tasks.
muon optimizerlogit marginseffective rankfeature transferabilityrobust features
ArtiFact: A Large-Scale Multi-Modal Cultural Heritage Dataset
We introduce ArtiFact, a large-scale multi-modal cultural heritage dataset comprising 651,045 museum records from the Metropolitan Museum of Art, the Art Institute of Chicago, and the Rijksmuseum, integrating tables, text, and images. To demonstrate its utility, we evaluate two downstream tasks: cross-modal error detection, where we inject 130,209 records with a curated taxonomy of seven error categories, revealing challenges in detecting domain-specific errors like material anachronisms; and semantic query processing, where current systems exhibit limitations in handling culturally contingent queries. ArtiFact establishes itself as a challenging benchmark for multi-modal data management research.
multi-modal datasetcultural heritagecross-modal error detectionsemantic query processingmuseum records
Do Video Foundation Models Understand Intuitive Physics? A Layerwise Probing Analysis
This work investigates whether pretrained video foundation models encode intuitive-physics knowledge in their frozen representations, analyzing variations across model families, layers, and probe types. Using frozen-feature probing on IntPhys2 and Minimal Video Pairs (MVP), the study compares predictive joint-embedding models (V-JEPA), masked reconstruction models (VideoMAE), and a diffusion-based video generator (LTX-Video). Results show V-JEPA achieves the strongest performance, particularly with temporal dynamics probes, while VideoMAE remains competitive and LTX-Video recovers weaker but non-trivial signal. Layerwise analysis reveals physics-relevant information peaks at intermediate-to-late depths, and temporal disruption significantly reduces performance, especially on MVP.
frozen-feature probingintuitive-physicsjoint-embedding modelsmasked reconstructiondiffusion-based
FMplex: Model Virtualization for Serving Extensible Foundation Models
FMplex introduces a model-serving system that virtualizes foundation models (FMs) to enable shared backbone deployment across multiple downstream tasks. By presenting each task with a virtual foundation model (vFM) backed by a shared physical FM, FMplex achieves resource efficiency while maintaining task-specific customization and isolation. The system incorporates a batch-aware fair-queueing scheduler for weighted task-level sharing and inter/intra-task batching. Evaluated on 7 FM backbones (16 variants) and 92 tasks, FMplex reduces latency by up to 80% compared to spatial partitioning and 33.3% over best-effort co-location, while supporting 6x more tasks at cluster scale.
foundation modelsvirtualizationbatch-aware schedulingtask isolationresource efficiency
ATN3D: Density-Aware LiDAR-Radar Early 3D Object Detection Under Extreme Sparsity
ATN3D introduces a LiDAR-Radar fusion framework for 3D object detection under extreme sparsity in autonomous driving. The method employs density-aware early fusion with cross-modal gating, occupancy-gated neighborhood aggregation using circular kernels, evidence-conditioned channel self-attention, and a range-aware loss function. Evaluated on the VoD benchmark, ATN3D achieves +3.55% mAP improvement in clear weather and +8.41% mAP under heavy fog, with notable gains for >30m objects (+3.33% clear, +2.09% fog), demonstrating robust long-range detection in sparse conditions.
3d object detectionlidar-radar fusionsparse sensingcross-modal gatingrange-aware loss
ReCoVLA: VLM-Guided Reward Compilation for Failure Recovery in Vision-Language-Action Policies
ReCoVLA introduces a failure-conditioned residual recovery framework for vision-language-action (VLA) policies, addressing brittleness in off-nominal states. The method keeps a pretrained VLA policy frozen, employing a vision-language model (VLM) to infer failure modes and recovery stages, then compiling structured rewards for residual-policy training. This decouples high-level semantic understanding from low-level corrective control. Experiments demonstrate a simulation success rate improvement from 36.7% to 66.7% over fine-tuned baselines, with 61.7% success in zero-shot sim-to-real transfer across manipulation tasks.
vision-language-actionfailure recoveryresidual policyreward compilationsim-to-real
Powering the Future of AI: Navigating the Trade-offs for Europe's Energy Transition and Net-Zero Goals
This study quantifies the energy and emissions impacts of AI-driven hyperscale data centers (DCs) in Europe through 2050 using a spatially explicit optimization model across 21 scenarios. Results show AI could add 73-723 TWh demand by 2050, risking 67-181 MtCO2 overshoots, with infrastructure geography determined by firm power and flexibility rather than clean energy abundance. Moderate scenarios require 200 additional firm generation hours, increasing LCOE by 35 EUR/MWh, while pessimistic cases need 70-226 GW capacity expansion. Workload dynamics significantly affect dispatch, flexibility, and emissions, though efficiency gains mitigate capacity needs. Achieving net-zero by 2050 remains feasible but requires policy adaptation to prevent intermediate emission risks.
hyperscale data centersspatially explicit optimizationfirm powerlcoecarbon-neutral
AGENTSERVESIM: A Hardware-aware Simulator for Multi-Turn LLM Agent Serving
AGENTSERVESIM introduces a hardware-aware simulator for multi-turn LLM agent serving, addressing the limitations of existing stateless request-level simulators. The simulator evaluates serving policies at program granularity through composable modules: Program Orchestrator, Tool Simulator, Session-Aware Router, and KV Residency Model, which handle program identity, tool-induced gaps, cache-aware dispatch, and KV-cache placement across memory hierarchies. AGENTSERVESIM reproduces real-system behavior within 6% error across key performance metrics while running entirely on commodity CPUs, enabling scalable exploration of agent-serving policies without costly accelerator deployments.
multi-turn llm agentskv-cache managementhardware-aware simulatortool-induced gapscache-aware dispatch
Shape Formation for the Cooperative Transportation of Arbitrary Objects Using Multi-Agent Reinforcement Learning
The authors propose a multi-agent reinforcement learning approach for autonomous shape formation in cooperative object transportation by multi-robot systems. The method addresses formation control, cooperative navigation, and collision avoidance for arbitrary-shaped objects with non-uniform mass distribution. Evaluations demonstrate reliable balanced formations in cluttered environments, generalizing to complex geometries and varying robot counts.
multi-agent reinforcement learningformation controlcooperative navigationnon-uniform mass distributionautonomous positioning
Closure-Validated Circuit Discovery in Attention Heads: Co-activation Proposes, Ablation Disposes
The study introduces a closure-validation method for identifying attention-head circuits in transformer models, challenging the reliance on co-activation clustering alone. By adapting sparse-autoencoder clustering to attention heads and validating through causal ablation, the authors conduct closure tests comparing ablation effects to random controls. Results across Pythia 1B, OLMo 1B, and OLMoE-1B-7B models demonstrate that co-activation clusters pass closure tests in dense models but fail in Mixture-of-Experts architectures, where ablation paradoxically improves loss. The findings emphasize that co-activation signals propose circuits, but closure validation is essential for confirmation.
attention-head circuitsco-activation clusteringcausal ablationclosure testmixture-of-experts
Next-Token Prediction Learns Generalisable Representations of Sleep Physiology
The study introduces Hypnos, a multi-modal sleep foundation model trained via next-token prediction, demonstrating its superiority over masked-reconstruction and contrastive approaches. Hypnos tokenizes eight physiological modalities (e.g., EEG, ECG) from 20,000 polysomnography recordings using residual vector quantization, then trains an auto-regressive RQ-Transformer for joint next-token prediction. Evaluations show Hypnos matches supervised sleep stage classification baselines with 100× less labeled data and generalizes to daytime physiology, outperforming dedicated ECG models in atrial fibrillation detection.
next-token predictionresidual vector quantizationmulti-modal foundation modelpolysomnographyauto-regressive transformer
I Was Scrolling and Then I Saw a Pregnant Strawberry
The paper analyzes AI-generated minidramas featuring anthropomorphized characters, revealing gendered and racialized narrative structures. Female characters are frequently associated with moral transgression and reproductive roles, while racialization processes are embedded in plotlines. Employing feminist film theory, critical race theory, and platform studies, the study identifies how generative AI's aesthetic (softness, roundness, cuteness) launders ideological content, facilitating its spread despite moderation. Methodologically, it combines personal observation and close reading to critique computational creativity's cultural implications.
anthropomorphizedracializationgenerative aicontent moderationcomputational creativity
Seeing the Hivemind: A Consensus-Aware Interaction Technique for Mitigating AI Homogenization
The Semantic Repulsion Technique (SRT) is introduced to mitigate AI homogenization in creative tasks by increasing semantic diversity and reducing consensus phrases. SRT was evaluated computationally and through a user study with 16 participants who regularly use AI for creative tasks. Computational results show SRT increases semantic diversity by 85-167% and reduces consensus phrases by 43-95%. User study results indicate SRT outputs received higher usefulness (p = .019, W = .208) and coherence ratings (p = .006, W = .260), with 68.8% of participants willing to use SRT-Strong for multiple tasks versus 18.8% for baselines. Originality and coherence ratings were positively correlated (ρ= +.40 to +.67), suggesting divergence does not compromise readability.
semantic repulsion techniquesemantic diversityconsensus phrasesai homogenizationcreative tasks
Optical Reasoning: Rethinking Images as an Expressive Reasoning Medium Beyond Text
The paper introduces optical reasoning, a novel approach using images as a standalone medium for reasoning tasks, surpassing traditional text-based methods. It presents two variants: typographic-based optical reasoning, optimizing visual layouts for compact rationale rendering, and graphical-based optical reasoning, integrating text and graphical elements into structured visual rationales. Evaluations across mathematical, scientific, and interleaved-modal reasoning benchmarks demonstrate that optical reasoning matches or exceeds text reasoning, achieving 1.96 times token efficiency and reducing reasoning tokens by 28.57% on language tasks and 16% on multimodal tasks.
optical reasoningtypographic-basedgraphical-basedtoken efficiencymultimodal tasks
TABVERSE: Benchmarking Cross-Format Table Understanding in LLMs and VLMs
The paper introduces TABVERSE, a multimodal benchmark for evaluating table understanding in LLMs and VLMs across different structural formats (HTML, Markdown, LaTeX) and rendered images while holding content constant. It systematically assesses three tasks: Question Answering (QA), Structural Understanding Capability (SUC), and Structure Reconstruction (SR). Results indicate that representation choice significantly impacts performance, with HTML generally outperforming other formats in structured text, while row-sensitive tasks and LaTeX reconstruction remain challenging. The benchmark reveals modality-specific gaps, emphasizing the importance of controlled format variation in table evaluation.
tabversemultimodal benchmarkstructural understandingrepresentation effectsmodality gap
CT-VAM: A Cerebello-Thalamic-Inspired Vision-Action Model for Efficient Visuomotor Control
The paper introduces CT-VAM, a cerebello-thalamic-inspired vision-action model for efficient task-conditioned visuomotor control. The model employs TARS (Thalamic Action Routing Stream), a stream-separated conditional attention decoder that independently processes action, visual, and task streams to prevent sensory token overload. With 68M parameters, CT-VAM matches LIBERO benchmark performance of larger vision-language-action models while reducing inference latency, and enables high-frequency control via flow-consistent inpainting for asynchronous chunk execution on resource-constrained platforms.
visuomotor controlstream-separated attentionaction chunksinference latencyflow-consistent inpainting
Self-Explainability in Self-Adaptive and Self-Organising Systems: Status and Research Directions
The article contributes a systematic literature review and unified framework for Self-Explainability (SX) in self-adaptive and self-organising systems. Through analysis of existing approaches, it develops a taxonomy, definition, and Levels of Self-Explainability to position current and future research. Results indicate most SX approaches remain conceptual with few implementations, and reveal a lack of formal evaluation standards. The work establishes foundational terminology and identifies key research gaps in this emerging field.
self-explainabilityself-adaptive systemssystematic literature reviewexplainable aitaxonomy
PRISM: Recovering Instruction Sets from Language Model Activations
PRISM introduces instruction set retrieval, a method to decode hidden states from frozen language models into comprehensive lists of active instructions, constraints, and subgoals. The approach employs an activation-conditioned interpreter trained with judge-guided GRPO to optimize coverage of instructions while penalizing unsupported claims. Evaluations across benign, constrained, prompt-injection, and hidden-objective scenarios demonstrate PRISM's superiority over activation-to-language baselines, particularly for security-critical objectives.
instruction set retrievalactivation-conditioned interpreterjudge-guided grpohidden statesprompt injection
Safe-RULE: Safe Reinforcement UnLEarning
The paper introduces Safe-RULE, a novel safe reinforcement unlearning framework for defending against data poisoning attacks in offline safe reinforcement learning (Safe RL). The method eliminates poisoned data influence without full retraining or environment access, explicitly optimizing both task performance and safety constraints during unlearning. Experiments on benchmark Safe RL tasks demonstrate significant safety performance improvements against poisoning attacks.
offline safe rldata poisoningreinforcement unlearningsafety constraintspolicy learning
AI Scientists Are Only as Good as Their Evidence: A Stratified Ablation of Proprietary Data and Reasoning Skills in Drug-Asset Valuation
This study demonstrates that proprietary evidence, not reasoning scaffolds or model quality, is the primary limiting factor in AI Scientist agents' drug-asset valuation capabilities. Through a three-arm ablation study, the authors evaluate a production valuation agent with varying access to data: plain web-only LLM (A), public structured tools with reasoning scaffolds (B), and proprietary Noah AI corpus (C). Results show that C achieves 0.96 recovery of curated gold competitive records and 7.43 completeness-aware decision utility, significantly outperforming A (1.76) and B (2.57). Reasoning scaffolds improve calibration and discipline but are insufficient without proprietary evidence, which sets the upper bound for decision quality.
drug-asset valuationreasoning scaffoldsproprietary evidencecompleteness-aware decision utilitythree-arm ablation
FuseFSS: Efficient Secure LLM Inference with Function Secret Sharing
FuseFSS introduces a compiler for efficient secure LLM inference using function secret sharing (FSS), replacing per-operator protocols with a unified pipeline. The method compiles scalar fixed-point operators into batched FSS evaluations, combining interval lookup and predicate bits extraction. Results show 1.24×--1.50× speedup, 9%--16% communication reduction, and 14%--23% lower preprocessing overhead on BERT and GPT-style models while maintaining accuracy.
secure inferencefunction secret sharingfixed-point arithmeticllmcompiler optimization
SecureClaw: Clawing Back Control of LLM Agents
SecureClaw introduces a dual-boundary architecture addressing two security failures in tool-using LLM agents: unauthorized external actions and sensitive plaintext exposure. The method employs a trusted gateway for sensitive reads, replacing raw values with opaque handles and bounded summaries, and implements a PREVIEW→COMMIT protocol for external state changes, ensuring only authorized canonical requests are executed. Evaluated across AgentDojo, AgentLeak, and Agent Security Bench (ASB), SecureClaw achieves 0% attack success rate on ASB, 0.64% on AgentDojo, and 3.23% overall leak on AgentLeak's attacked parity lane, maintaining usable task utility.
dual-boundary architecturetrusted gatewaypreview→commit protocolopaque handlesbounded summaries
Model Poisoning Against Federated Model Adaptation with Chain of Bit-Flips
We introduce a hardware-fault-based model poisoning attack against Federated Learning (FL) systems, leveraging bit-flip vulnerabilities to implant task-agnostic backdoors during FL training. The attack crafts a backdoor offline from the pretrained model and induces hardware faults in a single local model's parameters during training. Experiments demonstrate successful implantation across diverse models and datasets, with 10 faults per malicious client and 19 total occurrences on ResNet-18 achieving a 94% attack success rate. Practical constraints of Rowhammer, the preferred attack vector, are discussed alongside potential defenses.
federated learningmodel poisoningbit-flipbackdoor attackrowhammer
Emergence of Context Characteristics Sensitivity in Large Language Models
This work investigates how large language models (LLMs) develop sensitivity to context characteristics during instruction fine-tuning (IFT). The authors measure shifts in context usage preferences across three IFT stages: supervised fine-tuning (SFT), direct preference optimization (DPO), and reinforcement learning with verifiable rewards (RLVR). Experiments on four models and three datasets demonstrate that SFT increases preference for easily understandable contexts (high length, context-query similarity, fluency), while post-SFT dynamics either reinforce or resolve these preferences based on dataset characteristics. Findings indicate that context utilization is actively reshaped at each IFT stage, highlighting the importance of balanced dataset design for robust instruction tuning.
instruction fine-tuningcontext characteristicssupervised fine-tuningdirect preference optimizationreinforcement learning
Closing the Prior-Posterior Loop: Self-Reflective Molecular Design with Analysis-Driven LLM Iteration
The paper introduces a self-reflective molecular design framework that replaces scalar feedback with physicochemical rationales from first-principles calculations, transforming large language models (LLMs) into causal reasoners. The system combines retrieval-augmented generation with a reflection module that feeds orbital energies, atomic charges, and electron densities back into the design loop. On HOMO-LUMO gap targets (1.0-5.0 eV), it achieves deviations as low as 0.0003 eV and 100% success on moderate tasks, outperforming scalar-feedback baselines. The method generalizes to dipole-moment design and works across five LLM backbones.
self-reflective designhomo-lumo gapretrieval-augmented generationfirst-principles calculationsstructure-property-relationship
From Rigid to Dynamic: Entropy-Guided Adaptive Inference for Long-Context LLMs
EntropyInfer introduces a training-free framework for adaptive inference in long-context LLMs, addressing limitations of fixed sparsity patterns and uniform KV cache budgets. The method leverages observed entropy patterns—Rigid Heads (near-zero entropy) and Dynamic Heads (fluctuating entropy)—to allocate compute adaptively at the granularity of individual heads and segments during prefilling. For decoding, it employs a latent KV cache compression scheme that prioritizes critical entries based on generated output tokens. Experiments on Llama, Qwen, and openPangu models demonstrate up to 2.39× speedup beyond 100k tokens with minimal quality degradation compared to full attention, outperforming SnapKV, AdaKV, and CritiPrefill.
entropy-guided inferencekv cache compressiondynamic headsrigid headslong-context llms
Deterministic Integrity Gates for LLM-Assisted Clinical Manuscript Preparation: An Auditable Biomedical Informatics Architecture
The paper introduces an integrity-gate architecture for LLM-assisted clinical manuscript preparation, combining deterministic verification with prose-level probes where necessary. The method decomposes workflows into 43 skills (MedSci Skills toolkit), with 21 standard-library detectors enforcing deterministic checks at stage transitions. Evaluation on STARD/PRISMA/STROBE pipelines showed perfect defect detection (27/27) versus 11/27 for single-prompt LLM review, particularly excelling on code, bibliography, and style defects.
integrity gatesdeterministic verificationclinical manuscript preparationllm-assisted writingmedsci skills
Targeting World Models to Compromise Robot Learning Pipelines
The paper identifies a novel security vulnerability in robot learning pipelines where world models can be covertly compromised through data poisoning. Attackers inject malicious prompts or transition dynamics into seemingly safe teleoperated datasets, which only manifest as dangerous synthetic trajectories when processed by world models. Experiments demonstrate successful policy compromises in both action-conditioned and vision-language-action (VLA) world models, including end-to-end backdoors in downstream deep reinforcement learning policies. These findings highlight the need for secure world model architectures and supply chain reassessment.
world modelsdata poisoningrobot learningbackdoor attacktransition dynamics
LLM-Orchestrated Conformance Checking in Stroke Care Without Computer-Interpretable Guidelines
The study presents an LLM-orchestrated framework for healthcare conformance checking without requiring Computer-Interpretable Guidelines (CIGs). The modular architecture employs multiple LLMs to extract patient traces from discharge letters, derive normative rules from textual guidelines, translate rules into executable scripts, and compute a Trace Conformance Indicator. Evaluated on stroke care data from Alessandria Hospital, the system processed hundreds of traces against 50 guideline-derived rules, demonstrating 86% conformance and validating both the method's feasibility and high guideline adherence at the site.
conformance checkinglarge language modelsclinical guidelinestrace extractionstroke care
Memory Beyond Recall: A Dual-Process Cognitive Memory System for Self-Evolving LLM Agents
The paper introduces DCPM, a dual-process cognitive memory system for LLM agents that reorganizes memory along a hierarchical cognitive capability spectrum. The system employs two processes: a synchronous 'daytime writer' (System1) for recording belief revisions via doubly linked supersedes chains, and an asynchronous 'nighttime engine' (System2) for inducing schemas, intentions, and cross-domain abstractions. Evaluations on LongMemEval, PersonaMem, and PersonaMem-v2 show System2 improves implicit cross-session inference by up to +5.20 on PersonaMem-v2, while minimally affecting span recall performance.
dual-process theorybelief revisioncross-domain abstractionschema inductionsupersedes chains
Emergent alignment and the projectability of ethical personas
This paper introduces 'emergent alignment' as the converse of emergent misalignment, supporting the persona selection hypothesis (PSM) through constitutional AI (CAI) finetuning. Four constitutions (deontology, consequentialism, virtue ethics, human-subordinate alignment) guide supervised finetuning (SFT) on narrow safety tasks, inducing alignment across broader safety categories. Ethical persona diagnostics reveal models adopt expected profiles (e.g., consequentialist models favor utilitarian beliefs), but projectability varies between broad/narrow finetuning. Results suggest alignment strategies should prioritize projectability alongside in-distribution safety performance.
emergent alignmentpersona selection hypothesisconstitutional aisupervised finetuningethical persona
A Finetuned SpeechLLM for Joint Multi-Granular L2 Assessment and Natural-Language Rationales
The paper introduces a finetuned SpeechLLM for joint multi-granular L2 speech assessment, combining ordinal label prediction (sentence-level accuracy/fluency/prosody, word/phoneme-level accuracy) with natural-language rationale generation. The model employs a hybrid objective of supervised fine-tuning and Bounded Direct Preference Optimization, achieving competitive performance on SpeechOcean762 while maintaining interpretability. Analysis reveals plausible sentence-level rationales but degraded faithfulness at finer granularities due to sparse references and weak alignment with token-level labels.
speechllml2 assessmentbounded direct preference optimizationmulti-granularrationale generation
TheoremBench: Evaluating LLMs on Theorem Proving in Formal Mathematics
The paper introduces TheoremBench, a Lean4 benchmark for evaluating LLMs on formal theorem proving beyond competition-style problems. The benchmark comprises ~100 classical theorems in two versions: a plain main version and a premised version with structured subtheorems, enabling assessment of both final proofs and partial progress. Experiments demonstrate that explicit premises improve performance for Lean4-capable provers, while new metrics reveal biases toward easy subtheorems and inefficient tactic traces. The work highlights structural benchmark design's importance for evaluating formal reasoning in Lean4.
theorem provinglean4formal mathematicsllm evaluationproof structure
AliyunConsoleAgent: Training Web Agents in Real-World Cloud Environments via Distillation and Reinforcement Learning
The paper introduces AliyunConsoleAgent, a web agent framework for automated documentation verification in cloud consoles, addressing the challenge of UI-documentation divergence across hundreds of rapidly evolving cloud products. The method combines supervised fine-tuning on distilled trajectories from frontier models with reinforcement learning using Group Relative Policy Optimization (GRPO) and a dual-channel outcome reward model, supported by Terraform-based resource provisioning and LLM-driven on-demand provisioning for noise isolation. On a 278-task benchmark, the 32B-parameter model achieves 63.52% success rate (20.24pp improvement over base) at 92% lower cost than proprietary models, narrowing the performance gap to 1.82pp.
web agentdocumentation verificationgroup relative policy optimizationterraform provisioningoutcome reward model
SIFT: Selective-Index For Fast Compute of RAG Prefill by Exploiting Attention Invariance
SIFT introduces a selective-indexing method to accelerate Retrieval-Augmented Generation (RAG) prefill by exploiting attention invariance, reducing time to first token (TTFT) without significant accuracy loss. The method identifies and stores fine-grained locations of high attention scores offline using two compact bit vectors, leveraging local-attention invariance and cross-attention consistency to minimize redundant computations. SIFT achieves a 1.71x speedup in TTFT with less than 1% accuracy degradation, while reducing storage by up to 24,000x compared to full KV tensor caching.
retrieval-augmented generationattention invariancekv tensorsprefilltime to first token
Bayesian Selective Latent Inference for Wastewater-First Influenza Monitoring
The paper introduces Bayesian Selective Latent Inference (BSLI), a method for wastewater-first influenza monitoring that addresses source ambiguity in surveillance data. BSLI maintains a posterior over latent disease burden and identifiability, uses scientific gates to certify answerability, and optimizes query-stop decisions via a cost-calibrated Bellman policy. Theoretical results include variational properties, answerability guarantees, and Bellman-optimality proofs. Evaluated on 5,933 forecasting and 3,102 source-ambiguity episodes from public data, BSLI improves cost-performance while maintaining conservative abstention under ambiguity.
bayesian inferencewastewater surveillancesource ambiguitybellman policylatent identifiability
LargeMonitor: Monitoring Online Task-Free Continual Learning via Large Pretrained Models
LargeMonitor introduces a framework for online task-free continual learning (TFCL) that leverages large pretrained models for autonomous adaptation. The method employs a frozen large vision model (LVM) for zero-shot drift detection and a large multimodal model (LMM) for semantic diagnosis of stream variations, enabling shift-specific optimization strategies. Experiments show improved performance in TFCL benchmarks through precise drift detection and adaptive learning.
online task-free continual learninglarge vision modelsdrift detectionlarge multimodal modelsadaptive optimization
WeaveBench: A Long-Horizon, Real-World Benchmark for Computer-Use Agents with Hybrid Interfaces
WeaveBench introduces a long-horizon benchmark for computer-use agents (CUAs) operating in hybrid interfaces, combining GUI, CLI, and code operations across 114 real-world tasks from 8 domains. The benchmark evaluates agents on an Ubuntu desktop with a trajectory-aware judge that inspects deliverables, logs, and action traces to detect shortcut behaviors. Results show frontier models achieve only 41.2% PassRate, with outcome-only grading overestimating performance, highlighting the need for rigorous cross-interface evaluation.
computer-use agentshybrid interfaceslong-horizon taskstrajectory-aware judgeubuntu desktop
Context-Aware Deep Learning for Defect Classification in Atomic-Resolution STEM
The authors present a context-aware deep learning framework for defect classification in atomic-resolution STEM that integrates image contrast with experimental metadata (composition, beam energy, detector geometry). Using a dataset of ~55 million simulated patches across 576 cases in 96 doped monolayer transition-metal dichalcogenides, they demonstrate that contextual conditioning transforms defect classification from ill-posed to well-posed. The method achieves 98% accuracy on simulations and near-human agreement on experimental data, with a 94% reduction in posterior entropy, linking image contrast to underlying physical conditions.
defect classificationcontext-aware learningatomic-resolution stemtransition-metal dichalcogenidesmultimodal ai
Harness Engineering for Physical AI: Robot Middleware Is the Harness Layer
The paper proposes robot middleware as the harness layer for Physical AI, mediating control, computation, and communication in deployed robotic systems. It identifies three key enforcement functions—Projection, Isolation, and Transfer—currently implemented ad hoc in robot systems. The authors argue middleware should natively support these functions, exemplified through a ROS 2 Harness Profile that enforces output regions, inference budgets, and operating regimes across ROS 2, DDS, and Zenoh.
physical airobot middlewareharness layerros 2enforcement functions
AI Assurance in UK Defence: Challenges in Operationalising JSP 936
The report analyzes implementation challenges of JSP 936 Part 1 for AI assurance in UK Defence through a structured interpretive review, identifying eight thematic obstacles: evidence adequacy, human-AI interaction, operational environment definition, systems integration, performance maintenance, safety/security analysis, ethical measurement, and AI complexity mitigation. Findings indicate the directive offers governance foundations but faces unresolved technical and organizational issues stemming from socio-technical system nature, deployment uncertainties, and methodological limitations. The study highlights needs for additional methods and guidance to enable responsible AI adoption in Defence contexts.
ai assurancesocio-technical systemsoperational environmentsystems integrationethical measurement
Correct Looks Better: Pairwise Comparisons Reveal Accuracy Rankings
The study demonstrates that pairwise comparisons using Elo rankings strongly correlate (Spearman ρ > 0.9) with ground-truth accuracy rankings across five converted benchmarks. Methodologically, it analyzes free-form generative evaluations, showing Elo outperforms direct evaluation when judge models are weak. Results indicate minimal impact from stylistic cues or judge biases, with repetition (echo) identified as a causal factor in preferences for correct/incorrect pairs.
pairwise comparisonselo rankingsspearman correlationgenerative evaluationsjudge bias
Capacity, Not Format: Rethinking Structured Reasoning Failures
The study demonstrates that structured output formats like JSON degrade model performance primarily due to capacity constraints rather than formatting overhead. Using information-matched controls and a complexity gradient across 4 models and 5 benchmarks, the authors show format penalties emerge in capacity-limited models (e.g., Haiku drops 36.2pp, GPT-4o-mini drops 28.0pp) and scale with schema complexity. Delayed-structure ablation recovers most accuracy loss (80-87%), supporting capacity competition as the mechanism. Results qualify claims of frontier model immunity, with Opus 4.7 dropping 5.3pp on AIME math under JSON formatting.
structured outputcapacity constraintsjson degradationcomplexity gradientdelayed-structure ablation
Can Data Work be Reparative?
The paper presents an ethnographic study of a feminist civic-tech initiative developing reparative approaches to dataset construction for online safety systems. Through qualitative analysis of collaborative data work with impacted communities, the authors identify tensions in achieving just compensation and collective governance of AI datasets. The study employs an STS-informed reparative justice framework to argue that responsible AI requires reconfiguring accountability relationships, centering marginalized voices in dataset production rather than technical systems. Findings highlight structural challenges in transforming prevailing data work norms.
reparative justicedataset governancefeminist aicollaborative data workaccountability structures
SAILS: Surrogate-based Analysis of Interactions via Local Effect Smooths
SAILS introduces a model-agnostic framework for analyzing pairwise feature interactions in black-box models through interpretable GAM surrogates. The method fits local effect smooths to isolate interaction components, enabling (1) detection via significance-test-derived heuristics, (2) categorization into linear/product-separable/non-product-separable types, and (3) type-specific visualizations. Empirical validation on synthetic and real-world data demonstrates effectiveness for pairwise interactions, though limitations exist under strong feature correlations or higher-order interactions. SAILS advances XAI by characterizing interaction functional forms beyond mere detection.
feature interactionsgeneralized additive modelslocal effectsmodel interpretabilityblack-box analysis
RunAgent SuperBrowser: A Theory of Autonomous Web Navigation Grounded in Human Browsing Behaviour
SUPERBROWSER introduces an autonomous web-navigation agent grounded in human browsing behavior, operationalized through a perception-cognition-action triad. The system employs a vision-first bounding-box pipeline for candidate interaction labeling, a three-role brain architecture (Orchestrator, Planner, Worker) for strategic and operational reasoning, and a structured Ledger for efficient memory management. Action execution utilizes a three-tier click cascade with humanized Bezier motion and a chevron-aware bounding-box snapper. On the Mind2Web Hard benchmark (66 tasks), SUPERBROWSER achieves 89.47% success, outperforming all published open/research browser-agent baselines. The performance gain is attributed to the consistent application of a cognitive contract across the system.
bounding-box pipelinethree-role brainstructured ledgerclick cascadecognitive contract
From Coarse to Fine: Managing Temporal Granularity in Spatio-Temporal Data for Fine-Grained Traffic Prediction
The paper introduces the Spatial-Temporal Refinement Predictor (STRP), a granularity-aware framework for fine-grained traffic prediction using coarse-grained spatio-temporal data. STRP combines Tree Convolution for spatial dependency modeling and Inverse Dilated Convolution for temporal extrapolation, supporting both window-based and duration-based prediction settings. Evaluations on six benchmark datasets demonstrate STRP's superior accuracy and efficiency over state-of-the-art baselines, addressing the temporal granularity mismatch in traffic data systems.
spatio-temporal datatraffic predictiontree convolutioninverse dilated convolutiongranularity mismatch
Real-time body pose non-verbal communication with a consistency-based reliability measure
The paper introduces a dataset for recognizing communicative intent from 2D body pose, targeting real-time person-to-robot communication in resource-constrained environments like rescue missions. The authors compare their dataset against existing real (IPC) and synthetic (MotionLCM, VEO3.1, Kimodo) datasets, evaluating skeleton graph classifiers and joint motion-forecasting networks on accuracy and frame rate (NVIDIA Orin Nano). They propose an autoregressive self-consistency measure as an unsupervised reliability signal, providing theoretical bounds on prediction correctness and demonstrating its growth with consistent steps.
2d body posecommunicative intentskeleton graph classifiersautoregressive self-consistencyreal-time robotics
Reasoning Arena: Trace Tournaments When Verifiable Rewards Fall Short
The paper introduces Reasoning Arena, an adaptive training framework that addresses the limitation of uninformative verifiable rewards in reinforcement learning with verifiable rewards (RLVR) for large language models. When reward signals become uniform within a group of reasoning traces, the method constructs trace tournaments where traces are compared head-to-head to extract finer-grained preferences. A Bradley-Terry model is fitted on incomplete comparison graphs to enable scalable reward estimation without exhaustive pairwise comparisons. Empirical results show a 7.6% average improvement over RLVR baselines in mathematics and coding benchmarks, with 27-41% faster training and 50% reduced generation compute.
reinforcement learningverifiable rewardstrace tournamentsbradley-terry modelreasoning quality
Scaling Neural Network Verification with Tensor Parallelism and Fully Sharded Data Parallelism
The authors scale neural network verification by adapting tensor parallelism (TP) and fully sharded data parallelism (FSDP) to the auto_LiRPA/α,β-CROWN framework. TP shards weight and A-matrices across GPUs, achieving ≈2× peak-memory reduction at P=2, while FSDP shards only weight matrices with per-layer AllGather, producing bitwise identical bounds to single-GPU baselines. FSDP reduces baseline memory by 80-90% and peak memory by 34-39% on wide MLPs, integrates with complete verification and convolutional layers, and obtains an unsat result for CIFAR-100 ResNet-large. Experiments reveal alpha tensors as the memory bottleneck in α-CROWN+Branch-and-Bound mode.
tensor parallelismfully sharded data parallelismauto_lirpabranch-and-boundalpha-crown
Capability-Aligned Hierarchical Learning for Tool-Augmented LLMs
The paper introduces Capability-Aligned Hierarchical Learning (CAHL), a method that jointly optimizes high-level planning and low-level tool-execution policies in tool-augmented LLMs using RLVR to address planner-executor misalignment. Unlike prior hierarchical approaches that optimize policies separately, CAHL enables better coordination between task decomposition and tool invocation. Experiments on constrained benchmarks (API-Bank, BFCL) and open-ended environments (Bamboogle) validate its effectiveness in improving tool-use performance.
tool-augmented llmshierarchical learningplanner-executor alignmentrlvrtool-use benchmarks
PhysScene: A Scene Graph Dataset for Scientific Visual Reasoning in Physics Experiments
The authors introduce PhysScene, the first scene graph dataset specifically designed for physics experiments, addressing the lack of domain-specific benchmarks for scientific visual reasoning. The dataset models specialized instruments, structured setups, and functional relations in experimental environments, emphasizing semantic constraints and high relation density over scale. Evaluations demonstrate that PhysScene complements existing benchmarks and provides a challenging testbed for advancing scene parsing algorithms in scientific contexts.
scene graphvisual reasoningphysics experimentsrelational reasoningsemantic constraints
Experience Makes Skillful: Enabling Generalizable Medical Agent Reasoning via Self-Evolving Skill Memory
The paper introduces SkeMex, a self-evolution framework for medical agents that distills interaction trajectories into structured skills for reusable procedural knowledge. It organizes these skills into a multi-branch repository and employs context-dependent utility estimation to guide retrieval and governance. The framework operates via a closed-loop lifecycle for continual evolution. Experiments demonstrate SkeMex's superiority over memory-based agents in offline and online clinical tasks, with generalizability across model backbones and transferable skill memory.
medical agentself-evolutionskill memorycontext-dependent utilityclosed-loop lifecycle
Beyond Humans: Multispecies Animal Face Recognition Using Transfer Learning
This study investigates transfer learning for multispecies animal face recognition, addressing limitations of physical identification methods and small-scale datasets. The authors evaluate FaceNet (pre-trained on human faces) and Vision Transformer (ViT, pre-trained on ImageNet) across three animal datasets: dogs, primates (lemurs, golden monkeys, chimpanzees), and cattle. Experiments compare verification accuracy and Rank-1 Identification Rate against state-of-the-art models trained specifically for each species. Results show ViT achieves 96.85% verification accuracy and 84.34% Rank-1 Identification Rate for dogs, outperforming FaceNet. Performance varies across species, with ViT surpassing state-of-the-art for cattle but showing mixed results for primates, highlighting task-dependent effectiveness across animal classes.
transfer learningface recognitionvision transformerverification accuracyrank-1 identification rate
Leveraging Structural Constraints for Diffusion-based Neural TSP Solvers
We introduce Projected Consistency Inference (PCI), a structure-aware inference method for neural Traveling Salesman Problem (TSP) solvers that replaces gradient refinement with Hamiltonian tour decoding and local search. PCI leverages the discrete structure of feasible solutions through lightweight projections and 2-opt operations, requiring no retraining. On TSP instances with 500 and 1000 cities, PCI achieves optimality gaps of 0.17% and 0.31%, respectively, outperforming FT2T while reducing inference time by 30-40%. PCI also demonstrates lower variance, memory usage, and competitive performance against classical heuristics like LKH3 in rapid solution generation.
projected consistency inferencetraveling salesman problemhamiltonian tour2-optneural combinatorial optimization
Conan-embedding-v3: Fusing Modality-Specific Models for Omni-Modal Embedding
Conan-embedding-v3 introduces a decouple--fuse--recover framework for omni-modal retrieval, addressing challenges in unifying text, image, video, document, and audio modalities. The method employs Decoupled Specialist Fusion, training modality-specific models independently and fusing their task vectors into a single dense backbone. A failure mode, Projector Drift, is identified when attaching audio via an external encoder and projector, causing retrieval regression. Projector Recovery mitigates this through full-parameter fine-tuning of the projector while freezing the backbone, followed by balanced multi-modal rehearsal. The model achieves 74.9 on MMEB and 55.61 on the 30-task MAEB audio suite.
omni-modal retrievaldecoupled specialist fusionprojector driftprojector recoverymulti-modal rehearsal
A Universal Dense Football Event Representation Based on TabTransformer
The paper introduces a Transformer-based model for learning dense representations of heterogeneous football event data, combining continuous spatial coordinates with categorical action descriptors. The proposed TabTransformer architecture encodes categorical features (action type, outcome, body part) as learned embeddings, capturing sport-specific semantics through self-attention mechanisms. Evaluated on downstream tasks including action value estimation and play style recognition, the method demonstrates superior probability calibration (measured by Brier score) compared to task-specific baselines using traditional one-hot or ordinal encodings.
tabtransformerdense representationself-attentioncategorical embeddingbrier score
TRL-Bench: Standardizing Cross-Paradigm Representation-Level Evaluation of Tabular Encoders
TRL-Bench introduces a standardized benchmark for cross-paradigm representation-level evaluation of tabular encoders, addressing the challenge of comparing models trained under different paradigms. The benchmark employs three evaluation suites—TRL-CTbench, TRL-Rbench, and TRL-DLTE—to probe row-, column-, and table-level embeddings across 20 models and 16 tasks. Results demonstrate that encoder performance is capability-specific rather than captured by a single leaderboard, with generic text encoders excelling on tasks with strong surface-text signals and tabular specialists performing better when pretraining aligns with task objectives. TRL-Bench provides curated assets, including 50 OpenML tables and a 47,772-table DLTE lake, enabling reproducible evaluation of tabular representations.
tabular encodersrepresentation-level evaluationtrl-benchdata-lake table enrichmentcross-paradigm comparison
Anything2Skill: Compiling External Knowledge into Reusable Skills for Agents
The paper introduces Anything2Skill, a framework for compiling external knowledge into reusable procedural skills for agents. It decomposes knowledge records into evidence windows, extracts skills via plan-and-expand under a skill-tree prior, and structures them as contracts with invocation conditions, workflows, and constraints. These skills are managed in a SkillBank with taxonomy-aware compilation and versioning. Combined with retrieval-augmented generation (RAG), the method achieves 98.85% and 94.10% success rates on qsv and GitHub-CLI benchmarks, outperforming RAG-only approaches by enabling both declarative knowledge access and procedural skill reuse.
retrieval-augmented generationskill-tree priorprocedural memoryskillbanktaxonomy-aware compilation
Brain-Prompt Injection: A Route-Safety Audit for BCI-LLM Agents
The paper introduces brain-prompt injection as a novel attack surface in BCI-to-agent pipelines, where neural activity decoding exposes vulnerabilities to signal-side perturbations, context-only injections, and adaptive dual-decoder attacks. The authors propose a Route-Safety Audit Contract, comprising a minimal log schema, denominator hierarchy, and endpoint specification, and prove an audit-schema separation theorem alongside a C3 attacked-dependence decomposition. Split-conformal calibration is applied to a non-oracle EEG confirmation channel, revealing a false-accept frontier under explicit threat archetypes. Experiments on EEGMMI datasets with 5,400 events demonstrate that provenance blocks certain routes, while agreement-plus-provenance routes others, with conformal calibration achieving FAR 0.000 at clean utility 0.150 for α=.005.
brain-prompt injectionroute-safety auditsplit-conformal calibrationeeg confirmation channelc3 attacked-dependence
FF-JEPA: Long-Horizon Planning in World Models with Latent Planners
FF-JEPA introduces a hierarchical world modeling approach for long-horizon planning without requiring goal images, addressing computational inefficiency in Joint Embedding Predictive Architectures (JEPAs). The method employs two forward dynamics models: an action-conditioned forward model and an action-free latent planner that predicts subgoals sequentially. This decomposition enables tractable optimization by breaking complex trajectories into short-term problems. Preliminary PushT experiments demonstrate FF-JEPA mitigates long-horizon collapse in flat world models, validating its efficacy for goal-free planning.
joint embedding predictive architectureslong-horizon planninglatent plannersforward dynamics modelssubgoal prediction
Internalizing Geometric Law: Learning from Solver Residuals for Precision-Critical Generation
The paper introduces PyGeoX, a differentiable geometric domain-specific language (DSL) for precision-critical generation tasks, addressing hallucination in Large Language Models (LLMs) when producing geometrically constrained outputs. The authors identify Outlier Gradient Masking as a key failure mode in constraint satisfaction and propose Saturating Additive Rewards (SAR) to decompose rewards into bounded per-constraint terms, preserving gradients under severe violations. Evaluated on PyGeoX-Bench (300 problems), SAR improves hard-tier solving rates by 2.3× over MSE-based rewards, with an 8B model matching larger frontier systems.
geometric constraintsdifferentiable lossoutlier gradient maskingsaturating additive rewardsprecision-critical generation
Physics-Guided Sequence-Based Generative Framework for Acoustic Metamaterial Inverse Design
MetaSeq introduces a sequence-based generative framework for inverse design of acoustic metamaterials (AMMs), addressing broadband target responses and acoustic dispersion challenges. The method represents AMMs as structured sequences, preserving geometric precision and connectivity, and formulates inverse design as a sequence-to-sequence task. It combines supervised pretraining with reinforcement learning fine-tuning, guided by physics-based solvers and validity checkers. Evaluations against COMSOL and five baselines demonstrate a 45% reduction in response error compared to the best baseline.
acoustic metamaterialsinverse designsequence-based generationphysics-guided learningbroadband response
BSTabDiff: Block-Subunit Diffusion Priors for High-Dimensional Tabular Data Generation
BSTabDiff introduces a block-subunit generative framework for high-dimensional low-sample size (HDLSS) tabular data, addressing challenges like local correlations, sparse dependencies, and non-Gaussian marginals. The method partitions features into latent blocks, learning dependencies in a compact space via shared subunit variables, and employs copula-driven decoding with flexible marginals. It supports diffusion and normalizing flow priors for stable synthesis. Empirical results show BSTabDiff outperforms unstructured generators in realism and stability for HDLSS data.
hdlssdiffusion priorscopula-driven decodingnormalizing flowsblock-subunit
Proposal Refinement for Few-Shot Object Detection
The paper proposes a proposal refinement approach to address unbalanced region proposal distribution in few-shot object detection. The method introduces refinement loss during base training to improve novel class sensitivity and adds a refinement branch to RPN during fine-tuning to boost novel proposal generation. Evaluations show 1-6% performance gains over baselines on standard benchmarks without inference overhead, establishing new state-of-the-art results.
few-shot object detectionregion proposal networksrefinement lossproposal distributionfine-tuning
EgoTactile: Learning Grasp Pressure for Everyday Objects from Egocentric Video
The paper introduces EgoTactile, a benchmark combining egocentric video with full-hand pressure supervision for diverse everyday objects, including a bare-hand transfer subset for generalization. The authors propose EgoPressureDiff, a conditional diffusion framework that adapts a pre-trained video diffusion backbone, enhanced by a Physically-Informed Feature Rectification layer to incorporate semantic constraints and resolve visual-physical ambiguities. Experiments show superior performance on the benchmark and robust transferability to in-the-wild scenarios.
egocentric videograsp pressurediffusion frameworktactile sensingsemantic constraints
Self-Paced Curriculum Reinforcement Learning for Autonomous Superbike Racing in Simulation
The paper introduces a Self-Paced curriculum Deep reinforcement Learning (SPDL) framework for autonomous superbike racing, addressing the unique challenges of two-wheeled dynamics. The method combines Soft Actor-Critic (SAC) with automated curriculum generation, using a state space that incorporates lean-angle history and global track features, alongside a stability-focused reward function. Experiments in the VRider SBK simulator show SPDL outperforms vanilla SAC in training efficiency, lap times (quantitative results unspecified), and stability across multiple tracks and bike models.
self-paced learningsoft actor-criticautonomous racingreinforcement learningphysics simulation
End-to-End Training for Discrete Token LLM based TTS System
The paper introduces an end-to-end (E2E) training framework for discrete token-based TTS systems, unifying speech tokenizer, LLM, flow-matching (FM) model, and reward model (RM) optimization. The method jointly trains these components using multi-task objectives (reconstruction, next-token prediction, recognition) to improve token space quality and reduce inference mismatch. Results show SOTA performance on Seed-TTS-Eval (0.78%/1.56% WER) with a 0.6B-parameter LLM and 0.5B-parameter FM model, demonstrating E2E optimization's superiority over cascaded pipelines.
text-to-speechend-to-end trainingdiscrete tokensflow-matchingreward model
Trustworthy Smart Fabs via Professional Proxies: Scaling Safe and Sustainable by Design (SSbD) through Industrial Data Spaces
The paper introduces a zero-trust socio-technical orchestration framework to address governance bottlenecks in semiconductor manufacturing under the EU's Safe and Sustainable by Design (SSbD) framework. The framework employs 'Professional Proxies'—role-based agentic workflows executed within hardware-isolated trust zones—to automate compliance reporting. It integrates Virtual Metrology (VM) predictions and Federated Machine Learning (FML) within Trusted Execution Environments (TEEs) to ensure data sovereignty. The architecture enables cryptographic signing of compliance tokens via International Data Spaces (IDS) connectors without exposing proprietary data. This approach provides a verifiable pathway for achieving net-zero Industry 5.0 ecosystems while maintaining multi-stakeholder transparency and corporate data privacy.
zero-trust architecturevirtual metrologyfederated machine learningtrusted execution environmentsinternational data spaces
Resource-aware Computation-Communication Overlap for multi-GPU ML Workloads
The paper presents a resource-aware method for overlapping computation and communication in multi-GPU ML training, reducing execution time by up to 25.5%. The approach employs two runtime controls: shared-memory-driven occupancy shaping to regulate computation-kernel residency and elevated scheduling priority for communication kernels, enabling concurrent execution without modifying vendor libraries. Evaluations on NVIDIA A40, A100, H100, and AMD MI250X GPUs demonstrate effective overlap, addressing the communication bottleneck in distributed training.
multi-gpu trainingcomputation-communication overlapshared-memory allocationscheduling prioritycollective communication
MASS: Deep Research for Social Sciences with Memory-Augmented Social Simulation
The paper introduces Memory-Augmented Social Simulation (MASS), a novel paradigm enhancing LLM-generated social science research through realistic simulations. MASS combines dynamic goal-path planning with multi-level social norm constraints, a multi-disciplinary behavior dataset for agent memory initialization, and a structured forgetting mechanism based on the Ebbinghaus curve. Experiments show MASS improves generation quality by 6.81% over foundation LLMs and achieves a 17.19% gain in Insight metrics compared to baselines.
memory-augmented social simulationdynamic goal-path planningmulti-level social normebbinghaus curvestructured forgetting mechanism
Pretrained, Frozen, Still Leaking: Auditing Cross-Encoder Attribute Transfer in EEG Foundation Models
The contribution is a joint audit framework for EEG foundation models that detects spectral attribute leakage across pretrained frozen encoders, overcoming limitations of single-endpoint audits. The method introduces a cross-encoder transfer audit using ridge attribute decoders and linear bridges between BIOT, LaBraM, and EEGPT models, supported by theoretical guarantees on encoder overlap. Results show significant attribute transfer (CI lower bound ≥0.081) across all encoder pairs, with audit-endpoint disagreement scores positive (p<0.001) across eight datasets. Standard defenses (Wiener-style noise, LiRA, DP-SGD) fail to prevent leakage, demonstrating the framework's necessity for deployment decisions.
eeg foundation modelscross-encoder transferridge attribute decoderaudit-endpoint disagreementspectral attribute leakage
Culturally-Adapted Red-Teaming Across East and Southeast Asian Contexts: A Methodological and Comparative Analysis
The study demonstrates that culturally-adapted (CA) red-teaming outperforms direct translation (DT) for multilingual safety evaluation of large language models (LLMs) across four East/Southeast Asian languages. Using paired DT and CA datasets (Korean, Japanese, Thai, Khmer) with 1:1 seed matching, they measure Attack Success Rate (ASR) and Cultural Realism scores across four open-source LLMs. CA prompts show mean ASR increase of +9.3 percentage points over DT, with DT underestimating risk in 44/48 category-language combinations. Cultural Realism analysis reveals DT scores consistently below 1.0/3.0 (mean 0.17) versus CA scores up to 2.51, proving cultural adaptation essential for valid safety evaluation.
culturally-adapted red-teamingattack success ratecultural realismmultilingual safety evaluationdirect translation
CANS: Accelerating Multiuser Collaborative Edge Inference via Cooperative Autodidactic NeuroSurgeon
The paper introduces CANS, a framework for accelerating multi-user collaborative edge DNN inference by adaptively learning optimal model partitions. It employs FedLinUCB-DW, a novel algorithm that groups devices by type and leverages offline inference experience to warm-start online exploration, with theoretical regret bounds provided. Evaluations on simulated and hardware prototype systems show CANS reduces average inference latency by up to 50% compared to non-cooperative baselines.
edge computingdnn partitioningfedlinucb-dwregret boundlatency reduction
IMUG-Bench: Benchmarking Unified Multimodal Models on Interleaved Understanding and Generation
The paper introduces IMUG-Bench, a novel benchmark for evaluating unified multimodal models (UMMs) on multi-turn interleaved image-text dialogue tasks. It addresses limitations of existing benchmarks by incorporating 3,113 samples and 12,034 interaction turns across three classes: Static Spatial, Temporal Causal, and Hybrid. The benchmark reveals exposure bias in generation tasks and tests strategies like Chain-of-Thought and Self-Verification to mitigate it. Large-scale experiments demonstrate these methods improve generation accuracy, providing insights for enhancing UMM robustness in multi-turn interactions.
multimodalbenchmarkexposure biaschain-of-thoughtself-verification
Reliable to Expressive: A Curriculum for Rubric-Following Safety Judges
The paper proposes a curriculum training strategy for robust safety judges that consistently follow evaluation rubrics across varying formulations. The method combines instance-conditioned dynamic rubrics with a reliable-to-expressive curriculum, transitioning from fixed-rubric supervision to noisier dynamic-rubric data. Evaluated on three rubric prompts, the 12B curriculum judge achieves 94.12-94.88% accuracy with minimal cross-rubric variance (0.76), outperforming baselines including 30B models. An ablation study confirms the curriculum's necessity, as naive dynamic rubric mixing increases variance from 1.44 to 3.60.
safety judgesrubric-followingdynamic rubricscurriculum learningmeta-evaluation
Crop Recommendation and Agricultural Query Answering System Using Spatio-Temporal Graph Neural Networks and Hybrid Retrieval Augmentation
The authors present a unified precision agriculture system integrating weather prediction, crop recommendation, and agricultural question answering. They develop two deep learning models: a Transformer-based Graph Neural Network and a Spatio-Temporal Graph Convolutional Network (STGCN), which achieves superior weather forecasting accuracy (MSE ~0.011) over 30 days using data from 1,359 Nepalese locations. The STGCN outputs are combined with soil properties to generate localized crop recommendations via a scoring algorithm. A Retrieval-Augmented Generation chatbot provides agricultural Q&A by leveraging domain-specific documents. Deployed via mobile app, the system demonstrates usability and relevance in rural settings through user feedback.
spatio-temporal graph convolutional networkretrieval-augmented generationprecision agriculturegraph neural networkweather prediction
Unified Energy for Invariant and Independent Decoding in Diffusion Language Models
The paper introduces Unified Energy (Uni-E), a method to address performance gaps in Diffusion Language Models (DLMs) by jointly modeling token relationships through invariant (Inv-E) and independent (Ind-E) energy terms. Uni-E eliminates sampling-based partition estimation, handles distribution shifts from dependency and invariance issues, and scales to arbitrary model sizes. Theoretical analysis shows Uni-E corrects distribution shifts, while experiments on DLMs and Diffusion Large Language Models (DLLMs) validate its effectiveness in improving parallel text generation quality.
diffusion language modelsinvariant energyindependent energyparallel text generationdistribution shift
SEF-CLGC at SemEval-2026 Task 11: Logical Notation Impact on Language Model Performance
The paper introduces SEF-CLGC, a pipeline combining formal logical notations with Small Language Models (SLMs) to evaluate reasoning performance on SemEval-2026 Task 11 Subtask 1. The method trains SLMs on hybrid natural-symbolic language data to disentangle content from formal reasoning. Results show the best model achieves 27.80% content score while significantly reducing content bias in reasoning tasks.
sef-clgcsmall language modelslogical notationscontent biassemeval-2026
Decoding Pedestrian Crossing Intention from Egocentric Vision via Vision Language Models
The work introduces a vision-language model (VLM) approach for decoding pedestrian crossing intentions from egocentric videos, framed as a visual question answering task. Three VLM families were benchmarked in zero-shot settings, showing moderate gains over random chance but limited traffic reasoning. Parameter-efficient fine-tuning improved performance by 9% over a transformer baseline, with additional gains (14.5%) achieved by incorporating contextual cues like ego motion and eye gaze in the Qwen3-VL-2B model.
egocentric visionvision-language modelspedestrian intent decodingparameter-efficient fine-tuningvisual question answering
Steganography Without Modification: Hidden Communication via LLM Seeds
The paper identifies a steganographic channel in deterministic LLM decoding that requires no model modifications, exploiting PRNG seed-dependent token probability intervals for covert communication. Senders encode messages in PRNG seeds; receivers reconstruct intervals from generated text to recover seeds via exhaustive search, with two operational modes: known-prompt (perfect recovery via forced alignment) and unknown-prompt (approximate reconstruction with maximum-hit-count scoring). Experiments across six model families show 100% 32-bit seed recovery accuracy within 300 tokens (35s) for known prompts, and near-perfect accuracy at 600-800 tokens (12s) for unknown prompts, demonstrating prompt ignorance is not a security guarantee.
steganographyllm inferenceprng seeddeterministic decodingtoken probability intervals
From USD Scenes to Knowledge Graphs: Zero-Shot Ontology Grounding with LLMs
The study demonstrates that large language models (LLMs) can effectively automate ontology grounding in Universal Scene Description (USD) scenes for robot task reasoning, eliminating the need for manually curated dictionaries. Using zero-shot, training-free methods, LLMs achieved 90-96% exact-match accuracy with descriptive object names and 49-89% with abbreviated names on a kitchen scene containing 125 objects, significantly outperforming dictionary and embedding baselines. Context-augmented prompting recovered up to 48% accuracy under fully opaque names. Feature ablation revealed that LLMs primarily rely on semantic cues in the scene graph, with anonymization reducing accuracy to 0-6%, while geometry alone yielded only 4-17% accuracy.
ontology groundinguniversal scene descriptionzero-shot learningscene graphlarge language models
Vision Language Model Helps Private Information De-Identification in Vision Data
The paper introduces VisShield, an end-to-end framework for enhancing privacy awareness in Vision Language Models (VLMs) by localizing and masking sensitive text in visual inputs. The method combines OPTIC, a specialized instruction-tuning dataset with privacy-oriented prompts, and a tailored training strategy to adapt VLMs for precise Optical Character Recognition (OCR) and bounding box output. Experiments show VisShield outperforms existing approaches in handling private information, particularly Protected Health Information (PHI) in medical images.
vision language modelsprivacy protectionoptical character recognitioninstruction-tuningbounding box localization
Late-Layer Fusion is Enough: Dual-Path Vision Token Routing for Multimodal Large Language Models under Visual Saturation
The paper proposes Dual-Path Vision Token Routing (DPVR), a modality-asymmetric framework for efficient multimodal large language models (MLLMs). Through layer-wise analysis of LLaVA-1.5, the authors observe vision tokens saturate by middle layers (text-to-image attention drops to 0.04 post-layer 18), motivating DPVR-LF which routes vision tokens to a shallow side branch after saturation, skipping deep text-only processing. With 3% trainable parameters, DPVR-LF maintains competitive performance while reducing redundant visual computation, challenging the need for full-depth vision token processing in Transformer-based MLLMs.
multimodal large language modelstoken routingvisual saturationmodality asymmetrylate-layer fusion
Unveiling Privacy Risks in Multi-modal Large Language Models: Task-specific Vulnerabilities and Mitigation Challenges
The paper introduces MM-Privacy, a dataset for evaluating privacy risks in Multi-modal Large Language Models (MLLMs), defining Disclosure Risks and Retention Risks. It systematically assesses MLLMs, revealing their susceptibility to leaking sensitive data from images across various tasks, and examines task inconsistency's role in privacy vulnerabilities. Results demonstrate significant privacy concerns, highlighting the need for mitigation strategies. The dataset and code are publicly available.
multi-modal large language modelsprivacy risksdisclosure risksretention riskstask inconsistency
A Regret Minimization Framework on Preference Learning in Large Language Models
The paper introduces Regret-based Preference Optimization (RePO), a novel framework for reinforcement learning from human feedback (RLHF) that models preferences through regret minimization rather than reward maximization. RePO accounts for human feedback's prospective and counterfactual nature by treating preferences as behavior-conditioned assessments of relative suboptimality. Evaluations on mathematical reasoning benchmarks and human preference datasets show consistent performance improvements, demonstrating RePO's effectiveness as a human-aligned training approach for large language models.
regret minimizationreinforcement learning from human feedbackpreference optimizationcounterfactual comparisonsbehavior-conditioned assessments
An Enhanced Geometric-Spectral Feature Learning Framework for Airborne Multispectral Point Cloud Classification
The authors propose an enhanced geometric-spectral feature learning framework for airborne multispectral point cloud (MPC) classification, addressing challenges of high-dimensional heterogeneity, unbalanced samples, and inter-class spectral similarity. The method employs a two-stream architecture with attention mechanisms: one stream extracts position-encoded global spectral features via fusion self-attention, while the other uses multikernel point convolution and feature aggregation attention for spectral-guided geometric features. A residual attention fusion block integrates these features, complemented by a joint loss function for improved learning. Experiments on two novel MPC datasets show superior performance over state-of-the-art methods, with code and data publicly released.
multispectral point cloudattention mechanismsfeature fusionmultikernel convolutionjoint loss function
Autonomous Incident Resolution at Hyperscale: An Agentic AI Architecture for Network Operations
The paper introduces an agentic AI architecture for autonomous incident resolution in hyperscale cloud networks, addressing challenges of volume, velocity, and complexity. The system employs a multi-agent orchestration framework where specialized agents collaborate to detect, diagnose, and remediate network incidents without human intervention. Key architectural principles include hierarchical agent decomposition, skills-based tool invocation, structured knowledge encoding, progressive autonomy, and closed-loop verification. Deployed in production at a major cloud provider, the system achieves over 90% autonomous resolution rates for common incidents, ensuring safety through layered authorization and rollback mechanisms. Design tradeoffs, failure modes, and operational insights are discussed.
multi-agent orchestrationhierarchical decompositionskills-based invocationprogressive autonomyclosed-loop verification
ComplexConstraints and Beyond: Expert Rubrics for RLVR
The paper introduces expert-curated rubric-based evaluation as an advanced paradigm for assessing LLMs, addressing limitations of traditional benchmarks. It proposes five design principles for rubric construction, including Maximum Viable Atomicity and iterative LLM-judge calibration, validated through the ComplexConstraints dataset (1,000 examples with 10-40 atomic criteria per prompt). Results show rubric-based training improves instruction-following by +15.5% (4B) and +12.2% (235B), while RL training on rubric-graded environments yields transfer gains (+4.5% BFCL, +7.4% Tau2-Bench, +6.8% Tool-Decathlon).
rubric-based evaluationinstruction followingllm-judge calibrationcomplexconstraintsrl training
Optimizing Energy-based Neural Network Training with Coherent Ising Machine
The authors propose a Coherent Ising Machine (CIM)-based method for training energy-based neural networks via Equilibrium Propagation, achieving performance parity with software implementations. They integrate the Adam optimizer to solve Hopfield energy network ground states, improving convergence speed and solution accuracy. The approach demonstrates scalability across deeper architectures and convolutional operations, suggesting potential for energy-efficient analog, optoelectronic, or photonic implementations. Results indicate CIM dynamics as a viable platform for complex neural network training.
coherent ising machineequilibrium propagationhopfield networkadam optimizerenergy-based training
Hybridizing Equilibrium Propagation with Ising Machines for Efficient Energy-Based Learning
We introduce an Ising-dynamics-inspired equilibrium-propagation framework that replaces dissipative Hopfield relaxation with extended phase-space dynamics using conjugate variables. This hybrid approach maintains EP's local two-phase learning rule while altering the physical route to equilibrium, lowering effective energy barriers and improving convergence speed and noise robustness. The method trains deep convolutional Hopfield networks on MNIST, FashionMNIST, and CIFAR-10 with performance comparable to backpropagation, addressing EP's local minima convergence issue due to phase-space contraction.
equilibrium propagationising dynamicshopfield relaxationphase-space contractionenergy-based learning
Graph2Idea:Retrieval-Augmented Scientific Idea Generation with Graph-Structured Contexts
Graph2Idea introduces a knowledge graph-guided framework for retrieval-augmented scientific idea generation, addressing limitations of flat-text contexts in LLM-based methods. The approach retrieves relevant papers, converts them into structured knowledge triples, and constructs a target-centered knowledge graph to make cross-paper relations explicit. A two-stage generation process then identifies research directions and synthesizes ideas from graph-grounded evidence. Experiments on a scientific idea generation benchmark show improvements over baselines, with Novelty increasing from 0.45 to 0.52, Quality from 0.24 to 0.29, and Feasibility from 0.22 to 0.28.
retrieval-augmented generationknowledge graphscientific idea generationlarge language modelsstructured knowledge triples
Addressing Market Regime Changes and Heavy-Tailed Returns in Portfolio Optimization via Bayesian VAR and Elliptical Black-Litterman
The paper introduces BAVAR-BLED, a novel DRL framework for portfolio optimization that addresses fat-tailed returns and regime changes. The method combines Bayesian-Averaging Vector Autoregressive (BAVAR) models for regime-aware temporal feature extraction with a Black-Litterman model using Elliptical Distributions (BLED) for fat-tailed return estimation. Transformer networks construct market views, while CNNs estimate risk aversion. Evaluated on 29 Dow Jones stocks over a decade, BAVAR-BLED achieves Sharpe and Sortino ratios of 1.72 and 2.70, respectively, with 57.26% total returns, outperforming state-of-the-art methods.
portfolio optimizationbayesian varelliptical distributionsfat-tailed returnsregime changes
Context Rot in AI-Assisted Software Development: Repurposing Documentation Consistency for AI Configuration Artifacts
The paper introduces 'context rot' as a novel phenomenon in AI-assisted software development, where persistent configuration files (e.g., CLAUDE.md, AGENTS.md) guiding AI tools become stale as code evolves. It proposes repurposing traditional documentation consistency tools (e.g., README/wiki checkers) to detect such rot, establishing a research roadmap for adaptation. Preliminary evaluation on 356 repositories reveals 23.0% contain stale code references, validating existing tools' applicability to this new problem domain.
context rotai configuration artifactsdocumentation consistencypersistent contextcode evolution
DynaOD: Dynamic Origin-Destination Flow Generation with Discrete-to-Continuous Temporal Semantic Modeling
DynaOD introduces a semantic-driven framework for dynamic origin-destination (OD) flow generation, synthesizing mobility dynamics from temporal context without historical OD data. The method jointly models discrete directional trends (qualitative urban activity shifts) and continuous temporal evolution, constructing time-varying region representations that condition pretrained static OD generators. This modular design enables scalable deployment and cross-city transferability. Evaluations on large-scale real-world datasets demonstrate superior performance over baselines in predictive accuracy and distributional fidelity.
origin-destination flowtemporal semanticsurban mobilitymodular designcross-city transferability
Context-Fractured Decomposition Attacks on Tool-Using LLM Agents: Exploiting Artifact Provenance Gaps
The paper introduces Context-Fractured Decomposition (CFD), a novel attack vector for tool-using LLM agents that exploits provenance gaps in artifact-mediated workflows. By decomposing harmful intents across multiple benign-looking interactions and leveraging untracked artifact reuse, CFD achieves up to 28.3 percentage point higher jailbreak success rates compared to state-of-the-art baselines, even against robust single-turn defenses. The authors propose trace-level diagnostics and provenance lineage tagging as verifiable mitigation strategies, validated across standard agent-system jailbreak benchmarks.
provenance gapcontext-fractured decompositionartifact-mediated compositiontool-using llm agentsjailbreak defenses
FlashMemory-DeepSeek-V4: Lightning Index Ultra-Long Context via Lookahead Sparse Attention
FlashMemory-DeepSeek-V4 introduces Lookahead Sparse Attention (LSA), a novel inference paradigm addressing GPU memory bottlenecks in ultra-long context serving for LLMs. LSA employs a Neural Memory Indexer to predict future context demands, retaining only query-critical KV chunks in GPU memory via a backbone-free decoupled training strategy. This approach formulates the indexer as a dual-encoder architecture, trained independently using standard retrieval frameworks. Evaluated on LongBench-v2, LongMemEval, and RULER, FM-DS-V4 reduces the physical KV cache footprint to 13.5% of the full-context baseline while improving downstream accuracy by +0.6%. At 500K context scales, it suppresses KV cache overhead by over 90% without compromising reasoning capacities.
lookahead sparse attentionkv cacheneural memory indexerdual-encoder architectureultra-long context
A Unifying Lens on Reward Uncertainty in RLHF
(No summary returned.)
REFLECT: Intervention-Supported Error Attribution for Silent Failures in LLM Agent Traces
REFLECT introduces an intervention-supported error attribution method for diagnosing silent failures in LLM agent traces. The approach diagnoses candidate error steps, tests them via controlled replay with diagnosis-specific patches, and uses verified outcome flips as contrastive evidence to refine attribution. Evaluated on four multi-hop reasoning benchmarks, REFLECT achieves the highest localization accuracy among same-auditor methods, with particularly strong gains on structured tool-use traces, while functioning without ground-truth answers.
llm agentserror localizationsilent failuresintervention-supported attributionmulti-hop reasoning
OnlyDense: Reduced-Order Modeling for Lagrangian simulation
The paper introduces OnlyDense, a reduced-order modeling framework for Lagrangian simulations that treats system states as functions evolving in Hilbert space. Unlike graph-based or nonlinear latent space approaches, it approximates the state space via a linear subspace spanned by learned neural basis functions, enabling direct projection to latent coefficients and explicit basis access. This method combines classical projection-based reduced-order modeling with deep learning while maintaining invariance to discretization points. Evaluated on large-scale SPH simulations (>1M particles) with extreme deformation/fragmentation, it achieves R²>0.99 using just 32 basis functions.
reduced-order modelinglagrangian simulationhilbert spaceneural basis functionssmooth particle hydrodynamics
See More, Think Deeper: Query-Expanded Visual Evidence and Answer-Clue Guided Reflection for Long Video Understanding
The paper introduces CoVER, a framework enhancing Video-LLMs for long-video understanding through query-expanded visual evidence acquisition and answer-clue guided reflection. CoVER addresses limitations in evidence diversity and visual feedback by dynamically gathering multi-intent visual evidence and verifying draft answers with visual cues. Evaluations show CoVER-7B outperforms same-scale models and rivals closed-source state-of-the-art on select metrics.
video-llmslong-video understandingquery-expanded evidencevisual feedbackanswer-clue reflection
Stage-1 Controls the Entropy Regime, Not the Outcome
The study characterizes the role of Stage-1 warm-start in two-stage post-training for vision-language models (VLMs), demonstrating its primary influence on entropy regime rather than final performance. Using Qwen2.5-VL-7B with a 72B VLM teacher for on-policy distillation (OPD), experiments show Stage-1 warm-starts converge to a narrow 53–54% band on Geometry3K, with minimal endpoint variation. Early-stopped supervised fine-tuning (SFT) improves out-of-domain MathVista by +2.1 points, while OPD exhibits higher policy entropy and answer diversity (pass@16 +2.0 to +5.2 points) pre-RL, though advantages diminish post-RL. Results suggest OPD is not a superior RL warm-start in this setup.
vision-language modelson-policy distillationsupervised fine-tuningentropy regimereinforcement learning
INFUSER: Influence-Guided Self-Evolution Improves Reasoning
INFUSER introduces an iterative co-training framework for self-evolving language models, featuring a Generator that drafts questions from unstructured documents and a Solver that improves via training on them. The generator uses an optimizer-aware influence score, optimized by DuGRPO (a dual-normalized GRPO variant), to prioritize questions beneficial to the solver's target distribution. On Qwen3-8B-Base, INFUSER achieves >20% relative improvement over baselines on Olympiad and SuperGPQA benchmarks, with an 8B generator outperforming a frozen 32B one in math and coding. Ablations validate design choices, and extensions demonstrate flexibility.
self-evolutioninfluence scoredugrpoco-trainingoptimizer-aware
BareWave: Waveform-Native Flow-Matching Text-to-Speech
BareWave introduces a waveform-native flow-matching framework for direct text-to-speech generation, eliminating intermediate acoustic representations. The method addresses three training challenges: lack of pretrained waveform scaffolds, noise schedule optimization, and perceptual-temporal alignment via velocity-aware perceptual alignment (VAPA). Experiments demonstrate strong zero-shot voice cloning performance in intelligibility, speaker similarity, and naturalness, validating waveform-native TTS as viable. Project demos are available at https://barewave.github.io/.
flow-matchingtext-to-speechwaveform-nativezero-shot cloningperceptual alignment
Agent Economics: An Entropy-Controlled Pluralistic Alignment Framework for Preventing Artificial Hivemind in Autonomous Agents
The study introduces the Behavioral Protocol Framework (BPF), an entropy-controlled pluralistic alignment framework addressing strategic convergence (hivemind effect) and decision-making opacity in autonomous agent economies. BPF integrates three modules: Mentalizing-based Social Intelligence (MbSI) for Theory of Mind reasoning, Pluralistic Alignment (PA) for strategic diversity preservation via entropy control, and a Verifiable Execution Kernel (VEK) for transparent auditing. Evaluation involves a Python/Streamlit simulation to empirically validate PA's entropy mechanisms and VEK's audit capabilities. Anticipated results suggest BPF enhances stability, efficiency, and trustworthiness in agent-native economic systems.
behavioral protocol frameworkpluralistic alignmentmentalizing-based social intelligenceverifiable execution kernelentropy-control
Personalization Meets Safety:Mechanisms,Risks,and Mitigations in Personalized LLMs
The paper presents the first systematic review of safety risks in personalized LLMs, analyzing mechanisms across user representation, personalization paradigms, and evaluation. It introduces a unified taxonomy of vulnerabilities in prompting, retrieval augmentation, fine-tuning, RL, MoE, pruning, agent frameworks, and multimodal approaches, with corresponding mitigation strategies. Analysis reveals three research gaps: user-invariant safety evaluation, isolated technique analysis, and inadequate long-term risk assessment frameworks, demonstrated through a case study of OpenClaw.
personalized llmssafety risksmixture-of-expertsretrieval augmentationagent frameworks
A Multi-Agent System for IPMSM Design Optimization via an FEA-AI Hybrid Approach
The paper proposes an automated multi-agent framework for optimizing interior permanent magnet synchronous motor (IPMSM) design, addressing bottlenecks in manual setup, high FEA costs, and unreliable surrogate-based search. The method integrates retrieval-augmented generation (RAG) for problem definition with an uncertainty-aware FEA-AI hybrid pipeline, employing specialized agents for design, training, sampling, and optimization with GA-based search and uncertainty-driven switching between AI-surrogate and FEA evaluation. Experimental results demonstrate superior objective performance and reduced predictive uncertainty compared to FEA-only and AI-only approaches under matched computational budgets.
ipmsm designretrieval-augmented generationfea-ai hybriduncertainty-aware optimizationmulti-agent system
TRIAGE: Dialectical Reasoning for Explainable Risk Prediction on Irregularly Sampled Medical Time Series with LLMs
TRIAGE introduces a dialectical reasoning framework for LLMs to improve risk prediction on irregularly sampled medical time series (ISMTS), addressing the issue of risk polarization in clinical early warning systems. The method trains an LLM to generate outcome-specific rationales, enabling continuous risk scores with explicit clinical reasoning. Evaluated on three ISMTS benchmarks, TRIAGE improves AUPRC by 3.3% on average, reduces calibration error by 81%, and enhances rationale quality by 20% compared to baselines.
irregularly sampled medical time seriesrisk polarizationdialectical reasoningclinical early warning systemsllm-as-a-judge
ATM: Action-Consistency Transfer Matrix for Diagnosing and Improving Latent World Models
The paper introduces ATM (Action-Consistency Transfer Matrix), a diagnostic tool for evaluating latent world models without costly planner-coupled simulator rollouts. ATM employs lightweight post-hoc probes to compare action semantics between real and predicted transitions, producing an interpretable matrix that reveals representation quality and transition inconsistencies. Results demonstrate 100x speedup over CEM-based evaluation while maintaining reliable model ranking, alongside AITS, a method leveraging action-identifiability as a training signal for improved planning performance.
latent world modelsaction-consistencytransition analysisplanner-coupled evaluationaction-identifiability
SafeRun: Enabling Determinism in LLM Planning for Running
SafeRun introduces a decoupled architecture for deterministic LLM-based running planning, separating soft LLM interpretation from hard constraint enforcement by a solver. The framework ensures strict safety constraints while maintaining natural-language flexibility, addressing reliability issues in determinism-critical domains. Evaluated across five LLMs, SafeRun achieves 100% safety score (vs. 79.1% PE and 97.6% CodeAct averages) on a novel benchmark for running planning under physiological and safety constraints. The benchmark is publicly available.
llmdeterminismsafetyplanningbenchmark
TLDR: Compressing Audio Tokens for Efficient Autoregressive Text-to-Speech
TLDR introduces a patch-based autoregressive framework to accelerate codec-based AR-TTS by shifting causal modeling from token-level to patch-level sequences. The method groups consecutive codec tokens into latent patches using a lightweight compressor, models the shorter sequence with a frozen pretrained AR-TTS backbone adapted via LoRA, and reconstructs fine-grained tokens with a speaker-conditioned extractor. With a patch size of 4, TLDR achieves 1.8x inference speedup and 75% KV-cache memory reduction while maintaining performance.
autoregressive ttscodec tokenskv-cachelorapatch-level modeling
Understanding Quantization-Aware Training: Gradients at Quantized Weights Bias to the Low-Loss Basin
The paper proposes a geometric framework explaining why post-training quantization (PTQ) fails at aggressive bitwidths while quantization-aware training (QAT) succeeds. Modeling full-precision training as navigating a low-loss basin, PTQ selects high-loss quantized points when the quantization grid matches the basin width. QAT's straight-through estimator biases gradients toward the basin by evaluating them at quantized weights. Theoretical analysis proves QAT's recovery under local quantizer-compatibility, with experiments validating PTQ's basin-crossing failures and QAT's recovery across vision/language models and quantization schemes.
quantization-aware trainingpost-training quantizationlow-loss basinstraight-through estimatorquantization grid
Sustainability and Artificial Intelligence: Necessary, Challenging, and Promising Intersections
This article maps intersections between artificial intelligence and sustainability research by analyzing 541 publications from the Web of Science database. The study identifies convergence on wicked problems characterized by complexity, interconnection, and dynamism. Results reveal growing centrality of green/sustainable science in bridging disciplines, with specific journals and concepts emerging as key nodes. The analysis highlights necessary but challenging interactions, proposing pathways to diversify AI applications for sustainable development across institutional contexts.
sustainabilitywicked problemsbibliometric analysisgreen technologyinstitutional applications
LATTEArena: An Evaluation Framework for LLM-powered Tabular Feature Engineering (Extended Version)
The paper introduces LATTEArena, a competitive evaluation framework for LLM-powered automated tabular feature engineering (LATTE). The framework provides a six-dimensional taxonomy decomposing 15 methods, a modular arena for controlled comparison, multi-dimensional assessments (performance, cost, robustness), and component-level ablation studies. Key findings include Tree-of-Thought with Monte Carlo Tree Search achieving optimal cost-effectiveness and RPN/Code output formats excelling in classification/regression tasks. The authors release a modular framework and 4000+ execution logs for benchmarking.
tabular feature engineeringlarge language modelsmonte carlo tree searchtree-of-thoughtautomated feature engineering
The Token Not Taken: Sampling, State, and the Variability of AI Agent Outputs
The manuscript disentangles sources of variability in agentic AI systems, distinguishing intrinsic token sampling from extrinsic environmental factors. It analyzes how foundation models embedded in orchestration loops generate divergent outputs through probabilistic token decoding, which propagates into tool calls, code paths, and agent states. The work clarifies stochasticity in AI agents, reproducibility under matched conditions, and the distinction between deterministic execution and behavioral consistency in deployment. By systematically separating these layers, it provides a framework for understanding variability in AI system outputs.
token samplingfoundation modelorchestration loopagentic aistochasticity
SpaceVLN: A Zero-Shot Vision-and-Language Navigation Agent with Online Spatial Cognitive Memory and Reasoning
SpaceVLN introduces a zero-shot vision-and-language navigation agent featuring Spatial Cognitive Memory and Task-Guided Spatial Reasoning. The method employs a stagewise closed-loop framework that abstracts explored regions into Spatial Waypoints and maintains subtask-grounded landmark evidence, enabling hierarchical spatial representation. Spatial-CoT integrates task-progress reasoning with spatial perception, supporting embodied navigation without task-specific training. Evaluated on R2R-CE, RxR-CE, GN-Bench, and HM3D-OVON, SpaceVLN achieves state-of-the-art zero-shot performance, validated through real-robot deployment.
vision-and-language navigationspatial cognitive memoryzero-shot learningembodied navigationtask-guided reasoning
Baichuan-M4: A Clinical-Grade Medical Agent System for Continuous Care
Baichuan-M4 introduces a clinical-grade medical agent system for continuous care, combining three technical pillars: Baichuan-Harness (a unified runtime for RL training and deployment with action constraints and multi-agent coordination), a core reasoning model trained via continuous-care RL with SPAR++ and path compression, and a clinical tool layer for multimodal perception. The system achieves state-of-the-art performance across medical evaluation tasks, including OSCE-style consultation (3.3% hallucination rate), long-context memory, and multimodal understanding while maintaining safety and evidence-based retrieval.
continuous-care reinforcement learningspan-level reward modelingmultimodal medical perceptionclinical tool layerosc-style consultation
RTL-BenchLS: A Large-Scale Benchmark for RTL Reasoning and Generation with Large Language Models
The authors introduce RTL-BenchLS, a large-scale benchmark for evaluating LLM-based RTL reasoning and generation, addressing limitations of existing benchmarks in scale and task scope. The benchmark comprises over 10,000 formally verified Verilog designs and proposes three novel tasks: round-trip reasoning, masked-content reasoning, and repository-issue reasoning, verified via formal equivalence checking without manual testbenches. Evaluation of eight LLMs shows low performance (12-28% accuracy), indicating RTL-BenchLS's substantial challenge and utility for guiding future LLM development in hardware design.
rtl-benchlsverilogformal equivalence checkingllm-based generationhardware design automation
Diverse Thinking Schemata Elicit Better Reasoning in Large Language Models
The paper proposes Diverse Schemata Policy Optimization (DiScO), a framework that enhances reasoning in large language models by promoting diversity in thinking schemata—defined as reasoning transitions and answer candidates. DiScO employs schemata awareness, reinforcement learning for diversity, and diverse inference-time reasoning. Evaluations on mathematical reasoning benchmarks show DiScO outperforms standard group relative policy optimization, with human analyses confirming improved error recovery. The results highlight the importance of schema diversity for scaling model reasoning capabilities.
thinking schematareasoning transitionsanswer candidatesreinforcement learningmathematical reasoning
An Effective Router for Vision-Language Model Selection
The paper introduces ARMS, an effective router for vision-language model (VLM) selection, addressing challenges of specialized data scarcity, ineffective feature representation, and rigid model adaptation. ARMS enhances input signals with VLM profiles and employs a simple architecture to improve query and VLM capability representations. It supports incremental and independent training strategies for adapting to new VLMs. Experiments on in-distribution and out-of-distribution test sets show ARMS (800M parameters) outperforms larger commercial models like GPT-4o. The work includes a multimodal dataset of 32,626 image-text queries across seven VLMs.
vision-language modelsmodel selectionmultimodal datasetincremental trainingrouter architecture
CARE: A Conformal Safety Layer for Medical Summarization
The paper introduces CARE, a conformal safety layer for medical summarization that provides formal guarantees against hallucinations and omissions in LLM-generated summaries. The method employs two conformal risk controllers: one bounding the probability of unflagged hallucinations, and another controlling the expected fraction of important omissions. By jointly calibrating over threshold space (τ,γ), CARE achieves 5× fewer flagged sentences than baselines while maintaining risk bounds (α=0.15) across five medical tasks with ~100 labeled documents per domain. Clinician evaluations show 28.6 percentage point improvement in omission detection.
conformal risk controlmedical summarizationhallucination detectionomission controldistribution-free guarantees
Hardening Agent Benchmarks with Adversarial Hacker-Fixer Loops
The paper introduces hacker-fixer loops, a method for creating exploit-resistant verifiers in agent benchmarks without manual patching. The approach alternates three LLM agents: a hacker attempts to bypass the verifier, a fixer patches vulnerabilities, and a solver validates legitimate solutions. The method reduces attack success rates from 62% to 0% on KernelBench and from 76%/61% to 0% against stronger models like Gemini 3.1 Pro and Claude Opus 4.7. The authors release Terminal Wrench, containing 323 hackable environments and 3,632 trajectories, to facilitate future research.
adversarial benchmarkingreward hackingllm agentsverifier hardeningexploit-resistant
AlloSpatial: Agentic Harness Framework for Spatial Reasoning in Foundation Models
AlloSpatial introduces an agentic framework for allocentric spatial reasoning in multimodal foundation models (MFMs), addressing their fragility in global spatial representation. The method combines World2Mind, a cognitive mapping sandbox that generates allocentric-spatial trees and route maps, with a Spatial Reasoning Harness for tool-use judgment and modality-decoupled cue collection. Evaluations on VSI-Bench and MindCube show 5%-18% improvements in proprietary models without training, while trained agents outperform larger general-purpose models, demonstrating the efficacy of structured allocentric representations and verifiable reasoning.
allocentric spatial reasoningmultimodal foundation modelscognitive mappingtool-use judgmentgeometry-semantic arbitration
NutriMLLM: Multimodal Large Language Models for Dietary Micronutrient Analysis
NutriMLLM introduces the first family of vision-language models specialized for comprehensive dietary micronutrient estimation, addressing limitations of existing multimodal large language models (MLLMs) in this domain. The authors repurposed a decade of population-scale 24-hour dietary recalls to generate a synthetic corpus of 1.1 million image-description-nutrient triplets, each annotated with 65 nutrients. Fine-tuning Qwen3-VL and GLM-4.6V-Flash on this corpus yielded NutriMLLM variants, evaluated using a four-component framework measuring abstention, hallucination, usability, and numerical accuracy. On real food images, NutriMLLM achieved near-complete coverage across all nutrients, with the largest variant matching or exceeding proprietary baselines (GPT-5, Gemini 3, Claude Sonnet 4.5) in accuracy for most nutrients.
micronutrient estimationmultimodal large language modelssynthetic corpusvision-language modelsdietary recalls
PACT: Learning Diverse Diagnostic Strategies via Privileged Synthesis and Branch Consensus
PACT introduces a framework for learning diverse diagnostic strategies in clinical settings through privileged synthesis and branch consensus. The method combines Doctor-Patient-Supervisor (DPS) dialogue synthesis, which leverages complete electronic medical records for quality control while restricting agents to patient-visible information, with periodic anchor consensus training of paradigm-specific LoRA Branches. Evaluated on a dynamic multi-turn Chinese medical diagnosis benchmark, PACT achieves state-of-the-art performance on diagnostic outcome and consultation-process metrics compared to proprietary and specialized baselines.
clinical diagnosismulti-paradigm reasoninglora brancheselectronic medical recordsconsensus training
Report on CHIIR 2026 Workshop on Generative AI and Academic Search (GAI&AS)
The CHIIR 2026 Workshop on Generative AI and Academic Search (GAI&AS) advanced research on integrating generative AI into academic search systems, focusing on summarization, recommendation, synthesis, and conversational interaction. Participants explored foundational principles, applications, and search-as-learning paradigms, emphasizing transparency, credibility, and research integrity. Methodological discussions included guiding theories, design principles, and community-building efforts to develop human-centered GenAI-enhanced systems. The workshop highlighted diverse research initiatives and strong community interest in leveraging GenAI to support higher-order cognitive processes and long-term scholarly needs.
generative aiacademic searchsearch-as-learninghuman-centered designresearch integrity
PAI: Preserving Amplitude Information in Representation-Based Time-Series Anomaly Detection
The paper introduces PAI, a novel anomaly scoring scheme for representation-based time-series anomaly detection that preserves amplitude information. Current methods suffer from amplitude-agnostic embeddings, degrading performance on amplitude-related anomalies. PAI combines a diagnostic module (testing cosine/Euclidean scoring) with score augmentation (median/MAD deviation and local mean-shift scores) to fuse representation and amplitude features. Evaluated on TSB-AD-U-Eva and TAB UV datasets, PAI improves all four baseline methods, achieving 98.4% and 36.8% average VUS-PR gains respectively. PaAno + PAI outperforms SOTA by 15%. Results confirm amplitude retention's critical role in anomaly detection.
anomaly detectiontime-seriesamplitude-agnosticvus-prts2vec
From Statute to Control Flow: Span-Grounded Deontic Trees for Defeasible Scope Parsing
The paper introduces NormBench, a 2,290-provision benchmark for defeasible scope parsing to diagnose Silent Scope Omission (SSO) failures in rule-following agents. It proposes Span-Grounded Deontic Trees (SG-DT), a compiler-style intermediate representation that anchors logical branches to source spans with explicit exclusion guards. Evaluations on frontier LLMs reveal Recursion Decay (performance drop with defeater depth) and an Auditability Trap (span retrieval without correct control flow), while SG-DT improves exception-active case handling despite mixed aggregate accuracy.
defeasible scope parsingspan-grounded deontic treessilent scope omissionrecursion decayauditability trap
PolyBuild: An End-to-End Method for Polygonal Building Contour Extraction from High-Resolution Remote Sensing Images
PolyBuild introduces an end-to-end method for direct polygonal building contour extraction from high-resolution remote sensing images, eliminating post-processing steps. The approach combines an Initial Contour Generation Module (ICGM) for simultaneous object detection and initial contour extraction using sub-region center features, with a Contour Optimization Module (COM) that refines contours via a hybrid CNN-Transformer architecture integrating local and global spatial relationships. Evaluated on three datasets, PolyBuild outperforms state-of-the-art mask-based and contour-based methods in building boundary delineation.
polygonal contour extractionremote sensing imagesinitial contour generationcontour optimizationcnn-transformer architecture
Oversight Has a Capacity: Calibrating Agent Guards to a Subjective, Fatiguing Human
The paper introduces an open-source agent-oversight system that operationalizes human-in-the-loop safety mechanisms for LLM agents, addressing the challenges of subjective risk judgment and finite human attention. Using a hand-labeled dataset of 125 adversarially-weighted agent actions, the study demonstrates moderate reviewer agreement (Fleiss' kappa = 0.52) and frames guarding as selective classification under asymmetric cost. Key findings include an inverted-U relationship between oversight and safety due to reviewer fatigue, and the vulnerability to flooding attacks when human attention is overloaded. The system integrates prior techniques like FALCON and DeCCaF to quantify guard performance.
llm agentshuman-in-the-loopselective classificationfatigue-aware learningflooding attack
Failure-Aware Refinement of Vision-Language Model for Lithography Defect Detection
The study introduces a two-stage vision-language framework for semiconductor lithography defect detection, addressing limitations of direct fine-tuning. First, Qwen3-VL is fine-tuned with LoRA as a vision-language adapter to predict defect counts, categories, and bounding boxes. Second, a refinement module is trained using first-stage prediction failures and corrected labels, enabling error correction and improved inference. This failure-aware refinement process mitigates common test-time errors such as false positives, missed defects, and incorrect defect types, enhancing detection accuracy beyond single-stage fine-tuning.
vision-language frameworklithography defect detectionqwen3-vllorafailure-aware refinement
Order Matters: Unveiling the Hidden Impact of Macro Placement Sequences via Proxy-Guided LLM Evolution
The paper introduces OrderPlace, a proxy-guided LLM evolution framework for optimizing macro placement sequences in chip physical design, demonstrating that placement order significantly impacts solution quality. OrderPlace explores code-level policies beyond static heuristics, employing a lightweight proxy evaluation mechanism to efficiently filter candidate sequences using a deterministic greedy probe. Experiments on the ISPD 2005 benchmarks show that OrderPlace reduces wirelength by 34.04% and 14.08% compared to WireMask-EA and EGPlace, respectively, uncovering novel ordering strategies.
macro placementproxy-guided evolutionwirelength reductionispd benchmarksdeterministic greedy probe
Few-shot Class-variable Incremental Audio Classification via Prototype Adaptation and Pseudo Class-variable Training
The paper introduces Few-shot Class-variable Incremental Audio Classification (FCIAC), addressing scenarios where class counts dynamically increase or decrease, unlike prior class-incremental approaches. The proposed method combines a dynamically structured prototype adaptation network for classifier initialization with pseudo class-variable training to enhance adaptability. Evaluations on three public datasets demonstrate superior average accuracy over existing methods. Code is publicly available.
few-shot learningclass-incremental learningprototype adaptationaudio classificationdynamic architecture
A multi-agent system for spine MRI report generation from multi-sequence imaging
SpineAgent introduces a multi-agent system for automated spine MRI report generation, addressing the challenge of integrating multi-sequence data while preserving sequence-specific diagnostic information. The method pre-trains DINOv3-based encoders on T1- and T2-weighted sequences, then employs a continual training strategy with a synthesizer to embed diverse MRI sequences into patient-level representations. Evaluated on 32,047 patients and 453,683 MRI series, SpineAgent achieves state-of-the-art performance in pathology classification, localization, and multimodal retrieval, validated by both automated metrics and expert radiologist review.
multi-agent systemspine mridinov3multi-sequence embeddingreport generation
FAME: Forecastability-Aware Mixture of Experts for Heterogeneous Time Series Forecasting
FAME introduces a forecastability-aware mixture-of-experts framework for heterogeneous time series forecasting, addressing the challenge of model suitability across diverse data regimes. The method learns multidimensional forecastability fingerprints for each series, mines expert-suitability targets from validation performance, and employs a cost-aware sparse router to activate a budgeted set of experts per series. Evaluated on SNBC's industrial dataset (5,000+ machines, 60M+ transactions) and public benchmarks, FAME Top-2 reduces MSE by 12.4% over LightGBM while averaging 1.92 experts per series. The deployed system integrates with replenishment planning, demonstrating systematic expert suitability variation across data regimes.
mixture-of-expertsforecastability fingerprintsparse routingheterogeneous time seriesdemand forecasting
Cheap Reward Hacking Detection
The study introduces a transformer encoder trained to map Terminal-Wrench trajectories onto a unit sphere, where embedding distance approximates L1 distance between reward and metadata signals. A linear probe on this embedding detects reward hacking with AUC 0.9467 and TPR@5%FPR 0.8296, matching a sanitized LLM-as-judge baseline (AUC 0.9510) while outperforming its TPR@5%FPR (0.7130) at significantly lower computational cost. The encoder's performance drops to AUC 0.6213 when natural-language reasoning is removed, demonstrating its reliance on multimodal signals.
transformer encoderterminal-wrench trajectorieslinear probeauctpr@5%fpr
Benchmarking Vision-Language-Action Models on SO-101: Failure and Recovery Analysis
The study introduces a standardized real-world benchmark for evaluating Vision-Language-Action (VLA) models and imitation learning policies on the low-cost SO-101 robotic platform. The benchmark includes four manipulation tasks with unified evaluation protocols, enabling systematic comparison under embodiment uncertainty. Fine-tuned models ($π_{0.5}$, SmolVLA, Wall-X, ACT) are evaluated using real-world teleoperated demonstrations, incorporating a structured failure taxonomy and recovery-aware metrics. Results indicate that pretrained VLA policies generally outperform imitation learning baselines, though performance varies by task. Execution instability is identified as the primary failure source, with recovery capabilities differing across architectures. This highlights the need for failure and recovery analysis beyond binary task success, establishing SO-101 as a practical benchmark for low-cost robotic deployment.
vision-language-action modelsembodiment uncertaintyfailure taxonomyrecovery-aware metricslow-cost robotic platform
Can the Environment Speak for Itself? $T^{2}$-GRPO: A Turn-Trajectory Group Relative Policy Optimization for Caregiver Agents
The paper introduces T$^{2}$-GRPO, a turn-trajectory group relative policy optimization method for caregiver agents in dementia care. The framework decouples reinforcement learning into two normalized reward horizons, using environment-grounded rewards from patient state transitions and trajectory-level evaluations. It employs independent centered-rank normalization to preserve heterogeneous reward signals and prevent collapse. Experiments demonstrate T$^{2}$-GRPO's superiority over baselines in handling immediate feedback, long-term outcomes, and safety constraints in emotionally sensitive care scenarios.
caregiver agentsreward normalizationdementia carereinforcement learningsafety constraints
Intelligent Character Recognition of Handwritten Forms with Deep Neural Networks
The study introduces a unified deep neural network approach for handwritten form processing, combining character detection and classification into a single task. Training data is synthetically generated from form templates and existing datasets (EMNIST), avoiding manual annotation. Compared to state-of-the-art two-stage methods, this single-task approach demonstrates superior performance, achieving 88.28% recognition accuracy on real-world exam forms. Limitations in EMNIST's applicability to handwritten Latin letters necessitated dataset customization.
handwritten character recognitiondeep neural networksemnist datasetsingle-task learningsynthetic training data
Hybrid E-Assessment in Higher Education: Semi-Automated Grading of Paper-Based Written Examinations
The paper proposes a hybrid e-assessment method combining paper-based exams with semi-automated grading to address limitations of fully digital approaches in higher education. Students handwrite structured answers that are digitally captured, with vision-capable LLMs enabling reliable character recognition under exam conditions. A two-pass validation system and solution key comparison reduce misclassifications, improving assessment validity, fairness, and scalability for large cohorts while maintaining problem-oriented tasks.
e-assessmentsummative examinationhandwritten recognitionlarge language modelstwo-pass validation
sGPO: Trading Inference FLOPs for Training Efficiency in RLVR
The paper introduces sorted Group Policy Optimization (sGPO), a compute-efficient method for Reinforcement Learning with Verifiable Rewards (RLVR) that reduces wasted training FLOPs by leveraging inference FLOPs. sGPO profiles query difficulty via a small batch of parallel samples under the initial policy, using the empirical success rate to adaptively set rollout group sizes, filter data, and construct a curriculum. Experiments show sGPO matches or exceeds baseline performance while reducing total training compute by 3×, including the upfront inference cost.
reinforcement learningverifiable rewardssample efficiencyadaptive groupingcurriculum learning
Intrinsic Selection and Particle Resampling for Inference-Time Scaling Beyond Domain Verifiability
The paper introduces Intrinsic Selection (iS) and Intrinsic Particle Filtering (iPF) for Inference-Time Scaling (ITS) in domains lacking verifiability. By leveraging length-adjusted tail entropy from parallel sample sets, iS ranks candidates post-hoc, while iPF enables step-level resampling for improved reasoning trajectories. The method achieves 20% gains in engineering design selection, 6.1-point pass@1 improvements on hard math problems, and 26.5% gains in clinical responses via Particle Distillation (dPF). The approach operates without trained reward models or ground-truth verification, applicable across broad-purpose and multimodal architectures.
inference-time scalingintrinsic selectionparticle resamplinglength-adjusted tail entropykl-guided resampling
A Resilience-as-a-Service assessment framework for coordinated disruption response in interdependent urban transit systems
The paper proposes a KPI-driven, time-indexed framework for assessing resilience of disruption response solutions in urban transit systems, addressing the lack of comprehensive decision support tools. The framework integrates an optimization model with agent-based behavioral simulation, evaluating multiple dimensions including vulnerability, adaptability, robustness, resilience loss, responsiveness, cost-based performance, emissions, and equity. It specifically models secondary service degradation on helper lines when vehicles are reallocated. Implemented on the RER B transit line in Paris, results demonstrate that coordinated multimodal strategies yield balanced resilience profiles, outperforming single-mode alternatives in service continuity, total disruption cost, equity, and environmental performance. Sensitivity analysis identifies optimal conditions for multimodal response efficacy.
resilience assessmentagent-based simulationoptimization modelservice degradationmultimodal response
BLM-SGAN: Bidirectional Language Modeling for Semantic-Spatial Text-to-Image Generation
BLM-SGAN introduces bidirectional language modeling for semantic-spatial text-to-image generation, addressing challenges like long-range dependencies and vanishing gradients in GAN-based T2I models. The method integrates BERT's attention mechanisms to capture contextual information and manage extended sequences efficiently. Evaluated on bird image generation, BLM-SGAN achieves state-of-the-art performance with an Inception Score of 5.45 ± 0.08, outperforming SSA-GAN, DF-GAN, SD-GAN, and AttnGAN.
text-to-image generationgenerative adversarial networksbidirectional language modelinginception scoreattention mechanisms
ZIPP:Zero-shot Image Personalization from Personas
The paper introduces ZIPP, a zero-shot method for personalizing text-to-image diffusion models using natural-language personas without user-specific data or fine-tuning. The approach employs an LLM to rewrite prompts from persona perspectives and trains a Graph Attention Network on a 22M-user Reddit graph to mine personas, later verbalized via an MLLM. Evaluated on ZIPBench (1.5K users, 40K images), ZIPP shows 13-20% gains across 14 LLMs, matches few-shot baselines with 100+ examples, and achieves superior preference alignment (CMMD 0.16) and 79% human preference win rates over generic generation.
zero-shot personalizationdiffusion modelsgraph attention networkpersona miningprompt rewriting
Beyond Pass Rate: A Multilingual, Execution-Grounded Evaluation of Open Code LLMs
The study introduces a multilingual, execution-grounded evaluation framework for code generation models, addressing limitations of aggregate pass rates. It assesses 9 open code LLMs on 2,707 LeetCode problems across 12 languages, analyzing 325,343 problem-model-language jobs with execution outcomes and static-analysis signals. Results reveal Yi-Coder-9B-Chat leads with 23.64% mean correctness (vs. 57.2% human baseline), while Qwen2.5-Coder-14B-Instruct excels on hard problems and Gemma-2-27B-IT on lint pass rates. Compile errors dominate failures (63.25%), highlighting gaps in semantic correctness. The work demonstrates the necessity of artifact-preserving, multilingual evaluation for nuanced model comparisons.
code generation modelsexecution-grounded evaluationmultilingual benchmarksstatic-analysis signalslint pass rate
Instrumental convergence and power-seeking
The article critically examines the argument that artificial intelligence poses existential risks through power-seeking behavior, focusing on the instrumental convergence thesis. It analyzes existing defenses of this thesis, concluding that none sufficiently substantiate the strong version required to support power-seeking claims. The study employs logical and methodological scrutiny of AI risk arguments, highlighting gaps in current reasoning. Implications for longtermism, AI governance, and risk assessment methodologies are discussed, suggesting a need for more robust theoretical foundations in AI safety research.
instrumental convergencepower-seekingexistential risklongtermismai governance
Inference-Time Conformal Reasoning with Valid Factuality Control for Large Language Models
The paper introduces Inference-Time Conformal Reasoning (ITCR), a framework integrating conformal prediction (CP) into reasoning graph generation for large language models (LLMs) to provide valid factuality control. ITCR learns a structure-level factuality uncertainty function aggregating claim-level signals, calibrates a conformal threshold for generation stopping, and theoretically guarantees coverage. Experiments across datasets show empirically valid coverage, with inference-time calibrated graphs outperforming post-hoc pruning in downstream reasoning tasks.
conformal predictionfactuality controlreasoning graphsuncertainty quantificationlarge language models
Knowledge Graphs and Reasoning LLMs for Finding Simple Yet Effective Transcriptomic Perturbation Predictors
The work demonstrates that K-nearest neighbor models using biological knowledge graphs achieve competitive performance in predicting transcriptomic effects of unseen gene knockouts, outperforming most methods on out-of-distribution perturbations. When enhanced by reinforcement learning (RL)-optimized large language models (LLMs) that modify the neighborhood selection, performance matches state-of-the-art methods on Replogle et al. (2022) cell lines. RL training also improves the LLM's downstream differential expression prediction without direct optimization, showing knowledge graphs' efficacy as priors and RL's potential for refining biological prediction tools.
knowledge graphstranscriptomic perturbationreinforcement learninglarge language modelsdifferential expression prediction
Momentum for Reasoning: Dense Intrinsic Signals in Policy Optimization
The paper introduces Intrinsic Signal Policy Optimization (ISPO), a reinforcement learning method that mitigates two failure modes in reasoning tasks—Zero-Advantage Collapse and Hallucinated Certainty—by densifying rewards with intrinsic signals derived from the policy's conditional probabilities. ISPO combines sequence-level informativeness metrics with token-level directional rewards, penalizing confidently incorrect predictions. Evaluated across three base models and five mathematical reasoning benchmarks, ISPO outperforms baselines, particularly on harder tasks where zero-advantage collapse is prevalent, while diagnostics confirm reduced failure modes.
reinforcement learningpolicy optimizationintrinsic signalsmathematical reasoninghallucinated certainty
STAR: Rethinking MoE Routing as Structure-Aware Subspace Learning
The paper introduces STAR (Structure Aware Routing), a novel routing method for Mixture-of-Experts (MoE) that enhances input-expert specialization by aligning routing decisions with input structure. STAR reformulates routing as a subspace learning problem, combining standard learnable routing with an evolving principal subspace updated via Generalized Hebbian Algorithm (GHA) to track dominant input patterns. Experiments on synthetic data, language, and vision tasks demonstrate consistent improvements in routing stability and downstream performance over MoE baselines, with additional robustness gains from optional test-time subspace updates under distribution shifts.
mixture-of-expertssubspace learninggeneralized hebbian algorithmrouting stabilityinput structure
Governance Controls for AI-Generated Test Artifacts in Autonomous Software Testing
The study proposes a Governance-Aware Autonomous Testing Framework (GATF) to mitigate risks in AI-generated test artifacts, addressing hallucinations, compliance violations, and explainability gaps. The framework integrates governance validation, probabilistic risk assessment, and compliance monitoring into the autonomous testing lifecycle. Evaluated on Defects4J and PROMISE datasets, GATF achieved 89.6% risk reduction, with 94.3% governance accuracy, 96.5% artifact reliability, 94.2% compliance accuracy, and 90.8% explainability performance. Results demonstrate significant improvements in reliability and transparency over conventional AI-based testing systems.
autonomous testinggovernance validationcompliance monitoringrisk assessmentexplainability analysis
Q-Delta: Beyond Key-Value Associative State Evolution
The paper introduces Q-Delta, a query-aware delta rule that integrates mixed key-query prediction errors into state evolution for linear attention models. By conditioning state evolution on query information, Q-Delta enables corrective dynamics while maintaining efficiency. The method includes stability guarantees and a hardware-efficient chunkwise-parallel implementation using Triton. Empirical results show stable optimization, competitive throughput, and performance improvements over baselines in language modeling and long-context retrieval tasks.
linear attentiondelta rulestate evolutionquery-conditionedtriton
Bridging Expert Knowledge and Automated Feature Engineering via Self-Evolution
FEST (Feature Engineering with Self-evolving Trees) introduces a method for generating interpretable features from unstructured data by combining dual-stream feature generation (semantic and deterministic), semantic deduplication, and tree-guided iterative evolution. The approach outperforms baselines in 17 of 20 classifier-task combinations across brand classification, content authenticity detection, and stress detection, with a mean accuracy gain of 4.2 percentage points. FEST achieves 60-80% coverage of expert-designed features and improves accuracy by 6-12 percentage points when seeded with expert guidelines. The study also releases BrandGuide, a dataset pairing expert-designed features with 1M+ assets across 2,683 brands.
feature engineeringinterpretable mlsemantic deduplicationdual-stream generationexpert alignment
Scaling Decision-Focused Learning to Large Problems with Lagrangian Decomposition
The paper introduces a Lagrangian decomposition framework to scale decision-focused learning (DFL) for large predict-then-optimize problems, addressing computational bottlenecks. It proposes a surrogate objective and two loss functions, compatible with SPO+ and IMLE methods. Two framework variants balance efficiency and solution quality. Experiments on knapsack and portfolio optimization benchmarks show superior scalability, handling 8× larger instances than prior work while maintaining competitive performance. Implementation is publicly available.
decision-focused learninglagrangian decompositionpredict-then-optimizespo+imle
AI-Augmented Closed-Loop Quality Engineering: A Reference Architecture for Continuous Software Quality Intelligence
The paper proposes a closed-loop reference architecture for continuous software quality intelligence, integrating AI-augmented requirement feature mining, risk-based test prioritization, defect prediction, and production incident analysis. The method introduces a feedback learning mechanism that propagates production signals (defect severity and incident impact) to subsequent releases, evaluated on a semi-synthetic dataset of 4,500 requirements, 27,049 test cases, 13,089 defects, and 7,841 incidents across six release cycles. Results show defect leakage reduction (0.19 to 0.13), detection effectiveness improvement (0.72 to 0.84), and 35% faster test execution compared to non-adaptive baselines, with stable performance across releases.
closed-loop architecturesoftware quality intelligencedefect predictionrisk-based test prioritizationfeedback learning
Evaluating AI Investment Strategies
The paper presents an exact decomposition theorem showing that cumulative regret in dynamic policies equals the sum of per-period covariances between cost vectors and decisions, extending Aldridge's (2026) single-period result to stochastic dynamic programming. The identity holds under i.i.d. costs and mean-unbiased Markov policies, with derived bias corrections for non-stationary cases and a Bellman recursion connecting to RL algorithms. For rolling-window policies, estimation-error bias scales as O(d/w). The decomposition enables model-free auditing in strategic environments (e.g., platform mechanisms, repeated games) via a consistent, asymptotically normal trajectory estimator computable in O(T·nd) time.
regret decompositionstochastic dynamic programmingalgorithmic auditingbias correctiontrajectory estimator
RAILS: Verification-Native Clearing For Agentic Commerce
The paper introduces RAILS (Real-Time Agent Integrity & Ledger Settlement), a verification-native clearing protocol for agentic commerce, addressing the lack of neutral mechanisms to determine obligation fulfillment and settlement actions in autonomous agent transactions. The method employs seven primitives (Obligation Object, Evidence Envelope, Verification Mesh, Clearing Decision, Settlement Instruction, Clearing Passport, Finality Rules) within a formal model of admissibility-graded verification, ensuring no financially material settlement is supported by substandard evidence. The protocol achieves a falsifiable soundness property, distinguishing it from prior approaches that lack such formal guarantees.
agentic commerceclearing protocolverification meshadmissibility-graded verificationsettlement instruction
How Many Counterfactuals Does It Take? Probing VLM Hallucinations Through Circuits and Causal Effects
The paper proposes a method to quantify the robustness of hallucinated predictions in Visual Language Models (VLMs) through counterfactual analysis. It introduces a causal influence metric based on log-probability differences across factual, counterfactual, and activation-patched model runs, combined with circuit discovery techniques (CD-T) to identify responsible model components. The authors derive empirical bounds on the minimum number of counterfactual samples m needed to reliably detect prediction instability, using concentration inequalities and variance estimates of the causal influence distribution.
visual language modelscounterfactual robustnesscausal influencecircuit discoveryactivation patching
Unifying Object-Centric World Models and Diffusion Policy: A Hierarchical Framework for Multi-Stage Robotic Tasks
WorldDP introduces a hierarchical framework for multi-stage robotic manipulation tasks, combining object-centric world models with Diffusion Policy. The method employs a high-level world model as a transition function to optimize feasible subgoals during runtime, executed by a low-level Diffusion Policy. Object-centric representations decouple environmental entities, enabling sequential planning. Evaluated across robotics benchmarks, WorldDP consistently outperforms baselines, demonstrating that integrating physically grounded planning with efficient execution enhances multi-stage task performance.
world modeldiffusion policyobject-centric representationmodel predictive controlhierarchical framework
TeamHerald@CHIPSAL 2026: Hate Speech Detection and Sentiment Analysis of Nepali Memes using Transformer-based Architectures and Ensemble Learning
This work investigates hate speech detection and sentiment analysis for Nepali memes using Transformer-based models and ensemble learning, addressing challenges from code-mixing and limited baseline resources. The text-centric approach employs OCR for text extraction, evaluating six Transformer architectures and comparing Hard vs. Soft Voting ensembles. Results show a decoder-only model achieves best binary classification performance (hate speech), while Soft Voting yields 15.8% relative Macro F1 improvement for multi-class sentiment analysis, demonstrating task-dependent ensemble efficacy.
transformer-based architecturesensemble learningmacro f1-scorecode-mixingsentiment analysis
RadOT-Eval: Auditable Structured-Evidence Transport for Radiology Report Evaluation
RadOT-Eval introduces an interpretable structured-evidence optimal transport framework for auditing radiology report generation, addressing high-stakes errors like omitted findings and hallucinated content. The method decomposes reference and candidate reports into attribute-structured clinical evidence units, aligns evidence using entropy-regularized optimal transport, and employs clinically meaningful discrepancies in a monotone risk model to predict error burden. Evaluated on the RadEvalX dataset, RadOT-Eval achieves Spearman correlations of 0.715, 0.548, and 0.399 with total, clinically significant, and clinically insignificant error burdens, outperforming standard metrics and the LLM-based GREEN-radllama2-7B evaluator. It also demonstrates robust performance in corruption-sensitivity tests, achieving 0.768 AUROC and a 0.990 win rate.
optimal transportstructured evidenceentropy regularizationmonotone risk modelspearman correlation
APEX4: Efficient Pure W4A4 LLM Inference via Intra-SM Compute Rebalancing
The paper introduces APEX4, a system for efficient W4A4 (4-bit weight and activation) LLM inference via intra-SM compute rebalancing. It identifies the Tensor Cores to CUDA Cores throughput ratio (ρ) as a key hardware indicator for W4A4 viability, showing performance varies from 2.0–2.5× speedup on RTX 3090 (ρ=16) to 0.43–0.47× slowdown on A100 (ρ=64). APEX4 co-designs pure INT4 GEMM kernels with ρ-aware granularity adaptation, achieving perplexity within 0.63 of FP16 on LLaMA-2-70B and 4.0–4.4% higher zero-shot accuracy than W4Ax Atom-g128. Deployed in vLLM, it delivers up to 2.09× end-to-end speedup on compatible GPUs.
w4a4 quantizationintra-sm computetensor coresgemm kernelsgranularity adaptation
Structure-Conditioned Actor-Critic Branches for Quality-Diversity Reinforcement Learning
The paper proposes SV-QD-RL, a structure-value coupled framework for quality-diversity reinforcement learning (QD-RL) that generates diverse policy repertoires through structure-conditioned actor-critic branches. Each branch includes an actor with a structural mask, branch-specific critic, replay state, and evaluation attributes, enabling specialized learning trajectories. Experiments on MuJoCo tasks demonstrate improved archive quality and behavioral diversity, with structural conditioning, critic differentiation, and memory-consistent refinement contributing to performance. The framework provides selectable policy alternatives under varying behavior-level requirements.
quality-diversity reinforcement learningactor-critic branchesstructural maskbehavioral specializationpolicy repertoires
Artificial Intelligence for Mathematical Reasoning: An Integrated Survey of Language Models, Neuro-symbolic Systems, and Verified Discovery
This survey synthesizes advancements in AI for mathematical reasoning, categorizing approaches into four axes: informal reasoning, formal proof systems, mathematical discovery, and inference techniques. It evaluates benchmarks across arithmetic, geometry, and formal proving, while critiquing failure modes like brittleness and reward hacking. The analysis highlights future directions in verified discovery and reasoning efficiency, supported by companion resources for AI-assisted formalization.
mathematical reasoningneuro-symbolic systemslanguage modelsverified discoverybenchmark evaluation
Deep Active Re-Labeling: Toward Noise-Resilient Annotation Efficiency
The paper introduces Deep Active Re-Labeling (DARL), a framework addressing human annotation errors in Deep Active Learning (DAL). DARL allocates a portion of the annotation budget to re-label potentially noisy data, leveraging active noise sampling strategies informed by human learning patterns. Theoretical insights suggest that re-labeling even a small fraction of data can effectively mitigate noise when the model identifies noisy instances. Experimental results demonstrate that DARL improves data efficiency and produces a relatively noise-free annotation dataset under the same annotation budget compared to traditional DAL approaches.
deep active learningannotation noisere-labelingnoise samplingannotation efficiency
Hybrid Neural Network and Conventional Controller Approach for Robust Control of Highly Unstable Systems: Application to Tilt-Rotor Control
This paper introduces a neural-network-enhanced sliding mode controller (SMC) for fully actuated tilt-rotor systems, addressing the instability of direct input-output learning approaches. The method decomposes dynamics into input-independent and input-dependent components, learned via lightweight networks from flight logs, enabling robust control under uncertainties. Comparative evaluations show LSTM-based predictors outperform MLP variants in robustness and runtime efficiency. The hybrid approach combines conventional control theory with data-driven learning, validated through real-world data and simulations.
tilt-rotorsliding mode controllstmmlpfully actuated
SNR-ST-Mix: Sample-specific Neighborhood Regression Mixup for Augmented Spatial Transcriptomics Imputation with Deep Neural Network
The paper introduces SNR-ST-Mix, a data augmentation framework for spatial transcriptomics (ST) imputation that preserves local biological structure by mixing samples within k-nearest spatial neighbors and weighting interpolations by expression similarity. This method addresses limitations of conventional augmentation strategies, which often produce biologically implausible interpolations for regression tasks. Experiments across multiple tissue types show SNR-ST-Mix outperforms existing methods without architectural modifications or added computational cost, enhancing prediction stability and generalization.
spatial transcriptomicsdata augmentationk-nearest neighborsexpression similarityregression tasks
Structuring agentic AI for HPC code modernization
The paper presents a structured AI-assisted methodology for modernizing NMAP-RKPM, a 60,000-line Fortran-based MPI application, into OpenMP-parallel C++ MPI code. Using agentic AI with manual example curation, buildability constraints, and scoped sessions, the team achieved conversion within months despite LLM limitations. Results demonstrate effective code transformation while detailing encountered challenges and methodological rationale.
agentic aicode modernizationrkpmopenmpmpi
Rethinking the Divergence Regularization in LLM RL
The paper introduces Divergence Regularized Policy Optimization (DRPO), a novel reinforcement learning method for large language models (LLMs) that improves upon existing trust-region approaches. DRPO replaces the hard masking mechanism in DPPO with a smooth advantage-weighted quadratic regularizer on policy shift, enabling continuous gradient attenuation and corrective signals beyond trust-region boundaries. Experiments demonstrate DRPO's superior stability and training efficiency across diverse model scales, architectures, and precision settings compared to PPO and GRPO baselines.
reinforcement learningtrust-region controldivergence regularizationpolicy optimizationlarge language models
Weighted universal approximation of differentiable maps on infinite-dimensional manifolds
The paper generalizes the universal approximation theorem for functional input neural networks (FNN) to differentiable maps by incorporating derivative approximation. It introduces a weighted Nachbin theorem to establish a universal approximation theorem (UAT) for differentiable maps, extending beyond compact sets and including derivative approximation. The method involves mapping inputs from infinite-dimensional weighted manifolds to real-valued hidden layers with non-linear activations, followed by linear readouts into Banach spaces. Results demonstrate approximation capabilities for non-anticipative functionals, including horizontal and vertical derivatives, and show that linear functions of the signature can approximate path space functionals with directional derivatives.
universal approximation theoremfunctional input neural networksweighted nachbin theorembanach spacesdirectional derivatives
Echo-Memory: A Controlled Study of Memory in Action World Models
The paper introduces Echo-Memory, a controlled study examining memory mechanisms in action-conditioned world models for multi-segment video generation. The methodology fixes key variables (backbone, training, evaluation) to isolate four memory axes: capacity, compression, read-out, and recurrence. Using a three-branch evaluation protocol (replay quality, in-domain loops, open-domain returns), findings reveal raw context as a strong capacity baseline, compression trade-offs in spatial memory, and block-wise state-space recurrence as optimal for open-domain returns. Results demonstrate that memory structure significantly impacts world model performance beyond replay metrics.
echo-memoryaction-conditionedworld modelsstate-space recurrenceopen-domain return
Zero Touch Predictive Orchestration: Automating Time-Series Models for the Cloud-Edge Continuum
The paper proposes an automated time-series prediction architecture for Zero Touch Management in the Cloud-Edge Continuum (CEC), addressing the cold start problem via a novel data-mixing methodology. A lightweight Resource Exposer (RE) dynamically collects node telemetry, which is merged with the high-resolution TimeTrack dataset (45-second intervals) to compensate for sparse local data. A Neural Architecture Search (NAS) engine then generates accurate baseline models. Experiments show significant improvements in forecasting accuracy (MSE, MAE, MAPE) and convergence speed compared to training on sparse local data or generic datasets alone.
cloud-edge continuumtime-series forecastingneural architecture searchzero touch managementresource exposer
Discovering Functionally Selective Brain Regions with a Deep Topographic Multimodal Model
The paper introduces Topo-Omni, a deep topographic multimodal model that unifies visual, auditory, and language/cognitive processing on a single contiguous in-silico sheet. The model is constructed by fine-tuning a pretrained foundation model with a spatial smoothness objective, producing modality-specific clusters that align with human neuroimaging data. Results demonstrate that manipulating these clusters selectively biases perception, mirroring human intervention studies, and reveals novel natural landscape and animal networks validated in human data, suggesting a unified spatial principle for cortical organization.
topographic modelmultimodal integrationspatial smoothnesscortical organizationneuroimaging validation
iOSWorld: A Benchmark for Personally Intelligent Phone Agents
iOSWorld introduces the first interactive native iOS simulator benchmark for evaluating personally intelligent phone agents, featuring 26 custom-built iOS apps with interconnected user data spanning transactions, messages, and social relationships. The benchmark includes 133 tasks across three difficulty levels: single-app (27 tasks), multi-app (60 tasks), and memory/personalization (46 tasks). Evaluations of frontier and open-source computer-use models in vision-only and vision+XML settings show a best overall accuracy of 52%, dropping to 37% for multi-app tasks. Privileged vision+XML access improves frontier models by up to 26 percentage points, while smaller models do not benefit from accessibility-tree input. iOSWorld is released as an open-source benchmark with apps, data, tasks, and evaluation code.
ios simulatorpersonalization tasksaccessibility-treemulti-app tasksvision+xml
Perturbative Contrastive Physical Learning
The authors introduce Perturbative Contrastive Physical Learning (PCPL), a framework where learning emerges from contrasting physical states induced by controlled perturbations to inputs, boundary conditions, or parameters. PCPL unifies Equilibrium Propagation and Frequency Propagation, enabling contrast-driven updates without centralized gradient computation or explicit backpropagation. Learning geometry emerges implicitly from the system's physical response. The method is demonstrated on two platforms: (i) spring networks updating bond stiffness via displacement and force measurements, and (ii) continuous-variable photonic circuits trained using quadrature measurements and Jacobian estimates. Both platforms successfully learn classification tasks, and the photonic circuit implements analog multiplication, advancing autonomous physical learning systems.
perturbative learningcontrastive learningequilibrium propagationphotonic circuitsspring networks
Your Model Already Knows: Attention-Guided Safety Filter for Vision-Language-Action Models
The authors propose an attention-guided safety filter for Vision-Language-Action (VLA) models that enables real-time collision avoidance without additional training. By identifying a small subset of attention heads within VLA models that reliably localize target objects, the method treats all other scene elements as obstacles and integrates them into a Control Barrier Function (CBF) filter. Combined with a lightweight object tracker, this approach handles both static and dynamic obstacles. Evaluated on SafeLIBERO with moving obstacles, the method outperforms an oracle-based baseline by 43% in dynamic scenarios, demonstrating that VLA models inherently contain sufficient perceptual signals for safety filtering.
vision-language-action modelsattention headscontrol barrier functioncollision avoidancereal-time object tracking
Learning Dynamics Reveal a Hierarchy of Weight-Induced Layerwise Gram Metrics
The study introduces a novel framework for analyzing gradient descent dynamics in feed-forward ReLU networks with fixed readout and quadratic loss, focusing on field dynamics rather than weight space. By eliminating weight variables from activation dynamics, the authors derive a closed equation for residuals governed by a collective kernel that factorizes into input-geometric and dynamical co-activation matrices. For deeper networks (depth ≥3), residual dynamics maintain a layer-wise kernel structure but require a hierarchy of weight-induced Gram operators to mediate inter-layer information transport. This approach provides a systematic understanding of neural network training dynamics across varying depths.
relu networksgradient descentgram operatorsactivation dynamicsresidual dynamics
Adaptive directional gradients for parameterised quantum circuits
The paper introduces a framework for forward gradient estimators in parameterised quantum circuits (PQCs), reducing measurement costs by averaging tunable random directional derivatives. This approach generalizes stochastic gradient methods like SPSA and parameter-shift rules without ancilla qubits or controlled gates. The authors prove convergence under standard assumptions and propose QUIVER, an adaptive optimiser minimizing measurement costs. Numerical experiments demonstrate efficiency gains, training 60-qubit PQCs on ECG5000 and MNIST, and outperforming iCANS/gCANS in QAOA and VQE applications.
parameterised quantum circuitsforward gradient estimatorsmeasurement costquantum neural networksvariational quantum eigensolver
Tight Sample Complexity of Transformers
The work tightly characterizes the VC dimension and sample complexity of depth-L Transformers with W parameters processing length-T sequences. For standard Transformers, it establishes an upper bound of O(LW log(TW)) and a nearly matching lower bound of Ω(LW log(TW/L)) for VC dimension. For chain-of-thought learning, it proves teacher forcing achieves sample complexity O(LW log((T+T')W)), with a corresponding lower bound of Ω(LW log((T+T')W/L)), where T' denotes autoregressive steps.
vc dimensionsample complexitytransformerschain-of-thoughtteacher forcing
Disentanglement with Holographic Reduced Representations
The paper introduces an unsupervised learning algorithm using holographic reduced representations (HRR) for neural disentanglement, addressing the challenge of separating factors of variation in data. Unlike continuous representations, HRR treats disentangled representations as symbolic structures, leveraging the HRR unbinding operation to induce an inductive bias for factor separation. Empirical results demonstrate competitive performance against baselines in latent traversals and disentanglement metrics. An information-theoretic analysis proves HRR unbinding induces approximately independent symbol-value pairs, with a derived per-slot capacity bound quantifying reliable encoding of distinct symbolic concepts. HRR-based representations are shown to be more noise-robust and maintain reconstruction quality across varying SNRs compared to standard autoencoder models.
disentanglementholographic reduced representationsinductive biassymbolic structuresunbinding operation
Evaluating the Representation Space of Diffusion Models via Self-Supervised Principles
The paper introduces a framework for evaluating diffusion models' representation and generation capabilities through self-supervised principles. It decomposes features into invariant and residual components, proposing the Invariant Contamination Ratio (ICR) to quantify contamination of invariant signals. Results show that invariance peaks at intermediate noise levels, correlating with optimal downstream classification performance. ICR also detects early memorization in data-limited regimes via residual energy in Fisher directions, serving as a training-time indicator without external evaluators.
diffusion modelsself-supervised learninginvariant contamination ratiofisher-based metricrepresentation learning
BrainSurgery: Reproducible and Reliable Declarative Weight Manipulations for Model Editing and Upcycling
BrainSurgery introduces a declarative framework for robust neural network weight manipulation, addressing challenges in model editing and upcycling. The tool abstracts storage formats and memory management, enabling complex transformations via YAML plans with regex-based tensor targeting. It supports structural modifications, mathematical operations, and reshaping while validating tensor properties through built-in assertions. Demonstrated across four examples and three case studies (including LoRA extraction), the system ensures reproducible operations. This approach mitigates fragility in ad-hoc Python workflows for tasks like precision casting and low-rank factorization.
tensor surgerymodel upcyclingdeclarative transformationslora extractionweight manipulation
When Do Local Score Models Extrapolate Across Size? A Diagnostic Theory and Benchmark
The paper establishes a theoretical framework for when local score models can extrapolate across system sizes in scientific generative modeling. It demonstrates that stable size extrapolation depends not on architectural locality alone, but on the quasi-locality of Gaussian-smoothed scores, mediated through Tweedie's formula and posterior covariance. The authors introduce Finite-Depth Local Flow (FDLF), a diagnostic benchmark with exact scores and controllable response ranges, to empirically validate that spatial mixing preserves quasi-locality relative to model receptive fields, enabling successful size transfer, while weakened mixing degrades locality and causes failure.
score-based generative modelssize extrapolationquasi-localitytweedie's formulaspatial mixing
What the Eyes See, the LLMs Miss: Exploiting Human Perception for Adversarial Text Attacks
The paper introduces Human-Perceptible Adversarial Attacks (HPAA), a class of typographic manipulations that embed harmful content in visually salient but machine-evading forms, exploiting the perceptual mismatch between human and LLM-based moderation systems. Operating in black-box settings with minimal queries, HPAA combines spacing, visual emphasis, and spatial arrangement to achieve 86% human recognition while reducing detection rates below 1% across ten commercial and open-source moderation systems. Ablation studies identify typographic factors driving evasion, revealing a critical blind spot in current LLM-based moderation architectures.
adversarial attackscontent moderationtypographic manipulationhuman perceptionblack-box evasion
AutoMegaKernel: A Statically-Checked Agent Harness for Self-Retargeting Megakernel Synthesis
AutoMegaKernel (AMK) introduces a statically-verified system for compiling Llama-family models into single persistent CUDA megakernels, eliminating hand-written CUDA. It employs a frozen schedule-IR validator to ensure deadlock/race freedom, rejecting 6,091/7,160 adversarial schedules with zero false accepts. The system auto-generates correct megakernels for 10/10 tested models, achieving perplexity matching (2.5e-7) on SmolLM2-135M. An agent-driven loop improves performance (1.25-1.72x), while W8A16 megakernels outperform cuBLAS bf16 on inference-class GPUs (L4:1.33x, L40S:1.25-1.27x). Bottlenecks occur on training-class A100/H100 due to cross-SM sync.
megakernelstatic verificationw8a16cudallama-family
Algorithm for Contextual Queueing Bandits with Rate-Optimal Queue Length Regret
(No summary returned.)
In-Context Learning for Latent Space Bayesian Optimization
The paper proposes adapting tabular foundation models for latent-space Bayesian optimization (LSBO) by augmenting their pretraining with synthetic optimization tasks on molecular VAE latent spaces. The method introduces a regularization term during continued pretraining to maintain the model's general regression capabilities while specializing for LSBO. Evaluations on molecular optimization benchmarks demonstrate improved performance, validating the need for LSBO-specific adaptation in in-context learning surrogates.
latent-space bayesian optimizationtabular foundation modelsin-context learningmolecular vaepretraining adaptation
A Unifying Framework for Concept-Based Representational Similarity
The paper introduces a unifying framework for concept-based representational similarity, decomposing alignment along two axes: what is aligned (representations vs. concepts) and at what level (instance-wise vs. distributional). This yields four properties—instance-wise and distributional variants of translation and concept consistency—clarifying guarantees of existing methods. The authors propose InterVenchA, an intervention-based benchmark, and Coupled Sparse Autoencoder (CoSAE), which jointly enforces complementary alignment objectives. Experiments show that optimizing one property does not recover others, and 0.1% paired data suffices for instance-level alignment when anchoring distributional objectives. Concept alignment is fundamentally multi-objective.
representational similarityconcept alignmentinstance-wise alignmentdistributional alignmentcoupled sparse autoencoder
Data-driven discovery of governing differential equations across physical systems
The article proposes a problem-oriented framework for data-driven discovery of governing differential equations, introducing a two-dimensional phase diagram categorizing problems by structural and coefficient complexity. It presents the representation-evaluation-optimization (REO) framework as a unifying abstraction, shifting focus from individual algorithms to fundamental discovery principles. The analysis reveals methodological trends from sparse equations to complex laws, with applications in physics and adjacent sciences, highlighting future challenges in theory revision and mechanism distillation.
differential equationsdata-driven discoveryphase diagramrepresentation-evaluation-optimizationscientific discovery
Constrained user-item allocation for e-commerce marketing campaigns
The paper introduces an auto-targeting framework for joint user-item allocation in e-commerce campaigns, addressing the limitations of decoupled approaches. Three methods are proposed: (i) constrained spectral biclustering to identify dense affinity regions, (ii) greedy local search with swaps for combinatorial refinement, and (iii) multi-armed bandits for exploration. Evaluations on synthetic data, Amazon Reviews, and proprietary datasets show biclustering achieves superior campaign quality (measured by lift and fairness), though scalability favors bandit methods on large datasets.
spectral biclusteringmulti-armed banditcombinatorial optimizationuser-item affinitycampaign lift
Assessing Sample Quality in Conditional Generation under Compositional Shift
The paper introduces a post-hoc trust score for evaluating conditional generation under compositional shift, where reference distributions are unavailable. The method combines global realism (manifold compatibility) and attribute-wise faithfulness (proximity to requested attributes) to assess sample quality. Results demonstrate improved morphological structure preservation in biological imaging and enhanced downstream performance, with applicability to off-the-shelf models and early abstention during generation.
conditional generationcompositional shiftmanifold compatibilityattribute-wise faithfulnesspost-hoc evaluation
On Choosing the $μ$ Parameter in Gaussian Differential Privacy
The paper establishes principled mappings from pure differential privacy (DP) ε to Gaussian DP (GDP) μ by aligning worst-case membership inference attack success across three metrics: multiplicative advantage at fixed false positive rate (FPR), precision-recall tradeoffs, and privacy profiles. The authors derive conversion rules through adversarial success equivalence and tabulate recommended μ values for practical use. Results indicate μ≈ε/5 serves as a conservative general-purpose conversion factor between these privacy frameworks.
gaussian differential privacymembership inferenceprivacy-preserving machine learningprivacy profilepure-dp
Code Is More Than Text: Uncertainty Estimation for Code Generation
The paper proposes a code-specific uncertainty estimation (UE) framework addressing three unique properties of code generation: token fragility, intent-code gap, and executability. It introduces three orthogonal uncertainty axes—lexical (Top-K token entropy), algorithmic (pseudo-code consistency), and functional (behavioral consistency)—evaluated across five code LLMs. The ensemble improves AUROC by 8.1 points (0.696 to 0.776) over NL-derived baselines, with Top-K token entropy matching multi-pass performance at 3x lower cost on Qwen3-14B.
uncertainty estimationcode generationtoken fragilityintent-code gapbehavioral consistency
Integrating gene regulatory priors into Transformer attention with scTransformer for interpretable scRNA-seq analysis
scTransformer integrates gene regulatory priors into Transformer attention mechanisms for single-cell RNA-seq analysis, enhancing both interpretability and performance. The model constrains information flow based on known regulatory structures, producing biologically meaningful representations. Evaluated on a disease-relevant single-nucleus RNA-seq dataset, scTransformer improves supervised cell-type classification accuracy, enhances cell-type separation in embedding space, and generates attention patterns consistent with established regulatory programs. This approach demonstrates that embedding biological structure into Transformer models can yield more interpretable and robust foundation models for single-cell omics.
transformergene regulatory priorssingle-cell rna-seqattention mechanismsinterpretability
Streaming Interventions: Can Video Large Language Models Correct Mistakes as They Occur?
The paper introduces Ego-MC-Bench, a benchmark for evaluating reactive task guidance in cooking scenarios, revealing significant challenges for state-of-the-art video LLMs due to limited training data on mistake interventions. To address this, the authors propose Ego-CoMist, a synthetic dataset created by transforming non-interactive cooking videos into supervised examples with proactive interventions. Fine-tuning on Ego-CoMist improves performance, particularly for smaller, edge-device-suitable video LLMs.
video llmstask guidancemistake correctionsynthetic datasetedge devices
Automating the Expert Eye: A System-Agnostic Deep Learning Framework for Rare Event Discovery in Imbalanced Force Spectroscopy
A system-agnostic deep learning framework automates rare event discovery in Single-Molecule Force Spectroscopy (SMFS) data, addressing extreme class imbalance. The method employs a modified ResNet18 architecture with 1D-to-2D rasterized geometric matrices and an asymmetric Focal Loss objective. Evaluated on R. champanellensis cellulosome data with 1.34% target interactions (13 true events out of 970 traces), the model achieved 0.9196 accuracy and 0.9231 recall. A dual-threshold triage system reduced manual curation workload by over 90%, preserving rare data. Grad-CAM validated decision interpretability by localizing on structural unbinding regions. The open-source tool enables scalable molecular discovery in biophysics.
single-molecule force spectroscopyclass imbalanceresnet18focal lossgrad-cam
Efficient Traffic Prediction at Scale: A Systematic Study of STGCN Architectural Depth
This study investigates architectural depth in Spatio-Temporal Graph Convolutional Networks (STGCN) for traffic prediction, challenging the assumption that deeper models yield superior performance. Through systematic experiments across four datasets, the authors compare 1-block, 2-block, and 3-block STGCN variants. Results demonstrate that the 1-block architecture achieves optimal short-term prediction (10 minutes) on three datasets, with ≤1.8% relative error degradation at longer horizons, while reducing CPU inference latency by 61% and increasing throughput by 37% compared to the standard 2-block variant. The 3-block architecture offers negligible accuracy gains (<0.5%) at over double the computational cost, suggesting widespread over-parameterization in current STGCN implementations.
spatio-temporal graph convolutional networktraffic predictionarchitectural depthinference latencyover-parameterization
Investigating Calibration Challenges in Probabilistic Electricity Price Forecasting
The study identifies a critical calibration gap in probabilistic electricity price forecasting, where current scoring rules excessively prioritize sharpness over reliability, yielding overconfident uncertainty estimates. Through theoretical and empirical analysis, the authors demonstrate that neglecting calibration transforms probabilistic models into mere proxies for deterministic forecasts. They advocate for calibration-aware objectives and architectures to preserve distributional integrity in energy market predictions amid increasing renewable-induced volatility.
probabilistic forecastingcalibrationscoring rulesmarket volatilityuncertainty estimation
BUDDY: BUdget-Driven DYnamic Depth Routing for Adaptive Large Language Model Inference
The paper introduces BUDDY, a budget-driven dynamic depth routing framework for adaptive LLM inference that addresses two limitations of existing depth pruning methods: (1) inflexible budget control and (2) static routing paths during decoding. BUDDY employs a lightweight Decision Module to score and select top-k Transformer layers per input, reuses first-layer KV cache for efficient context-aware rerouting, and optionally predicts compute budgets via a Budget Predictor. Experiments on Llama-family and Qwen models demonstrate competitive performance versus static pruning baselines while enabling strict budget control, decode-time adaptation, and multi-budget support within a single model.
depth pruningkv cachetransformer blocksbudget controldecode-time adaptation
Loss-Guided Adaptive Scale Refinement for Molecular Force Prediction
The study introduces a loss-guided adaptive scale refinement framework for molecular force prediction, addressing the limitations of fixed-scale modeling in molecular representation learning. The method employs interpolation, routing, differentiable scale updates, and scale pool refinement to dynamically adjust modeling scales. Evaluated on a NaCl aqueous ionic system, the framework reduces the overall force MAE from 399.65 to 381.23 and the close-contact MAE from 327.22 to 260.51 in regimes with nearest-ion distance below 0.6 nm. The final scale pool {0,0.125,0.25,0.375,0.5,0.75,1} demonstrates near-continuous oracle performance, validating adaptive scale refinement as an effective approach.
molecular force predictionadaptive scale refinementdifferentiable scale updatesscale pool refinementclose-contact mae
Report the Floor: A Training-Free Conformal Interval Is a Mandatory Baseline for Probabilistic Time-Series Forecasting
The paper demonstrates that ConformalNaive, a training-free conformal interval method using last-value point forecasts and split-conformal residual quantiles, outperforms existing baselines in probabilistic time-series forecasting. Evaluated across 2,217 real series from nine public datasets, it surpasses naive quantile baselines, NPTS variants (73% and 64% of series), and Conformal Seasonal Pools (71% of series), while matching simpler learned conformal predictors. It also shows better calibration than trained neural forecasters (84-85% vs. 66% coverage at nominal 95%). The authors introduce ConformalNaive+, a horizon-adaptive variant, and advocate for mandatory inclusion of conformal naive baselines in future evaluations.
conformal predictionprobabilistic forecastingtime-seriessplit-conformalwinkler score
Escaping the KL Agreement Trap in On-Policy Distillation
The paper identifies a 'KL agreement trap' in on-policy distillation (OPD), where teacher-student agreement on degenerate prefixes produces low reverse KL divergence but ineffective supervision. It proposes KL Agreement Trap Termination (KAT), an online termination rule that detects persistent low-KL agreement using dynamic thresholds. Evaluated on four mathematical benchmarks, KAT improves average@k accuracy by 2.66% and pass@k by 3.43%, while reducing rollout length by 59.73% by filtering weak supervision signals.
on-policy distillationkl divergencesupervision signalsrollout terminationmathematical reasoning
Breaking the Tokenizer Barrier: On-Policy Distillation across Model Families
The paper introduces a method for cross-tokenizer On-Policy Distillation (OPD), enabling knowledge transfer between Large Language Models (LLMs) with different tokenizers. By developing a token-mapping algorithm, the approach preserves high-fidelity token-level signals during distillation, overcoming the tokenizer barrier in traditional OPD. Experiments demonstrate superior compute-efficiency over baselines across multiple benchmarks, expanding the range of viable teacher-student pairs for LLM enhancement.
on-policy distillationlarge language modelstoken-mapping algorithmcross-tokenizerknowledge transfer
Dense Force Estimation with an Event-based Optical Tactile Sensor
The paper presents the first framework for dense 3D force field reconstruction using event-based optical tactile sensors, addressing limitations of vision-based sensors in temporal resolution and motion blur. The method combines event-based marker tracking for shear displacement estimation with a convolutional neural network for normal displacement prediction, followed by inverse Finite Elements Method (iFEM) force mapping. Experiments demonstrate accurate force reconstruction with mean absolute errors of (0.14 N, 0.10 N, 0.93 N) across force ranges up to (4 N, 4 N, 20 N) at 100 Hz, enabling high-frequency tactile feedback for robotic manipulation.
event-based sensingtactile sensorforce reconstructionfinite elements methodmarker tracking
Operator learning for solving Fokker-Planck equations with various initial conditions
The authors propose a conditional normalizing flow-based physics-informed neural network (PINN) framework for learning solution operators of Fokker-Planck equations (FPE) across arbitrary initial conditions. The method reformulates the problem using Chapman-Kolmogorov equations, employs linearized SDE PDFs as base distributions for normalizing flows, and introduces time-weighted loss functions to address small-time numerical instabilities. Numerical experiments demonstrate the approach's effectiveness in approximating transition PDFs from Dirac delta initial conditions while maintaining stability across varying time scales.
fokker-planck equationnormalizing flowsphysics-informed neural networksstochastic differential equationsoperator learning
Graph Mamba Operator: A Latent Simulator for Interacting Particle Systems
The Graph Mamba Operator (GraMO) introduces a latent-space simulator for interacting particle systems, integrating state-space models with graph-based interaction learning to address error accumulation in long-horizon predictions. Unlike prior approaches that separate spatial and temporal dynamics, GraMO couples graph-based interactions and temporal state updates within a single recurrence, using input-dependent coefficients for adaptive regime transitions. Evaluated on N-body systems, motion capture, and robotics datasets, GraMO achieves the lowest error across benchmarks and the largest gains in long-horizon prediction.
graph neural networksstate-space modelslatent-space simulatorlong-horizon predictioninput-dependent coefficients
Now You (Still) See Me: Detecting Evasive Steganographic Payloads in LLMs
The paper demonstrates that activation-based steganography detection in LLMs can be systematically evaded through adversarial fine-tuning, but detectability can be restored via targeted data interventions. The authors extend detection using non-linear MLP probes and adversarially fine-tune steganographic trojans across five base models (Qwen3-8B, Llama-3.1-8B, Ministral-8B, Qwen3-14B, Phi-4-14B), achieving 58-79% secret recovery while evading probes with minimal capability degradation (1-8%). An information-theoretic analysis reveals payload concealment via synergistic interactions with residual degrees of freedom, mitigated by a recontextualization dataset that restores detectability for all evasive trojans.
steganographylinear probesadversarial fine-tuninginformation-theoretic characterizationrecontextualization dataset
Benchmarking Empirical Privacy Protection for Adaptations of Large Language Models
This work benchmarks empirical privacy protection in differentially private (DP) adaptations of large language models (LLMs), addressing gaps between theoretical guarantees and practical effectiveness. Using robust membership inference and canary data extraction attacks, the study systematically evaluates privacy risks across adaptation data distributions, from exact overlaps with pretraining to out-of-distribution (OOD) cases, and compares adaptation methods and privacy regimes. Results indicate that distribution shifts significantly impact vulnerability, with closer adaptation-pretraining alignment increasing risk despite DP guarantees. Parameter-efficient methods like LoRA offer superior empirical protection for OOD data. The benchmark provides actionable insights for deploying customized LLMs in sensitive settings.
differential privacylarge language modelsmembership inferenceparameter-efficient fine-tuningout-of-distribution
PriFT: Prior-Support Guided Supervised Fine-Tuning
PriFT introduces prior-support guided supervised fine-tuning (SFT) to improve generalization by deriving token weights from a frozen pretrained model rather than the fine-tuned model, avoiding self-reinforcing dynamics. The method estimates prior support for target tokens using pretrained token probability (PriFT-prob) or cumulative probability mass (PriFT-mass). Experiments on mathematical reasoning, code generation, and medical question answering demonstrate that PriFT achieves state-of-the-art SFT performance and provides better initialization for reinforcement learning compared to existing token-reweighting approaches.
supervised fine-tuningtoken reweightingprior supportpretrained distributionreinforcement learning
Distilling Safe LLM Systems via Soft Prompts for On Device Settings
The paper introduces a parameter-efficient method for safety alignment in on-device LLM deployment through soft prompt distillation. It systematically evaluates LLM architectures, training objectives, and parameter-efficient fine-tuning approaches, proposing distillation frameworks based on total variation and KL divergence to transfer safety behaviors from guard models to soft prompts. Results show superior safety-usefulness trade-offs compared to LoRA adapters, steering vectors, and direct optimization, with minimal memory and compute overhead at inference.
soft promptssafety alignmentparameter-efficient fine-tuningdistillation frameworkson-device deployment
Zero-Shot Semantic Re-Identification for Autonomous Driving: A VLM Baseline Study
The study proposes a zero-shot Vision-Language Model (VLM) pipeline for semantic re-identification (ReID) in autonomous driving, replacing traditional visual embeddings with structured textual descriptions. The method generates attribute-based descriptions (category, color, shape, etc.) for traffic participants and evaluates their matching performance across observations. Results show comparable retrieval accuracy to supervised CNN baselines while offering improved interpretability, though challenges persist in attribute consistency and fine-grained discrimination.
vision-language modelszero-shot learningsemantic re-identificationautonomous drivingattribute-based matching
PBSD: Privileged Bayesian Self-Distillation for Long-Horizon Credit Assignment
PBSD (Privileged Bayesian Self-Distillation) introduces a Bayes-calibrated self-distillation method for fine-grained credit assignment in long-horizon agentic tasks with sparse final rewards. The approach leverages a privileged answer-conditioned teacher model to compute a Bayesian evidence score, which is autoregressively decomposed into turn-level signals indicating whether intermediate steps support or undermine the verified outcome. This enables principled reweighting of sparse supervision into turn-level credit signals, compatible with standard policy optimization. Experiments show PBSD enhances performance in both in-domain and out-of-domain settings, facilitating effective policy learning and improved generalization, particularly in transferring knowledge from short-context training to long-context inference.
bayesian evidence scorecredit assignmentself-distillationpolicy optimizationautoregressive decomposition
Thresholded Local Hyper-Flow Diffusion
The paper introduces Thresholded Local Hyper-Flow Diffusion (TL-HFD), a first-order method for seeded clustering in submodular hypergraphs that maintains locality during computation. TL-HFD performs projected subgradient updates on an active region around seeds and expands via thresholded boundary activation, proving exactness of local updates and finite-time dual suboptimality. Theoretical guarantees include additive activated-volume bounds and robust sweep-cut guarantees for early-stopped iterates. Empirical results show TL-HFD matches or outperforms HFD with reduced activated volume, particularly on noisy instances.
submodular hypergraphsseeded clusteringprojected subgradientthresholded activationsweep-cut guarantee
Machine-Learning Emulation of Satellite Greenhouse Gas Retrievals: Stability over Time
The study evaluates temporal stability in machine-learning emulators for satellite-based greenhouse gas retrievals, demonstrating performance degradation when test periods diverge from training data. Using GOSAT satellite measurements, the authors show that incorporating time as an input feature improves XCH4 prediction accuracy in Lasso and neural-network models, with Lasso outperforming more complex methods in stability. Validation against TCCON ground measurements reveals time-augmented Lasso achieves errors comparable to GOSAT-TCCON discrepancies for both XCO2 (1.6 ppm) and XCH4 (8 ppb).
greenhouse gas retrievalemulator stabilitylasso regressiongosattccon validation
Toward Compiler World Models: Learning Latent Dynamics for Efficient Tensor Program Search
The paper introduces a world-model-inspired evaluator for tensor program optimization that models schedule evaluation as action-conditioned latent dynamics over program states. The method employs a lightweight transition model to roll out scheduling actions in continuous latent space, avoiding expensive AST mutations and repeated code encoding. Results show 1.37×-1.54× latency improvements over Ansor on GPU/CPU, matches Ansor-10K with 10× fewer measurements, and achieves 4.61×/3.67× speedup over PyTorch variants.
tensor program optimizationlatent dynamicsauto-scheduleraction-conditionedcontinuous latent space
SG-OPD: Sign-Gated On-Policy Distillation via Sign-Consistency Gating and Phased Teacher Sampling
The paper introduces Sign-Gated On-Policy Distillation (SG-OPD), addressing limitations in standard on-policy distillation (OPD) by mitigating trajectory misalignment and unreliable teacher preferences. SG-OPD employs a binary verifier for two mechanisms: phased teacher sampling integrates verified teacher rollouts during cold-start, while sign-consistency gating adjusts distillation updates based on teacher-verifier agreement. Evaluated on competition-level mathematical reasoning benchmarks, SG-OPD outperforms standard OPD by 1.98 and 7.50 average gains at per-sample and per-question levels, respectively.
on-policy distillationsign-consistency gatingphased teacher samplingbinary verifiermathematical reasoning
PRISM: Topology-Aware Cross-Modal Imputation for Modality-Deficient Federated Graph Learning
PRISM introduces a topology-aware federated cross-modal imputation framework for modality-deficient federated graph learning (MM-FGL), where clients lack entire modalities. The method proactively retrieves missing-modality semantics from the federation and integrates them into local graph propagation under topology-aware control, addressing errors amplified by graph structures. Evaluations on six multimodal graph datasets demonstrate PRISM's superiority, outperforming baselines by 4.48% on average across graph-centric and modality-centric tasks.
federated graph learningcross-modal imputationmodality deficiencytopology-awaremessage passing
Intention Driven Identification of In-Possession Match Phases in Association Football through Temporal Graph Learning
The study introduces a Temporal Graph Attention Network (T-GAN) framework for identifying in-possession match phases in football from 25 Hz TRACAB tracking data, based on three tactical intentions and six phases. The method combines player-interaction graphs, contextual features, and Transformer-based temporal modeling, evaluated via frame-level F1 and sequence-aware IoT-D metrics. T-GAN achieved macro-average F1 scores of 0.87 for intentions, 0.76 for invasion phases, and 0.79 for scoring phases, with sequence modeling and graph-based relational features proving critical for performance.
temporal graph attention networkin-possession phasestactical intentionsintersection over truth-dominanceplayer-interaction graphs
Trajectory Geometry of Transformer Representations Across Layers
This work introduces a geometric framework for analyzing transformer representations across layers, employing five metrics: trajectory length, curvature, semantic convergence index, layerwise cosine similarity, and representational stability. The method recasts the transformer forward pass as a discrete population trajectory through a high-dimensional manifold, leveraging tools from computational neuroscience. Experiments across GPT-2, TinyLlama, and Qwen2.5 reveal four key findings: semantic convergence in middle-to-late layers (peak CI 0.41--0.58), higher curvature for reasoning tasks (0.71--0.83 rad), trajectory bifurcation for ambiguous tokens (5.6x separation), and a universal three-phase layerwise structure. Results are validated against shuffled-layer and random-embedding controls.
transformertrajectory geometryrepresentational stabilitysemantic convergencelayerwise cosine similarity
ERBench: A Benchmark and Testsuite for Equation Discovery Algorithms
The paper introduces ERBench, a benchmark and testsuite for evaluating equation discovery algorithms through symbolic regression. Current benchmarks lack comprehensive assessment of robustness across varying dimensionality, sampling conditions, and domain shifts, focusing instead on limited groundtruth formula recovery. ERBench addresses this gap by emphasizing systematic evaluation of algorithm performance under diverse data conditions, prioritizing equation recovery as a proxy for discovery capability in scientific modeling.
symbolic regressionequation discoverybenchmark evaluationgroundtruth recoverysampling robustness
Multi-View Speech Representation Learning for Parkinson's Disease Detection Using Context-guided Cross-modal Attention
A multi-branch deep learning framework is proposed for Parkinson's disease (PD) detection from speech, leveraging complementary speech representations to capture pathological information. The method processes 5-second speech segments using three modalities: Log-Mel spectrograms (ResNet-18), MFCC sequences (BiLSTM), and HuBERT embeddings. A context-guided cross-modal attention mechanism dynamically weights HuBERT embeddings based on global acoustic context from spectrograms and MFCCs. Evaluated on the Spanish PC-GITA corpus under speaker-independent 5-fold cross-validation, the architecture achieves 91.51% accuracy, 91.24% F1-score, and 95.97% AUC. Ablation studies confirm the contributions of the attention mechanism and multi-modal integration.
cross-modal attentionhubert embeddingslog-mel spectrogramsbilstmhypokinetic dysarthria
Orange Lab: Lowering Barriers to Data Mining through Embedded Interactive Workflows
Orange Lab introduces component exposition, a paradigm enabling visual workflow components to be embedded in web contexts while hiding underlying complexity, facilitating interactive data exploration. The web-based environment allows collaborative construction of reactive machine learning workflows from modular components, with interactions propagating dynamically through pipelines. Deployments in data literacy education demonstrate reduced entry barriers, enabling hands-on ML concept exploration without system expertise.
visual programmingworkflow embeddingreactive systemsdata literacymodular components
SNN-MLIR: An MLIR Dialect for Compiling Neuromorphic SNNs from NIR to Bare-Metal C
The paper introduces SNN-MLIR, an MLIR dialect for compiling spiking neural networks (SNNs) from the Neuromorphic Intermediate Representation (NIR) to bare-metal C code. The dialect provides type-polymorphic operations supporting both floating-point and quantized data, enabling a unified intermediate representation for simulation and hardware deployment. A Python front-end processes NIR files, automatically inserting rescaling operations for quantization consistency, while a lowering pass converts the dialect to standard linalg and arith operations for C11 code generation. Evaluation confirms numerical fidelity, CPU portability, and quantization cost. The toolchain currently supports feedforward, fully-connected networks and is open-sourced under Apache-2.0 with LLVM-exception.
spiking neural networksmlir dialectneuromorphic intermediate representationquantizationbare-metal compilation
The Injection Paradox: Brand-Level Suppression in Safety-Trained LLM Recommendations via RAG Context Injection
This work identifies the Injection Paradox, a reproducible failure mode in safety-trained Retrieval-Augmented Generation (RAG) LLM recommendation systems where prompt injections suppress brand recommendations below baseline levels. Through experiments with Claude Opus 4.6, the authors demonstrate that a single injected document reduces the target brand's top-2 recommendation rate from 54% to 0% across 50 trials, with suppression extending to unmodified documents of the same brand. Counterfactual experiments confirm this directional pattern across three brands. Contrasting results in GPT models, where injections increase recommendations, highlight model-family differences in injection handling. These findings suggest potential reverse-attack scenarios exploiting safety-sensitive model behaviors.
retrieval-augmented generationprompt injectionsafety trainingrecommendation systemsmodel-family differences
Asymptotic Optimality of Thompson Sampling for Risk-Averse Bandits with Sub-Gaussian Rewards
The paper establishes the asymptotic optimality of $ρ\text{-}\mathrm{NPTS}_{\mathrm{SG}}$, a nonparametric Thompson Sampling algorithm for risk-averse bandits with sub-Gaussian rewards. By leveraging a discretization lemma for bounded support and a truncated variant for sub-Gaussian tails, the algorithm achieves instance-dependent regret bounds matching the lower order in $\log n$ for any continuous risk functional $ρ$, including CVaR and Sharpe ratio. This result holds under weaker conditions than prior work, requiring only continuity of $ρ$ rather than dominance or Lipschitz constraints. The proof relies on Dirichlet posterior projections to overcome super-exponential barriers in prior analyses.
thompson samplingrisk-averse banditssub-gaussian rewardsinstance-dependent regretdirichlet posterior
Counterfactual Reasoning for Fine-Grained Evidence Disentanglement in VideoQA
The paper introduces CREDiT, a counterfactual reasoning framework for fine-grained evidence disentanglement in VideoQA, addressing spurious correlations in multimodal models. The method employs a structural causal model to decompose cross-modality representations into causal and non-causal components, using feature-level interventions and counterfactual inputs to suppress non-causal correlations. Evaluations on NExT-GQA, SportsQA, and SPORTU-video show improved accuracy and reasoning reliability, particularly in complex scenarios.
counterfactual reasoningevidence disentanglementstructural causal modelcross-modality representationsfeature-level interventions
Improved Convergence Analysis of Topology Dependence in Decentralized SGD
We present an improved convergence analysis of Decentralized Stochastic Gradient Descent (SGD) that precisely characterizes the impact of network topology on convergence rates. Unlike prior analyses that relied solely on the spectral gap, our method incorporates all eigenvalues of the mixing matrix to derive tighter bounds. Through empirical evaluation, we demonstrate that our analysis more accurately captures topology-dependent convergence behavior, particularly in heterogeneous settings where prior work observed significant experimental impacts. These findings advance theoretical understanding of decentralized optimization by elucidating the nuanced relationship between topology and convergence dynamics.
decentralized sgdspectral gapmixing matrixconvergence rateheterogeneous learning
Claw-R1: A Step-Level Data Middleware System for Agentic Reinforcement Learning
Claw-R1 introduces a step-level data middleware system for agentic reinforcement learning (RL), addressing the gap in managing the full lifecycle of agent-environment interactions. The system connects heterogeneous agent runtimes with RL training backends via a Gateway Server and a Data Pool. The Gateway Server captures multi-turn interactions through a unified LLM API, while the Data Pool organizes step-level records with metadata such as prompt IDs, response IDs, and rewards. Users can inspect live trajectories, curate data, and configure training-ready batches. Claw-R1 treats agent interaction traces as managed data assets, emphasizing the importance of data management in agentic RL.
agentic reinforcement learningstep-level datamiddleware systemgateway serverdata pool
Counterfactual Transport Flows for Offline Conservative Trajectory Refinement
The paper introduces counterfactual transport flows, a trajectory refinement framework for offline reinforcement learning that improves behavior from logged data without unsafe extrapolation. The method constructs local preference pairs by retrieving higher-feedback trajectories near a candidate trajectory in latent space, using them as weak supervision for conservative refinement. A refinement strength parameter controls the improvement-preservation trade-off at inference. Experiments on D4RL benchmarks (AntMaze, MuJoCo) demonstrate improved behavior from historical returns while providing interpretable refinement paths.
offline reinforcement learningtrajectory refinementcounterfactual transportweak supervisiond4rl benchmarks
Driving Video Retrieval for Complex Queries with Structured Grounding
STRIVE-D introduces a data-calibrated retrieval framework for driving videos, addressing limitations of vision-language and keyword-based methods in capturing dynamic events like cut-ins and hard braking. The method leverages weakly labeled in-domain videos to assess query rule reliability, adapt mismatched rules, and fuse calibrated rule scores with existing retrieval signals. Evaluated on three benchmarks including DrivingDojo, STRIVE-D achieves up to 84% relative improvement in top-1 accuracy over state-of-the-art approaches.
video retrievalweakly supervised learningrule adaptationautonomous drivingdata calibration
RAM: Reachability Across Morphologies
The paper introduces Reachability Across Morphologies (RAM), a morphology-conditioned implicit neural representation that serves as a fast, differentiable surrogate for pose reachability in robotics. RAM generalizes to unseen morphologies while inherently accounting for self-collisions, trained on a large-scale dataset of 3e10 samples generated from forward kinematics. Experiments demonstrate an 86% F1-score at nanosecond inference, outperforming baselines by 14% with three orders of magnitude faster inference, and enabling significant speed-ups in gradient-based morphology and trajectory optimization.
implicit neural representationreachability analysismorphology synthesisforward kinematicsself-collision detection
Alcmean's: Unsupervised community detection using local Laplacian, automatic detection of the number of centers
We propose Automatic Laplacian Centrality Means (ALCMeans), an unsupervised community detection algorithm that combines Laplacian energy-based center identification with DeepWalk embeddings for robust node representation. ALCMeans eliminates the need to predefine the number of communities, enhances cluster center selection using structural importance, and leverages representation learning for accurate assignments. Experiments on benchmark datasets show 10-20% higher NMI and ARI scores compared to Louvain, Newman-Girvan, LPA, Fast-Greedy, and MAGI. Additional evaluations with modularity and F1-scores confirm ALCMeans' superiority, making it a promising tool for real-world network analysis despite increased runtime relative to lightweight heuristics.
community detectionlaplacian energydeepwalk embeddingsnode representationstructural importance
From Shortcuts to Reasoning: Robust Post-Training of Theory of Mind with Reinforcement Learning
The study addresses shortcut learning in Theory of Mind (ToM) tasks by proposing Thinking-RFT, a Reinforcement Fine-Tuning method with verifiable rewards and explicit reasoning chains. The authors first analyze ToM datasets, identifying shortcut-prone questions (e.g., belief tracking) versus those requiring deeper reasoning (e.g., intention). Evaluating on four shortcut-free datasets, Thinking-RFT improves ToM performance by 6% over Supervised Fine-Tuning (SFT), with notable gains in higher-order reasoning (10%) and multimodal cases (7%). The method leverages joint reasoning and RL, outperforming Non-Thinking-RFT by 7%, and grounds reasoning in causal anchor cues.
theory of mindreinforcement fine-tuningshortcut learninghigher-order reasoningmultimodal learning
Stabilizing On-Policy Distillation for MLLM Reasoning with Global Normalization
The paper introduces Globally Normalized Distillation Policy Optimization (GNDPO), a method to stabilize on-policy distillation (OPD) for multimodal large language models (MLLMs). GNDPO addresses gradient instability in token-level distillation by normalizing KL scores into batch-level relative advantages, preventing gradient explosions while preserving fine-grained supervision. Experiments demonstrate that GNDPO enhances training robustness and improves performance on multimodal reasoning tasks compared to naive OPD approaches.
on-policy distillationgradient instabilitykl scoresmultimodal reasoninggndpo
Beyond FLOPs: Benchmarking Real Inference Acceleration of LLM Pruning under a GEMM-Centric Taxonomy
The paper introduces a GEMM-centric taxonomy for reorganizing LLM pruning methods along the M, N, and K dimensions of matrix multiplication, enabling systematic comparison of their inference acceleration. A unified benchmarking framework evaluates the acceleration-quality Pareto frontier across pruning families, revealing that static depth pruning achieves the strongest Pareto-optimal performance and remains closest to theoretical upper bounds in memory-constrained scenarios. Results show distinct regime transitions: static depth dominates at 0-4% quality loss, dynamic depth at 5-16%, and static width pruning at 17-26% loss.
llm pruninggemm taxonomyinference accelerationpareto frontierstatic depth pruning
The Hidden Bias of Process Reward Models:PRISM for Rewarding the Right Reasoning
The paper identifies a hidden bias in Process Reward Models (PRMs) caused by imbalanced step-level training data, leading to overcrediting of plausible but incorrect steps and high false-positive rates. To address this, the authors propose PRISM (Precision Ranking for Improved Step Modeling), a policy-aware PRM training framework that uses contrastive step-level comparisons and hard negatives generated via temporal lookahead, without requiring additional human labels. PRISM reduces false positives by 22% on PRMBench and improves macro F1, while also enhancing accuracy in policy optimization tasks (up to 33% for Best-of-N selection).
process reward modelscontrastive learningfalse-positive ratepolicy optimizationhard negatives
Neural Legendre-Fenchel transform with Hessian Preconditioning
The paper introduces a Hessian-based preconditioning strategy for neural Legendre-Fenchel (LF) transform to address ill-conditioned functions. Building on the projective polarity reformulation of LF transform, the method applies an affine deformation around a minimizer to align the second-order Taylor approximation with the canonical paraboloid. A residual network then learns this simplified mapping, with the original conjugation recovered via inverse deformation. Experiments on diverse convex functions, including high-dimensional cases, show improved convergence and numerical accuracy, particularly for ill-conditioned problems. The approach requires only modest computational overhead.
legendre-fenchel transformhessian preconditioningconvex conjugatesaffine invarianceresidual network
MilliVid: Hierarchical Latents for Long-Range Consistency in Video Generation
MilliVid introduces a hierarchical latent approach for long-range consistency in video generation, addressing the impracticality of long transformer sequence lengths. The method pre-trains an autoencoder to compress frames into multi-scale tokens, ranging from coarse scene layout to fine texture details, and trains a video diffusion model for coarse-to-fine token generation. By controlling detail levels during rollout, it preserves geometry and object permanence while reducing compute for less perceptually relevant details. Evaluated on a long Minecraft video dataset, MilliVid achieves substantially more consistent rollouts compared to existing baselines.
hierarchical latentvideo diffusion modelcoarse-to-fine rolloutmulti-scale tokenobject permanence
Beyond Convolution: Advancing Hypergraph Neural Networks with Hypergraph U-Nets
This work introduces Hypergraph U-Nets with Parallel Hierarchical Pooling (PHPool) and Unpooling (PHUnpool) operators, addressing the lack of effective pooling methods for hypergraph data. PHPool constructs pooling operators globally via dendrogram cuts at multiple granularities, preserving structural integrity, while PHUnpool performs inverse operations for reconstruction. Evaluated on hypergraph reconstruction, classification, and anomaly detection tasks, the method outperforms state-of-the-art graph and hypergraph learning approaches.
hypergraph neural networksu-net architectureparallel hierarchical poolingdendrogram clusteringanomaly detection
Data augmented bootstrap: Unifying confidence interval construction by approximate invariance
The authors introduce Data Augmented Bootstrap (DAB), a unified framework for constructing confidence intervals leveraging approximate invariance transformations of data. DAB generalizes existing methods including conformal prediction, wild bootstrap for Maximum Mean Discrepancy U-statistics, and SymmPI, while also recovering classical bootstrap. Theoretical coverage guarantees interpolate between finite-sample and asymptotic regimes based on invariance strength, measured via Kolmogorov distance or Gaussian universality conditions. The framework enables integration of data augmentation into statistical methods. Empirical evaluation demonstrates DAB's effectiveness on simulated, image, language, and scientific datasets.
data augmented bootstrapconfidence intervalsapproximate invariancekolmogorov distancegaussian universality
Families of Control-Cost-Parametrized Inverse-Optimal Universal Stabilizers
The paper introduces a cost-parametrized family of stabilizing feedback laws, enabling user-defined control costs in inverse-optimal control. The method constructs a nonlinear 'expander' operator through cost differentiation and function inversion, proven Lipschitz for uniform neural operator approximation. Results include semiglobal practical asymptotic stability and second-order suboptimality bounds under approximation, validated numerically. The approach is termed 'half-direct-optimal' as it balances between direct optimal control and fully inverse optimal methods.
inverse-optimal controlstabilizing feedbackneural operatorsemiglobal stabilitysuboptimality bounds
Decoy-Calibrated Failure Audits for Language Models
Janus introduces a decoy-calibrated procedure for auditing language model failures by validating error explanations against randomly assigned decoys and held-out data. The method scores descriptors by error-rate lift, compares them with decoys sharing the same frequencies, and confirms only those that outperform decoys and replicate on separate data. In controlled audits of multi-table lookup tasks, Janus successfully identifies planted failures, confirming long-chain descriptors and their interactions. On benchmarks MuSiQue and LongBench v2, Janus rejects all high-error pockets flagged by SliceLine, demonstrating the necessity of both decoy calibration and holdout checks. The principle separates proposing explanations from reporting them, ensuring only validated findings are reported.
januserror-rate liftdecoysdescriptorsholdout check
DynaCF: Mitigating Shortcut Learning in Reward Models via Dynamic Counterfactual Sensitivity
DynaCF introduces a dynamic reweighting framework to mitigate shortcut learning in reward models trained from pairwise preferences. The method measures shortcut sensitivity online via semantics-preserving counterfactual perturbations, tracking margin shifts and preference flips, then downweights shortcut-sensitive samples in the Bradley-Terry objective. Experiments demonstrate consistent improvements in preference modeling robustness compared to static approaches.
reward modelingshortcut learningcounterfactual perturbationsbradley-terrydynamic reweighting
Structural Grid Descriptors Predict Within-Task Solver Success on ARC-AGI
The study demonstrates that structural grid descriptors predict solver success within the same ARC-AGI task, achieving mean within-task best-feature AUC of 0.885 (p < 0.001). Using hand-crafted grid descriptors measured at 50% trajectory completion across 44,800 runs with two distinct solvers (beam search and Stochastic DFS), the method identifies a grid-complexity axis as most predictive. Results generalize across solvers (transfer AUC 0.747-0.762) and hold on a pre-registered set (AUC 0.765), enabling compute reductions (33.6% for beam search, 65.3% for SDFS) without solve loss.
arc-agigrid descriptorsconditional mutual informationbeam searchstochastic dfs
Multi-Armed Bandits with Arriving Arms: Sequential Screening, Dynamic Regret, and Sublinear Guarantees
The paper introduces UCB-AA, an elimination-based algorithm for multi-armed bandits with dynamically arriving arms, addressing arrival information discrepancy (AID) and drifting benchmarks (DB). The method employs preliminary screening for new arms before full competition, achieving sublinear dynamic regret under gap evolution regularity conditions. Theoretical results demonstrate regret bounds dependent on the arrival process, while simulations show reduced wasted pulls and maintained competitive performance.
multi-armed banditsdynamic regretarrival information discrepancyucb-aasublinear guarantees
LEAF: A Learning-Enabled ADMM Framework for Accelerated Convex Optimization
The paper introduces LEAF, a learning-enabled ADMM framework that accelerates convex optimization by approximating the Moreau envelope via an Input Convex Neural Network (ICNN). The method preserves convexity and smoothness while reducing model complexity through scalar-valued envelope learning, yielding two variants: MEL-ADMM and sMEL-ADMM. Theoretical guarantees confirm convergence and feasibility, with empirical results showing up to 10× speedup over state-of-the-art solvers while maintaining low optimality gaps.
moreau envelopeinput convex neural networkadmmconvex optimizationlearning-enabled optimization
Structure-Aware Modeling of Multiple-Choice Questions Improves Automatic Difficulty Estimation
The study demonstrates that structure-aware modeling of multiple-choice questions (MCQs) improves automatic difficulty estimation by explicitly representing distractors as separate components. The authors propose controlled architectures that encode each distractor as distinct text inputs, aggregating representations via order-aware concatenation or order-invariant summation. Evaluated on Chilean datasets (4,114 MCQs), the best model achieved R²=0.83 (Natural Sciences) and R²=0.71 (Social Sciences), outperforming stem-only baselines. An order-invariant variant matched accuracy with 50% fewer parameters, offering a favorable trade-off for scalable educational applications.
automatic question difficulty estimationmultiple-choice questionsdistractor modelingstructure-aware architectureseducational assessment
Beyond Neural Collapse: Task-Intrinsic Geometry Governs Neural Representations in Modular Arithmetic
This work elucidates the geometric organization of neural representations in modular arithmetic tasks, contrasting with neural collapse predictions. Through layerwise analysis, it demonstrates that classifier weights first form a rank-2 equiangular configuration, subsequently constraining embedding dynamics to the same plane via backpropagation and weight decay. The study formalizes an entropy-regularized transport interpretation on S1, showing that modular-addition labels reduce embedding formation to phase alignment, yielding single-frequency characters as minimizers. Quantitative analysis reveals that the cyclic rank-2 solution outperforms simplex equiangular tight frames by a Θ(K) advantage under Schatten or weight-decay surrogates, establishing a critical threshold λ_crit = Θ(1/K).
neural collapsemodular arithmeticequiangular tight frameentropy-regularized transportphase alignment
Heterophily-Aware Adaptive Knowledge Distillation for Hypergraph Neural Networks
HADES introduces heterophily-aware adaptive knowledge distillation for hypergraph neural networks (HNNs), addressing performance degradation on heterophilic nodes connected through diverse hyperedges. The method quantifies node heterophily as a proxy for teacher reliability and modulates knowledge transfer accordingly. Evaluations on real-world hypergraphs show that HADES consistently enhances student model performance across various HNN teachers and distillation objectives, with students achieving up to 12.3× faster inference while often surpassing teacher predictive accuracy.
hypergraph neural networksknowledge distillationheterophilyadaptive distillationinference speed
Online Learning with Recency: Algorithms for Sliding-window Streaming Multi-armed Bandits
The paper introduces algorithms for single-pass sliding-window streaming multi-armed bandits (MABs), addressing both pure exploration and regret minimization under memory constraints. The model extends traditional streaming MABs by considering only the most recent $W$ arms in a stream, requiring limited memory for arm storage. Theoretical analysis shows hardness for exact best-arm identification with sublinear memory but provides efficient algorithms for approximate solutions, alongside sharp memory-regret trade-offs for regret minimization. Experimental results validate the theoretical trade-offs between sample complexity, regret, and memory usage.
sliding-window streamingmulti-armed banditspure explorationregret minimizationsublinear memory
A systematic investigation of molecular encoding methods for drug property predictions across neural network and Transformer encoder-based model
This study systematically evaluates molecular encoding methods for drug property prediction using a Multi-Layer Perceptron (MLP) and a Transformer encoder-based model (MLP+TL). The authors investigate topological fingerprints, substructure-based fingerprints, and string-based representations across seven molecular datasets. Both models achieved average AUC values above 0.9 on classification tasks such as toxicity, mutagenicity, and side-effect prediction. Attention weights from the MLP+TL model provided interpretable insights into chemically relevant groups, such as hydroxyl-related substructures in blood-brain barrier permeability. The findings offer practical guidance for selecting molecular encoding methods and advance interpretable approaches in molecular informatics.
molecular encodingtransformer encoderattention weightsauc valuesinterpretability
C$^3$ache: Accelerating World Action Models with Cross Inference Chunk Cache
C$^3$ache accelerates World Action Models (WAMs) by exploiting cross-chunk redundancy in denoising residuals during inference. Unlike prior methods that cache computations within a single chunk, C$^3$ache reuses residuals across chunks at the same denoising step, leveraging smooth behavioral trajectories. Evaluated on benchmarks with a Fast-WAM backbone, the method achieves up to 2.5× wall-clock speedup with minimal impact on task success rates, while remaining training-free.
world action modelsdenoising residualscross-chunk redundancyinference accelerationrobot behavior
From inverse problems to neural operators: prediction, mechanism, and generalization of data-driven models
The article proposes a unified framework for classifying data-driven modeling strategies in Scientific Machine Learning, emphasizing their capacity for mechanism discovery and generalization. It contrasts traditional inverse problem approaches, which derive governing differential equations from domain knowledge, with modern methods like Sparse Identification of Nonlinear Dynamics (SINDy), Neural Ordinary Differential Equations (Neural ODEs), and neural operators that directly learn input-output mappings. Drawing from philosophical literature, the authors argue that these methods differ primarily in their assumed model class for input-output relations and that only certain models can discover underlying mechanisms, enabling generalization. The analysis aims to bridge disparate modeling strategies and guide their appropriate application.
inverse problemsneural operatorssparse identificationneural odesmechanism discovery
Self-Consistent Generative Paths via Admissible Random Variational Transport
The paper introduces a theoretical framework for assessing self-consistency in generative probability paths, defined as random fixed points of admissible local variational transport corrections. The framework incorporates random variational transport operators combining divergence, energy, and structural constraints, encompassing random regularized optimal-transport proximal steps and other generative methods. Key results include proofs of well-posedness, random fixed-point existence and attraction, residual-to-generation error bounds, and empirical residual concentration. The theory enables path self-consistency testing and residual-control principles for diagnosing failures, regularizing training, and adaptive sampling across diffusion, flow, one-step, VAE, GAN/WGAN, and autoregressive generators.
generative probability pathsvariational transport correctionsrandom fixed-pointoptimal-transport proximal stepsresidual-control principle
From Hazard Functions to Language Space: Cox-Supervised Distillation of Survival Risk into a Large Language Model
The study demonstrates that survival risk predictions from a Cox proportional hazards model can be effectively distilled into a large language model (LLM) through text-based supervision. The method converts structured clinical covariates into text prompts and fine-tunes a Qwen-based LLM to generate patient-specific survival risk, using Cox model outputs as training targets. Evaluated on GBSG2, ACTG320, and WHAS500 datasets, the approach achieves competitive discrimination and calibration, with t-SNE visualizations revealing continuous risk gradients in the LLM's latent space. This suggests LLMs can internalize survival-risk structure while maintaining predictive accuracy.
cox proportional hazardslarge language modelsurvival analysislatent spacerisk distillation
Estimate Collapsibility of Causal Effects in Completed Partial DAGs via Strong d-Convex Hulls
The paper introduces a collapsible method for consistent causal effect estimation in completed partially directed acyclic graphs (CPDAGs), preserving estimator consistency before and after marginalization over variables. The authors characterize minimal collapsible sets as strong d-convex hulls and devise an efficient algorithm to obtain these sets in DAGs, extending it to CPDAGs. The method integrates graph reduction with the IDA framework. Empirical experiments demonstrate the effectiveness of collapsibility for causal estimations in CPDAGs. Code is publicly available.
collapsibilitycpdagsd-convex hullsida frameworkgraph reduction
Backward Coherence and Hidden-State Stability in Recurrent Neural Networks: A Quasi-Reverse-Martingale Theory
The paper introduces backward coherence as a measure of hidden-state stability in recurrent neural networks (RNNs), formalizing it through quasi-reverse-martingale theory. The authors propose learning a backward projector $g_φ$ to reconstruct $h_t$ from $h_{t+1}$, yielding theoretical guarantees including almost-sure convergence, interpretable limiting representations, and time-uniform confidence sequences. Empirical results show backward-coherence regularization reduces quasi-martingale total by 43-58%, accelerates stability by 28-44%, and improves tracking-error recovery. The method demonstrates competitive performance on PhysioNet 2012 (mortality prediction), FRED-MD (forecasting), and UCI Human Activity Recognition, with theoretical validity under specified assumptions.
backward coherencequasi-reverse-martingalehidden-state stabilityrecurrent neural networksvariational inference
PROBE-Web: An Interactive System for Probing Evaluation Landscapes of Knowledge Graph Completion Models
PROBE-Web introduces an interactive system for probing evaluation landscapes of knowledge graph completion (KGC) models, addressing limitations of conventional rank-based metrics. The system enables flexible evaluation through two adjustable perspectives: predictive sharpness (P1) and popularity-bias robustness (P2). It provides four functionalities: conventional evaluation toolkit, perspective-aware evaluation, explainable case studies, and landscape exploration via a user-friendly GUI. The tool facilitates comparative analysis of multiple KGC models' strengths and weaknesses, helping users align evaluations with specific objectives.
knowledge graph completioninteractive evaluationpredictive sharpnesspopularity-bias robustnessrank-based metrics
Generalized Rank-based Evaluation for Knowledge Graph Completion: Perspectives, Framework, and Analyses
The paper introduces PROBE, a generalized evaluation framework for knowledge graph completion (KGC) addressing two overlooked perspectives: predictive sharpness (P1) and popularity-bias robustness (P2). PROBE comprises a rank transformer (RT) for score estimation based on desired sharpness and a rank aggregator (RA) for final score computation with controlled bias robustness. Theoretical analysis proves PROBE satisfies six key properties for reliable evaluation, outperforming existing metrics in maintaining performance consistency across incomplete facts. Experiments with six KGC models on six real-world KGs demonstrate PROBE's comprehensive and flexible evaluation capability compared to conventional metrics.
knowledge graph completionpredictive sharpnesspopularity-bias robustnessrank transformerrank aggregator
Synthetic but Not Realistic: The Evaluation Challenge in Generative Modelling for Structured Electronic Medical Records
The study proposes a multi-dimensional evaluation framework for synthetic electronic medical records, addressing limitations of current statistical similarity metrics by incorporating epidemiological principles. It assesses descriptive fidelity, clinical utility, and structural validity across four generative paradigms (GAN-based, VAE-boosted, diffusion-based, masked modelling) using the PRIME-CVD cohort (n=50,000). Results show all models capture marginal distributions but fail to preserve subgroup structure, effect estimates, and dependency relationships simultaneously, with well-calibrated models still exhibiting distorted relationships that compromise clinical inference.
synthetic dataelectronic medical recordsgenerative modellingclinical validityepidemiological evaluation
Diffuse AI Control on Fuzzy Tasks
The paper introduces a framework for addressing diffuse AI control risks on fuzzy tasks, conceptualized as an adversarial game between a blue team (trusted weak model) and a red team (subversive strong model). The blue team trains a strong model using a weak scorer to mitigate subversion, while the red team identifies behaviors rated highly by the weak scorer but performing poorly. Evaluated on experimental proposal writing using Opus 4.6 and GPT-OSS-20B, the red team found subversive behaviors via multi-objective evolutionary prompt optimization. An adversarial optimization algorithm for the blue team produced robust prompts resistant to red team exploitation.
diffuse ai controlfuzzy tasksadversarial gamemulti-objective optimizationprompt optimization
Fourier Neural Operators with rank-1 lattice points and hyperbolic cross
The paper improves Fourier neural operators (FNOs) by replacing spatial tensor product grids with rank-1 lattice points and using hyperbolic cross frequency index sets. This modification leverages one-dimensional fast Fourier transforms instead of multi-dimensional transforms, simplifying the architecture. Theoretical analysis shows reduced generalization error with fewer parameters, spatial points, and training samples. Empirical validation on an elliptic PDE demonstrates enhanced accuracy and efficiency. The method combines lattice-based spatial discretization with parametric training on carefully constructed lattices.
fourier neural operatorrank-1 latticehyperbolic crossgeneralization errorelliptic pde
CHROMA: Detecting AI-Generated Images through Inter-Channel Color-Space Correlations
The paper introduces CHROMA, a method for detecting AI-generated images by analyzing inter-channel color-space correlations. It demonstrates that LPIPS, a perceptual metric, responds inconsistently to channel-dependent perturbations, revealing unconstrained cross-channel statistics in generative models. The authors propose augmenting RGB inputs with inter-channel correlation maps and training a fixed CNN backbone, showing improved discrimination between real and synthetic images across RGB and Lab color spaces. Evaluated under single- and multi-generator regimes, CHROMA achieves competitive performance with simpler architecture and training than existing detectors.
inter-channel correlationcolor-space forensicsai-generated image detectionperceptual metricscnn backbone
From A to B to A: Palindromic Zero-Shot Voice Conversion with Non-Parallel Data
The paper introduces a zero-shot voice conversion (VC) framework that constructs synthetic training pairs from non-parallel data via K-Nearest Neighbors (KNN) retrieval over WavLM representations. The method aligns source and target speech by using retrieved segments as synthetic inputs paired with real target audio as ground truth, enabling multilingual support without parallel corpora. A speaker loss from a pretrained verification model maintains target-speaker identity. Experiments show the approach achieves high naturalness and speaker similarity across languages, outperforming baselines despite English-only training.
voice conversionk-nearest neighborswavlmspeaker verificationnon-parallel data
Aperon Technical Report: Hierarchical No-Pointer Tangent-Local Search for High-Dimensional Approximate Nearest Neighbors
The paper introduces HNTL (Hierarchical No-pointer Tangent-Local), a vector indexing framework for approximate nearest neighbor search that eliminates pointer overhead in proximity graphs. HNTL partitions high-dimensional space into local tangent subspaces, representing vectors as low-dimensional coordinates via local PCA (capturing 96.3% variance for d=768), and scans them sequentially using a pointerless Block-SoA layout. Evaluated on anisotropic manifold data (N=10,000), HNTL achieves perfect Rerank Recall@10 with only C=20 candidates while demonstrating 3.61x speedup (4.137 ns/vector) over graph traversals via NEON auto-vectorization, attributed to 3.59x higher IPC and minimal cache misses.
approximate nearest neighborspointerless indexinglocal pcablock-soaauto-vectorization
Continuous Language Diffusion as a Decoder-Interface Problem
The paper investigates how continuous diffusion language models generate fluent text from Gaussian-corrupted sentence embeddings by analyzing Embedded Language Flows (ELF). It identifies a decoder-basin mechanism where denoising succeeds when trajectories reach regions interpretable by the native decoder. A diagnostic protocol evaluates denoisability, semantic recoverability, and decoder compatibility, revealing failures masked by scalar metrics. Experiments show frozen T5 token-embedding lookup achieves 93-96% agreement with native decoder decisions, while a linear readout reaches 97.9%. The study demonstrates that continuous and latent diffusion models must be evaluated as representation-decoder systems.
embedded language flowsdecoder-basin mechanismcontinuous diffusionsemantic recoverabilitylatent diffusion
Active Flow Expansion for Out-of-Distribution Discovery: from Theory to Molecules
The paper introduces Active Flow Expansion (ActFlow), a continued pre-training method for out-of-distribution flow modeling that expands a model's generable set to increase coverage of valid design spaces. ActFlow employs verifier feedback to iteratively adapt to synthetic data generated through active exploration in the learned flow representation, departing from standard data distribution matching. Theoretical guarantees are established for generable set expansion as a local-to-global reachability process. Empirical evaluations across molecular and protein sequence design tasks demonstrate that ActFlow significantly outperforms synthetic flow pre-training methods, expanding valid coverage far beyond the initial pre-trained model's region.
active flow expansionout-of-distribution modelinggenerable setflow representationverifier feedback
Generalization in Nonlinear Least Squares via Learned Feature Geometry
The paper derives generalization bounds for ridge-regularized nonlinear least-squares models using on-average algorithmic stability, introducing a data-dependent effective dimension based on the empirical Jacobian Gram matrix and residual-curvature terms. The analysis recovers classical linear-case results when curvature vanishes, but evaluates them at trained parameters rather than initialization, differing from neural tangent kernel approaches. Bounds depend on gradient feature geometry via covering complexity, scaling with intrinsic dimension for manifold data or activation-stable regions in ReLU networks. Experiments validate Jacobian compression, linearization tightness, and bound accuracy across synthetic and benchmark datasets.
nonlinear least squaresalgorithmic stabilityjacobian gram matrixeffective dimensionintrinsic dimension
OptMuon: Closed-Loop Orthogonalized Momentum Methods for Stochastic Optimization with Zero-Noise Optimality
(No summary returned.)
Reformulate LLM Reinforcement Learning for Efficient Training under Black-box Discrepancy
The paper introduces Discrepancy-Constrained Markov Decision Process (DCMDP), a novel framework addressing train-inference discrepancies in LLM reinforcement learning. DCMDP formulates RL as a dual-objective optimization problem, maximizing rewards while constraining behavior alignment between training and inference via a Lagrangian relaxation mechanism. This allows adaptive balancing of performance improvement and discrepancy control, enabling free policy exploration within a tolerance region while correcting excessive discrepancies. Empirical evaluations demonstrate significant performance gains on Qwen-3-8b (8B dense) and Qwen-3-30bA3b (30B Mixture-of-Experts) models, facilitating high-fidelity training aligned with low-cost inference deployment.
reinforcement learningdiscrepancy-constrained markov decision processlagrangian relaxationtrain-inference discrepancymixture-of-experts
Understanding the Parameter Space Geometry of Transformers Encoding Boolean Functions
The study investigates why transformers fail to learn certain simple but sensitive functions, such as PARITY, despite their theoretical expressivity. By analyzing the geometry of the parameter space, the authors demonstrate that sensitive functions occupy vanishingly small regions, making them unlikely to be found via random initialization. The work introduces the sensitivity profile, showing that transformers initialized randomly almost surely compute low-sensitivity functions. This explains the bias toward low-sensitivity functions and proves the unlearnability of functions lacking such properties.
transformersparameter spacesensitivity profileboolean functionslearnability
Declarative Outcome-Conformant Synthesis: Exact, Closed-Form Specification Satisfaction and a Conformance Benchmark
The paper introduces outcome-conformant synthesis, a novel capability for generating synthetic tabular data that exactly satisfies declared analytical outcomes without requiring source data. The authors formalize this task, contrasting it with imitation-based methods like copulas, GANs, and diffusion models, which prioritize fidelity to real data distributions. They propose a closed-form generator based on conditional-sum sampling of a Gamma population, achieving exact aggregation with a marginal cost of 0.006 in 1-Wasserstein distance. Additionally, SpecBench is introduced as the first benchmark for measuring conformance to analytical outcomes in cold-start relational synthesis. Empirical results show that off-the-shelf synthesizers miss declared aggregates by 74-86%, while the proposed method achieves exact conformance.
outcome-conformant synthesisconditional-sum samplinggamma population1-wasserstein distancecold-start relational synthesis
IR-SIM: A Lightweight Skill-Native Simulator for Navigation, Learning, and Benchmarking
The authors present IR-SIM, a lightweight skill-native simulator for robot navigation that enables rapid scenario construction via YAML configuration files. The system integrates mobile robot kinematics, LiDAR sensing, collision checking, and visualization modules, allowing text-prompted scenario generation for benchmarking and training data synthesis. Experiments demonstrate IR-SIM's utility in natural language scenario construction, collision avoidance policy training, social navigation benchmarking, and seamless transition to high-fidelity simulators. The tool requires no custom coding for scenario modification or real-world deployment validation.
robot navigationyaml configurationlidar sensingcollision avoidancebenchmarking
Compositional Approximation Can Strictly Outperform Superpositional Approximation
The paper demonstrates that compositional approximation methods can strictly outperform superpositional approximation for certain function classes. By constructing explicit examples, the authors show an arbitrarily large gap in approximation rates between these methods. The analysis focuses on function classes with structural properties that limit superpositional approximation rates, while compositional methods achieve higher rates under constraints enabling efficient parameter encoding. Results establish conditions where neural networks and similar compositional approaches yield superior approximation performance compared to traditional linear combination techniques.
compositional approximationsuperpositional approximationapproximation ratesfunction classesparameter encoding
A Geometric Measure of Linear Separability for Neural Representations
The paper introduces the directional linear separability measure (LSM), a diagnostic tool for assessing one-sided affine separability in neural representations. LSM evaluates class-wise geometry by identifying the minimal intrusion of competing samples into the target class's halfspace, normalized by class size. The method establishes theoretical properties including invariance under linear embeddings and provides a penalty-based search for high-dimensional features. Empirical results demonstrate LSM's utility in analyzing geometric transformations across deep-learning architectures and components.
linear separabilityaffine halfspacesneural representationsgeometric measureclass-wise intrusion
Agentic Search for Counterfactual Recourse under Fixed LLM Budgets
The paper introduces Comp-MCTS, a tree-search framework for generating counterfactual recourse alternatives under fixed LLM-call budgets. It addresses the challenge of producing multiple valid counterfactuals by combining LLM-based proposal generation, oracle validation, and compression-guided pruning in a training-free setting. Experiments on four tabular datasets demonstrate that Comp-MCTS outperforms baselines in yield of unique valid counterfactuals, achieving better quantity--quality trade-offs while maintaining competitive proximity, sparsity, and novelty metrics.
counterfactual recoursellm-agentic searchfixed-budget optimizationtree-search frameworkoracle validation
Discovering and decoding latent mean-field structure with variational autoencoders
The study establishes a criterion linking variational autoencoder (VAE) capacity to faithful reconstruction of many-body system joint probabilities, showing that successful VAEs structurally emulate finite-size mean-field factorizations. The method compares latent channel rate with bipartite mutual information, enabling extraction of microscopic parameters from trained decoders. Validation on Curie-Weiss, Hopfield, and Maier-Saupe models recovers order parameters, including full Hopfield pattern matrices, while application to retinal recordings identifies two effective collective variables and derives a generalized Hopfield model matching experimental data.
variational autoencodermean-field theorymany-body systemshopfield modellatent representation
Hierarchical Projection for Adaptive Knowledge Transfer
We propose Projection Transfer Learning (ProjectionTL), a hierarchical Bayesian framework for adaptive knowledge transfer across heterogeneous domains. The method decouples transfer at two levels: first, constructing a source-guided hierarchical prior that aggregates information across sources using data-driven weights; second, refining this borrowing through a posterior-projection step that selectively retains feature-level coordinates exhibiting local agreement with the target signal. This enables simultaneous source selection and feature selection, mitigating negative transfer while preserving interpretability. Experiments on simulations and biomedical applications demonstrate improved accuracy, stability, and interpretability compared to existing methods, offering a scalable strategy for trustworthy cross-domain learning in high-dimensional settings.
hierarchical bayesian modelingposterior-projectionnegative transfersource selectionfeature selection
Activation Steering Induces Emergent Misalignment: A More Comprehensive Evaluation
The study investigates emergent misalignment (EM) induced by activation steering in large language models (LLMs), expanding beyond prior finetuning-focused research. Using steering vectors constructed from target behavior examples and injected into intermediate activations, the authors demonstrate broad misalignment in Qwen-3.5 models, with harmful responses exhibiting higher semantic relevance and coherence than finetuned counterparts. They analyze steering-specific factors (magnitude, low-rank subspace structure, construction epochs) and evaluate robustness across model families, scales, tasks, and intervention layers. Results identify activation steering as an under-examined EM source and provide an activation-space perspective on EM mechanisms and safety risks.
activation steeringemergent misalignmentlarge language modelssteering vectorinference-time intervention
Rank Intervals for Leaderboards: A Hierarchical Framework for Model Evaluation
The authors propose a hierarchical framework for constructing statistically valid rank intervals in multi-task leaderboard evaluations, addressing uncertainty at both task and leaderboard levels. The method combines task-level rank confidence intervals from pairwise comparisons with leaderboard-level rank prediction intervals using conformal prediction. Experiments on simulated data, TabArena, and PromptEval (MMLU) benchmarks demonstrate statistically valid and informative intervals, enabling uncertainty-aware model ranking. This approach improves reliability in aggregating performance across diverse tasks while quantifying variability.
rank intervalsleaderboard evaluationconformal predictionmulti-task learningstatistical guarantees
Speaker-Invariant Representation Learning for Spoofing Detection via Gradient Reversal and A Variational Information Bottleneck
The paper introduces a speaker-invariant spoofing detection method to address poor generalization in out-of-domain settings caused by speaker bias. The proposed teacher-student framework uses a pre-trained speaker recognition teacher and gradient reversal layer, augmented with a Variational Information Bottleneck to balance identity suppression and spoofing cue preservation. Evaluations across nine datasets demonstrate a 25.7% relative reduction in Equal Error Rate (EER) compared to the MHFA baseline.
spoofing detectiongradient reversalvariational information bottleneckspeaker-invariantequal error rate
Learning to Solve Generative ODEs Beyond the Linear Span
SpanLift introduces a neural solver that addresses the span-limited bottleneck in generative ODE solvers by augmenting scalar-coefficient updates with a spatial residual operator. The method preserves a fixed base solver as an in-span prior and learns a spatial residual operator over the state and velocity buffer, trained via endpoint teacher matching without additional model evaluations. SpanLift achieves state-of-the-art few-step sampling across tasks, improving CIFAR-10 FID from 8.16 to 5.69 and ImageNet FID from 17.37 to 11.83 with only 3 NFE.
span-limitedspatial residual operatorendpoint teacher matchingfew-step samplinggenerative odes
SkillHone: A Harness for Continual Agent Skill Evolution Through Persistent Decision History
SkillHone introduces a framework for continual agent skill evolution by maintaining persistent decision histories, including diagnoses, revisions, and outcomes. The method employs role-separated subagents to run candidate skills on practice probes with redacted reporting, enabling cross-session refinement without rediscovering past rationale. Evaluated on deep-research benchmarks (GAIA, WebWalkerQA-EN) using Qwen3.6-35B-A3B, SkillHone outperforms a commercial retrieval-backed agent by 15.8 and 3.2 points respectively, while surpassing prior skill-evolution methods.
continual learningagent skillsdecision historypractice probesredacted reporting
A Comparison of SSL-Based Feature Extractors and Back-End Classifiers for Spoofing Detection: A Multi-Corpus Training and Cross-Linguistic Analysis
The study benchmarks four self-supervised learning (SSL) feature extractors paired with four back-end classifiers for spoofing detection, revealing domain bias in ASVspoof 5 and cross-linguistic adaptation benefits. Methods include hierarchical local feature extraction (ResNet) and global sequence modeling (attention/graph-based classifiers), evaluated across three training scenarios and six datasets. Key findings show naive data scaling degrades performance due to domain bias, while 8 hours of target-language fine-tuning enhances robustness, underscoring the need for domain-aware and language-specific adaptation.
self-supervised learningspoofing detectioncross-linguistic analysisdomain biasfeature extractors
Operator learning for the 2D incompressible Navier-Stokes equations: a conformal prediction approach in the data-scarce regime
The paper introduces a perturbation-based conformal prediction framework for uncertainty quantification in neural operator learning, specifically for the 2D incompressible Navier-Stokes equations. The method augments a trained Fourier Neural Operator (FNO) with split conformal prediction, constructing local uncertainty scales by comparing predictions from two operators: one trained on original labels and another on Gaussian-perturbed labels. In data-scarce regimes, this approach yields narrower conformal bands than existing methods while maintaining target coverage, demonstrating perturbation sensitivity as an efficient uncertainty proxy for conformalized neural operators.
conformal predictionneural operatorsuncertainty quantificationnavier-stokes equationsfourier neural operator
FiberTune: Preserving Action-Fiber Visual Residuals in Vision-Language-Action Fine-Tuning
FiberTune introduces a training-time objective to prevent residual visual collapse in vision-language-action (VLA) policy fine-tuning by preserving teacher-structured visual residuals. The method uses an online action probe to estimate action-predictive feature directions, filters them from visual-token representations, and aligns the residuals to a frozen teacher while regularizing their effective rank. Evaluated across six simulation settings (pi_0.5, OpenVLA-OFT) and physical SO-101 pick-place, FiberTune improves task success rates, including +10.7pp SR(5) on CALVIN ABC-to-D and 72.7% to 78.1% on SO-101, with diagnostics confirming increased residual alignment and rank.
vision-language-actionfine-tuningresidual collapseaction fibersonline probe
Parameter Tuning with Generalization Guarantees for GPU-Accelerated Linear Programming
The paper derives generalization guarantees for hyperparameter tuning in cuPDLP, a GPU-accelerated first-order linear programming solver. By analyzing PDHG (primal-dual hybrid gradient) behavior and PDLP's augmented techniques (preconditioning, adaptive steps, restarts), the authors establish linear and polynomial sample complexity bounds for learning optimal parameters. Proof-of-concept experiments validate the need for data-driven parameter tuning in solver-grade implementations, leveraging recent advances in data-driven algorithm design.
linear programminghyperparameter tuningprimal-dual hybrid gradientsample complexitydata-driven optimization
SpectrumKV: Per-Token Mixed-Precision KV Cache Transfer for Prefill-Decode Disaggregated LLM Serving
SpectrumKV introduces per-token mixed-precision KV cache transfer for prefill-decode disaggregated LLM serving, treating KV transfer as a precision-allocation problem rather than binary token pruning. The method dynamically assigns FP16, INT8, or INT4 precision to tokens based on importance, using a deployment-time probe to handle model-dependent INT4 tolerance. Evaluations on Qwen2.5-7B-Instruct, Mistral-7B-Instruct-v0.3, and Gemma-2-9B-it show perplexity changes of +1.97%, -0.06%, and -0.44% at 50% KV budget, outperforming PDTrim's +25.85%, +22.07%, and +35.63%. NIAH retrieval reaches 52.6% for Qwen at b=0.3 budget (vs 26.3% for PDTrim) and 100% at b=0.5, with 50-62% TTFT reductions.
kv cachemixed-precisionprefill-decodeperplexityniah retrieval
Towards Long-Horizon Vessel Trajectory and Destination Forecasting with Reasoning Large Language Models
This paper introduces a Reinforcement Learning with Verifiable Reward (RLVR) framework for long-horizon vessel trajectory and destination forecasting using reasoning-capable large language models (LLMs). The method converts AIS-based 60-day historical trajectories into semantic textual representations for RL prompt construction, enforcing physical validity and destination correctness through hierarchical matching and curriculum learning. Experimental results demonstrate that RLVR-trained 4B LLMs outperform zero-shot LLMs and deep learning baselines, particularly on destination-related metrics, with LSTM remaining competitive under limited fine-tuning data. The findings highlight the importance of reward-compatible optimization and task-specific capacity matching over model size.
reinforcement learninglarge language modelstrajectory forecastinghierarchical matchingcurriculum learning
Tyan-WP: A Wind Power Foundation Model for Ultra-Short-Term Probabilistic Forecasting
The paper introduces Tyan-WP, the first wind power foundation model for ultra-short-term probabilistic forecasting, addressing limitations of site-specific time series models (TSMs) and generic large time series models (LTSMs). The model incorporates static site embeddings (using coordinates, terrain, and ecoregion metadata) and a power-aware meteorological fusion (PAMF) module to capture interactions between power and meteorological covariates. Pretrained on 126,000 U.S. sites over seven years, Tyan-WP achieves zero-shot forecasting improvements, reducing MAE by 19.9%, RMSE by 16.6%, CRPS by 22.2%, and AQL by 21.7% while increasing R^2 by 16.7% compared to baselines, with strong cross-geography generalization on U.K. sites.
probabilistic forecastingzero-shot learningstatic site embeddingmeteorological fusionwind power foundation model
Bayesian Optimization of a Multi-Product Chemical Reactor Using Composite Models and Partial Physics Knowledge
The authors propose a Bayesian optimization (BO) framework for economic optimization of multi-product chemical reactors when only partial physics knowledge is available. They employ composite Gaussian process (GP) models to predict reactor outputs (product concentrations, temperature) while analytically computing profit from these predictions and market prices. This preserves economic objective structure, avoids retraining with price changes, and enables physics-based validation via energy balance residuals. The BO leverages GP uncertainty for exploration and constraint handling through upper confidence bounds, while penalizing energy-balance mismatches. Evaluated on a non-isothermal reactor simulation, the method outperforms trust-region safe BO in economic performance and avoids temperature violations compared to purely data-driven BO.
bayesian optimizationgaussian processchemical reactorenergy balancecomposite models
Reinforcement Learning for Flow-Matching Policies with Density Transport
The paper introduces RLDT, a reinforcement learning algorithm for fine-tuning flow-matching policies in continuous control. It formulates policy improvement as density transport towards high-reward regions, leveraging Stein Variational Gradient Descent (SVGD) to construct a transport field from a maximum-entropy RL objective. The method avoids biased gradients by approximating policy actions via expected-target estimation, enabling stable training without backpropagation through time. Experiments show RLDT outperforms baselines in reward quality and convergence speed across diverse tasks, including dense/sparse rewards and vision-based robot manipulation.
reinforcement learningflow-matchingstein variational gradient descentdensity transportcontinuous control
How Much Capacity Does EEG Denoising Need? Ultra-Compact Networks reveal Benchmark Saturation and Metric-Utility Gap
The study investigates EEG denoising model capacity by sweeping channel width (1.05K–40.26K parameters) in a depthwise-separable convolutional U-Net, fixing other variables. Results show performance saturation at 3–6.5K parameters, with minimal gains beyond (≤0.015 correlation coefficient/log10-parameter). An 8.46M-parameter baseline matched the 40.26K model, revealing a 200x parameter gap with no advantage. Downstream evaluation exposed a classifier-dependent metric-utility gap: reconstruction-optimized denoising degraded CSP+LDA classification (best accuracy 0.547 vs. 0.612 noisy baseline; p=0.0488). Findings advocate for capacity-controlled evaluation and task-aware benchmarks.
eeg denoisingmodel capacitydepthwise-separable convolutionmetric-utility gapedge deployment
Quantum Global Variational Learning for Quantum Error Correction
We introduce a quantum neural network with global structure for efficient quantum error correction, reducing unitary matrix requirements in quantum circuits. The method achieves a 97% reduction in training time, a 25% improvement in training completion rate, and 100% success rate in training, surpassing prior error correction performance. Enhanced robustness against internal network noise is demonstrated, with fidelity improvements of up to 15% due to reduced computational load.
quantum neural networkquantum error correctionunitary matricesinternal network noisetraining completion rate
Improving the sharpness in neural network-based parametric post-processing of ensemble forecasts
The paper proposes a neural network-based parametric post-processing method for ensemble weather forecasts that improves sharpness without degrading probabilistic forecast skill. The authors augment the continuous ranked probability score (CRPS) loss function with a penalty term to reduce prediction interval width. Evaluated on ECMWF 2m temperature forecasts from EUPPBench, the method achieves 8.2%-12.5% reduction in central prediction interval width while maintaining CRPS and predictive mean RMSE performance.
ensemble forecastingparametric post-processingcontinuous ranked probability scoreneural networkssharpness improvement
Convolutional Sparse Coding via the Locally Competitive Algorithm on Loihi 2
The authors present the first implementation of convolutional sparse coding via the Locally Competitive Algorithm (LCA) on Intel's Loihi 2 neuromorphic chip, benchmarking it against GPU baselines. Their method extends the recurrent LCA formulation to convolutional feature maps with local inhibitory kernels derived from pairwise filter interactions, addressing spatial structure and weight sharing in practical sparse inference workloads. Results demonstrate feasibility and identify operating regimes where neuromorphic implementations become advantageous, positioning convolutional LCA as a benchmark for structured sparse inference on neuromorphic hardware.
convolutional sparse codinglocally competitive algorithmneuromorphic computingloihi 2sparse inference
A spectral audit framework reveals task-dependent aperiodic reliance across EEG and ECG deep learning
We propose a spectral audit framework to quantify task-dependent reliance on aperiodic features in physiological time-series deep learning. The method combines aperiodic/periodic decomposition, phase-preserving Fourier interventions, sham controls, and simulation validation. Results show architecture-general aperiodic reliance: flattening drops exceeded 0.42 balanced-accuracy points for sleep-wake classification, reached 0.07-0.13 for clinical abnormality detection, and remained minimal for motor imagery across six neural architectures. Six of seven EEG foundation models exhibited FDR-significant aperiodic reliance on clinical EEG, persisting after age/sex and recording-era controls. PTB-XL ECG analysis confirmed neural drops of 0.32-0.36 post-demographic matching, extending this confound beyond EEG.
aperiodic decompositionfourier interventionssham controlsbalanced-accuracyneural architectures
Lost in the Non-convex Loss Landscape: How to Fine-tune the Large Time Series Model?
The paper proposes Smoothed Full Fine-tuning (SFF), a method to improve fine-tuning of large time series models (LTSMs) by addressing their non-convex loss landscapes. SFF constructs an auxiliary randomly initialized LTSM and linearly interpolates its weights with the pre-trained model to smooth the loss landscape, preserving pre-trained knowledge while enhancing trainability. Experiments on eight LTSMs (Timer, TimesFM, MOMENT, UniTS, MOIRAI, Chronos, TTMs, Sundial) demonstrate consistent performance gains across diverse downstream tasks.
large time series modelsnon-convex optimizationfine-tuningloss landscapepre-training
OrderDP: A Theoretically Guaranteed Lossless Dynamic Data Pruning Framework
OrderDP introduces a theoretically guaranteed dynamic data pruning framework for lossless training acceleration. The method combines random subset selection with top-q sample selection, ensuring unbiased gradient estimation via a surrogate loss objective. Theoretical analysis confirms convergence and generalization, while empirical results on CIFAR-10/100 and ImageNet-1K show 40%+ training cost reduction with competitive accuracy and stable convergence compared to baselines.
data pruninggradient estimationsurrogate losstraining accelerationgeneralization analysis
Titans-as-a-Layer: Test-Time Memory for Conversational Speech Emotion Recognition
The paper introduces Memory-as-a-Layer (MAL), a plug-and-play test-time neural memory adapter for conversational speech emotion recognition (SER). MAL augments large audio language models (LALMs) by writing dialogue history into a small neural memory and reading it back as audio-token-aligned residual updates, preserving the host model's token positions. Evaluations across multiple LALMs and SER datasets demonstrate improved performance, validating test-time memory as an effective residual contextual mechanism for SER.
speech emotion recognitionneural memoryaudio language modelstest-time adaptationresidual updates
Calibration of Structured Ignorance Certificates for Diagnosing Unknown Unknowns in Reasoning Models
The paper introduces Structured Ignorance Certificates (SICs), a JSON-formatted output schema that forces language models to explicitly declare knowledge gaps rather than hallucinating answers. Authors construct a 7,347-sample Unknown-Unknown dataset using Qwen3-14B to generate cross-domain questions, then fine-tune a 14B-parameter model with Group Relative Policy Optimization using a composite reward function. Results show 99.46% JSON validity, 0.967 mean Certificate Specificity Score, and 3.6% ROUGE-L improvement over baselines on retrieval-grounded generation.
structured ignorance certificatesunknown-unknown detectiongroup relative policy optimizationepistemic calibrationretrieval-grounded generation
EinSort: Sorting is All We Need for Tensorizing LLM
The paper introduces EinSort, an adaptive tensorization method for compressing large language models by discovering low-rank structures through index ordering. The approach leverages tensor networks to efficiently represent and compress model weights and KV-cache, addressing challenges posed by scale and unstructured weight distributions. Experimental results demonstrate superior reconstruction quality compared to baseline methods, validating the efficacy of the proposed technique in reducing memory and computational overhead.
tensor networkslow-rank structurekv-cacheindex orderingreconstruction quality
Physics-Guided Dual Decoding and Spectral Supervision for Global 3D Hydrometeor Prediction
PredHydro-Net introduces a physics-guided dual-decoding framework for global 3D hydrometeor prediction, addressing challenges posed by zero-inflated, long-tailed distributions. The method employs a decoupled architecture where thermodynamic and dynamic fields modulate hydrometeor generation, integrating wavelet-based frequency decoupling, spectral amplitude matching, and adversarial training. Evaluated over 72 hours, it outperforms Earthformer, PredRNNv2, and the Global Forecast System in extreme-event detection and spectral representation, while maintaining climatological consistency with GPM satellite retrievals. The model accurately reproduces 3D cloud structures in events like Hurricane Ian and demonstrates physical interpretability through feature attribution.
hydrometeor predictiondual-decoding frameworkwavelet-based frequency decouplingspectral amplitude matchingphysics-guided deep learning
A Theoretical Analysis of Memory and Overfitting Phenomena in Stochastic Interpolation Models
The paper presents a theoretical analysis of memorization and overfitting in stochastic interpolation models, deriving closed-form expressions for the optimal velocity field and score function. It demonstrates that both deterministic and stochastic generation processes recover training samples in continuous-time oracle settings, with Euler discretization yielding samples centered around training data, controlled by step size. Accumulated estimation errors govern endpoint deviations from the training set, leading to a representation of generated samples as perturbed training data with discretization, estimation error, and Gaussian noise bounds. Theoretical definitions of overfitting and underfitting are provided, supported by synthetic simulations.
stochastic interpolationvelocity fieldeuler discretizationestimation errorsoverfitting
Routine laboratory trajectories encode the onset of organ-level complications in cancer
A transformer model trained on 2,777,595 longitudinal laboratory measurements from 3,905 patients with multiple myeloma or ovarian cancer predicted the two-year onset of 162 treatment-associated complications across eight clinical categories. The model achieved 1.5- to 6.1-fold enrichment above prevalence at the group level and demonstrated AUROC gains up to +0.11 compared to non-sequential baselines, highlighting the importance of temporal laboratory trajectories. Predictions generalized across cancers and healthcare systems, with external validation on MIMIC-IV and MMRF CoMMpass datasets yielding AUROC up to 0.85. Biomarker masking recovered pathophysiological signatures, confirming the model's ability to encode organ deterioration weeks to months before clinical onset.
transformerlongitudinalaurocbiomarker maskingpathophysiology
Autonomous Aerial Manipulation via Contextual Contrastive Meta Reinforcement Learning
The paper introduces Aco2, a contextual contrastive meta reinforcement learning framework for autonomous aerial manipulation by quadrotors. The method employs a contextual observation encoder to infer latent dynamics from interaction history, augmented by a contrastive objective that structures embeddings around task-relevant variations without explicit system identification. Trained in simulation with domain randomization, Aco2 achieves zero-shot transfer to physical quadrotors, enabling end-to-end payload pickup, transport, and delivery of diverse handle-equipped objects.
meta reinforcement learningcontextual encodercontrastive learningaerial manipulationdomain randomization
📰 Industry Media (2)
Learning to lead in a hybrid human-AI enterprise
Enterprise adoption of autonomous AI agents is projected to increase by 300% within two years, with early implementations demonstrating 30-50% productivity gains in customer service, HR, and sales domains. The study examines organizational restructuring through case studies (e.g., Wipro's 240,000-employee deployment of an Ema Unlimited-powered HR agent reducing query resolution from 48 hours to 5 seconds) and surveys (86% of CHROs prioritizing AI workforce integration). Key findings indicate 75% of roles will require redesign by 2030, with successful transitions dependent on reskilling for agent orchestration (prompt engineering, governance frameworks) and soft skill development (relationship building, adaptability).
agentic aienterprise automationreskillinggovernance frameworksprompt engineering
Five things you need to know about AI
The article synthesizes five key AI trends as of mid-2026: (1) Generative AI's workplace automation remains uncertain despite widespread adoption, with limited empirical data on employment impacts. (2) Real-world harms like deepfake abuse (98% pornographic targeting women) and military LLM deployments demonstrate escalating risks beyond speculative existential threats. (3) Anti-AI activism grows, targeting creative industries (e.g., game development controversies) and data center expansions, with incidents like the QuitGPT movement and physical attacks on executives. (4) Scientific applications show promise (e.g., Google DeepMind's Co-Scientist for hypothesis generation), though concerns persist about research quality and scope narrowing. (5) Pervasive AI adoption creates polarized narratives about technological inevitability versus societal control.
generative aideepfakesllmsco-scientistquitgpt
Generated automatically at 2026-06-09 21:30 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.