Daily Digest — 2026-07-08
338 items · 5 research labs, 326 arxiv papers, 7 industry media
🏛️ Research Labs (5)
Australian Payments Plus moves faster with ChatGPT and Codex
Australian Payments Plus (AP+) demonstrates accelerated workflows in regulated payment systems through strategic deployment of OpenAI's ChatGPT Enterprise and Codex. The method integrates AI tools for technical investigations, document synthesis, and product prototyping, maintaining human oversight for risk validation. Results include 77% of employees saving 2+ hours weekly, 80% reporting improved creativity, and Codex reducing simulation build time from days/weeks to 1 day. Complex reconciliation investigations decreased from 4 hours to 30 minutes. AP+ emphasizes secure, governed AI adoption through team-specific examples and champion-driven change management, scaling innovation while preserving accountability in critical payment infrastructure.
chatgpt enterprisecodexpayment systemsreconciliationthreat modeling
Hugging Face Models on Foundry Managed Compute
Microsoft Foundry integrates Hugging Face models into its managed compute platform, enabling enterprise-grade deployment of open-source AI models. The platform supports diverse modalities (text, vision, audio) and optimizes runtime selection (vLLM, SGLang, TensorRT-LLM) based on model architecture. Models undergo a curation pipeline involving security screening, license review, and performance validation before deployment. Foundry provides unified endpoints, SDKs, and observability across pay-per-token, provisioned throughput, and managed compute options. Results include seamless integration of Hugging Face models with Foundry Agents, global deployment capabilities, and automatic runtime upgrades without redeployment.
managed computeruntime selectioncuration pipelinefoundry agentssecurity screening
Run AI workloads on any cloud, store on Hugging Face: zero-egress storage with SkyPilot
Hugging Face and SkyPilot introduce zero-egress storage integration, enabling AI workloads to run on any cloud while storing models and datasets on Hugging Face Hub. The method leverages SkyPilot's multi-cloud scheduler (20+ providers) with Hugging Face's Xet-backed storage, using FUSE mounts (hf-mount) for lazy loading and content-defined chunking for deduplication. Results show free model loading (500 MB/s) across clouds, checkpoint streaming at 170 MB/s, and 10x storage savings for incremental updates. The solution eliminates egress costs and supports read-write operations via hf:// URLs with existing HF_TOKEN authentication.
zero-egress storagefuse mountscontent-defined chunkingmulti-cloud schedulinghf-mount
Expanding Managed Agents in Gemini API: background tasks, remote MCP and more
The Gemini API introduces enhanced Managed Agents capabilities, including background task execution, remote Model Context Protocol (MCP) server integration, custom function calling, and network credential refresh. These features enable asynchronous, production-ready agents by allowing long-running tasks to execute remotely, direct access to private databases via MCP, local execution of custom functions alongside sandbox tools, and seamless credential rotation. Developers can now build autonomous agents that operate within real development environments without blocking applications, leveraging the Gemini Interactions API for advanced agent definitions and environment configurations.
gemini apimodel context protocolbackground executioncustom function callingnetwork credential refresh
Catch up on the Dialogues stage at Google I/O 2026.
Google I/O 2026's Dialogues stage featured interdisciplinary discussions on AI advancements, with keynotes from Google and DeepMind leadership. Sessions covered proactive AI agents for productivity (Josh Woodward et al.), quantum-AI convergence (Hartmut Neven, James Manyika), scientific problem-solving (Demis Hassabis), embodied physical AI in robotics (Kanishka Rao, Alberto Rodriguez), and AI-augmented cinematic storytelling (Doug Liman et al.). The event highlighted emerging technical synergies across domains without presenting quantitative results.
proactive ai agentsquantum-ai convergenceembodied physical aicinematic storytellingscientific problem-solving
📜 arXiv Papers (326)
From Fixed to Free Cameras: Calibration-Free View-Robust Vision-Language-Action Model
We propose Camera-Centric Vision-Language-Action (CamVLA), a calibration-free view-robust model for robot manipulation that decouples action prediction from camera geometry. CamVLA predicts (1) camera-centric end-effector actions in the local camera frame and (2) a 6-DoF hand-eye matrix relating cameras to the robot base, composing these via geometric transformation into base-frame actions. This disentangles pose-independent action generation from camera-perspective grounding, enabling deployment with single monocular RGB images and task instructions. Evaluations on simulation and real-world robot data demonstrate improved success rates across diverse unseen viewpoints compared to existing view-robust VLA policies.
vision-language-actioncamera-centric6-dofhand-eye matrixmonocular
Weak-to-Strong Generalization via Direct On-Policy Distillation
The paper introduces Direct On-Policy Distillation (Direct-OPD), a method for transferring reinforcement learning (RL) improvements from weak to strong language models without costly RL reruns. Direct-OPD distills the policy shift between pre- and post-RL weak model checkpoints as an implicit reward signal for the stronger student model, applied to its own on-policy states. Experiments show Direct-OPD boosts Qwen3-1.7B's performance on AIME 2024 from 48.3% to 62.4% in 4 hours, outperforming direct RL and enabling sequential policy shifts.
reinforcement learningpolicy distillationlanguage modelsweak-to-strong generalizationimplicit reward
Interpretable Human-Label-Free Deep Learning for Real-Bogus Classification with Uncertainty Quantification
We present an interpretable, human-label-free deep learning framework for Real-Bogus classification in time-domain surveys, addressing costly labeling and noisy community labels. The method combines simulated transient injections with contaminated survey data, employing a dual-network model trained via asymmetric co-teaching for robustness under class contamination. Uncertainty quantification is achieved through a hybrid strategy leveraging MC dropout and deep ensembles. Evaluation on benchmark data shows strong classification performance, high-fidelity recovery of transient light-curve classes, and competitive calibration. Latent-space analysis reveals uncertainty alignment with decision boundaries and bogus subclasses. The approach enables scalable, consistent classification without human-labeled data and supports transfer to future surveys.
real-bogus classificationuncertainty quantificationasymmetric co-teachingtransient injectionslatent-space visualization
LLM-as-a-Verifier: A General-Purpose Verification Framework
LLM-as-a-Verifier introduces a general-purpose verification framework that leverages LLMs to assess solution correctness without additional training, scaling along score granularity, repeated evaluation, and criteria decomposition. By computing expectations over scoring token logits, it generates continuous scores, improving calibration and separation between positive and negative solutions. The framework achieves state-of-the-art performance on Terminal-Bench V2 (86.5%), SWE-Bench Verified (78.2%), RoboRewardBench (87.4%), and MedAgentBench (73.3%). It also provides fine-grained feedback for RL, enhancing sample efficiency in SAC and GRPO on robotics and mathematical reasoning benchmarks.
verification frameworkscoring granularitytoken logitssample efficiencycriteria decomposition
Search Beyond What Can Be Taught: Evolving the Knowledge Boundary in Agentic Visual Generation
The paper introduces SearchGen-20K and SearchGen-Bench, a dataset and benchmark with 20,839 prompts across 12 failure categories and 22 domains, revealing a 40-point performance drop (21-28/100) in frontier open visual generators for world-knowledge-grounded requests. To address this, the authors propose a teach-then-search co-training framework that dynamically discovers the generator's knowledge boundary, improving performance by selectively retrieving external context. The work includes SearchGen-Corpus-1M for reproducible research and demonstrates monotonic improvement through co-training, enabling recursive self-improvement in agentic visual generation.
visual generationknowledge boundaryagentic systemsmultimodal searchco-training framework
What Does a Discrete Diffusion Model Learn?
The paper rigorously analyzes discrete diffusion models by deriving the continuous-time Markov chain ELBO with boundary terms, proving the Oracle Distance theorem: the negative ELBO equals data entropy plus the path KL divergence from the oracle reverse process to the learned one. The method identifies the optimizer as the conditional expectation of the true reverse jump rate and shows the irreducible cost is the information destruction rate of the forward process. Results include exact coordinate conversions for token-factorizing noise, unifying existing methods (MDM, UDM, SEDD, GIDD) and explaining parameterization effects on ELBO behavior, validated numerically on a solvable model.
discrete diffusionmarkov chainelbooracle distancetoken-factorizing noise
Cortex: A Bidirectionally Aligned Embodied Agent Framework for Long-horizon Manipulation
Cortex introduces a bidirectionally aligned embodied agent framework for long-horizon manipulation, addressing the semantic-kinematic gap in hierarchical VLA models. The method standardizes 32 canonical skill primitives, injects tractability principles into data generation, and employs event-balanced sampling for subtask transition ambiguity. Evaluations show Cortex outperforms monolithic baselines by 3.1% on Libero-long and 4.1% on RoboTwin, with zero-shot capability for unseen real-world tasks like multi-stage chemistry experiments.
vision-language-action modelsskill primitivestractability principlesevent-balanced samplinglong-horizon manipulation
GaP: A Graph-as-Policy Multi-Agent Self-Learning Harness For Variational Automation Tasks
Graph-as-Policy (GaP) introduces a multi-agent coding harness for Variational Automation (VA) tasks, addressing reliability gaps in model-free policies for open-world adaptability. GaP generates directed computation graphs with perception, planning, and control nodes from a Modular Open Robot Skill Library (MORSL), then refines graph structures and parameters through parallel internal simulation rehearsals. Evaluated on 8 VA benchmarks (4 in-simulation, 4 real-world), GaP significantly outperforms baselines in success rates and throughput. The approach integrates Task and Motion Planning (TAMP) and Robot Operating System (ROS) principles for interpretable robot programming.
variational automationgraph-as-policytask and motion planningrobot operating systemmodular open robot skill library
SPEARBench: A Benchmark for Naturalness Evaluation in Streaming Speech-to-Speech Language Models
SPEARBench introduces a novel benchmark for evaluating naturalness in streaming speech-to-speech language models, addressing gaps in existing benchmarks by measuring conversational qualities like timing, prosody, and interpersonal dynamics. The method constructs controlled dialogues from the Seamless Interaction corpus, evaluates multiple models across dimensions including latency, speech quality, dialect consistency, and emotional adaptation, and compares them to human reference answers. Results indicate current models achieve high signal quality and ASR robustness but lag in human-like conversational behaviors such as emotional adaptation and stance dynamics.
speech-to-speechnaturalness evaluationconversational aidialogue systemsprosody
REDDIT: Correcting Model-Generated Timestamp Drift in ASR without Forgetting via Replay-Based Distribution Editing
The paper introduces REDDIT, a replay-based distribution editing framework for correcting timestamp drift in autoregressive ASR systems without catastrophic forgetting. The method first edits timestamp targets under replayed decoder context while preserving the base distribution on non-timestamp tokens, then applies edited-prefix refinement. Using VAD-trimmed speech spans and synthetic gaps for supervision, REDDIT achieves 95.0% mIoU on long-gap benchmarks (from 38.7%) with only 1.6% parameter updates, reducing alignment error from 2752ms to 223ms while maintaining 41.3% MER on CV-en.
autoregressive asrtimestamp driftreplay-based learningdistribution editingvad-trimmed
SovereignPA-Bench: Evaluating User-Owned Personal Agents under Evolving Intent, Platform Mediation, and Consent Constraints
We introduce SovereignPA-Bench, an executable benchmark for evaluating user-owned personal agents under evolving intent, platform mediation, privacy boundaries, consent constraints, evidence requirements, and burden tradeoffs. The benchmark separates ObservableState from HiddenLabels, reports component metrics for task success, alignment, privacy, consent, evidence, manipulation, burden, and auditability, and preserves paired scenario ordering for model and policy comparisons. Evaluating 120 sovereignty stress scenarios across 4 model families and 8 policy baselines yields 3,840 frozen-prompt trajectories with raw prompts, outputs, provider-form responses, parsed actions, recomputable metrics, hard-set analyses, qualitative cases, and a blinded 3-annotator audit over 240 items. Full-sovereign scaffolding improves sovereignty score over baselines while reducing privacy leakage, consent violation, over-concession, and manipulation capture, highlighting the need for consent-aware, evidence-grounded action in personal-agent evaluation.
sovereignpa-benchobservablestatehiddenlabelsfrozen-promptfull-sovereign
Graph Sparse Sampling: Breaking the Curse of the Horizon in Continuous MDP Planning
Graph Sparse Sampling (GSS) introduces a branch-free graph-based planning algorithm for continuous Markov Decision Processes (MDPs), mitigating the exponential computational complexity of tree-based methods like Monte Carlo Tree Search (MCTS). GSS shares sampled futures across candidate decisions, enabling GPU-friendly batch processing and heuristic-guided computation. Theoretical analysis provides finite-sample performance guarantees with polynomial dependence on the planning horizon under specific regularity and action-coverage conditions. Empirical evaluations demonstrate GSS outperforms tree-based planners in continuous-control simulations, achieving near-optimal performance on long horizons.
graph sparse samplingcontinuous mdpsmonte carlo tree searchplanning horizongpu-friendly batch
Selective Disclosure Watermarking for Large Language Models
We propose Hierarchical Vocabulary Routing (HeRo), a watermarking framework enabling selective disclosure of metadata in text generated by large language models (LLMs). The method recursively partitions the vocabulary and distributes watermark information across hierarchical layers, allowing verifiers to decode only payload portions corresponding to their access level. HeRo preserves the unbiasedness of the underlying sampling process, maintaining text quality. Experiments demonstrate fine-grained access control with high detection accuracy and low latency. Code is publicly available.
hierarchical vocabulary routingselective disclosurelarge language modelswatermarkingaccess control
Multiplayer Interactive World Models with Representation Autoencoders
The paper introduces the first multiplayer world model for dynamic environments with complex physical interactions, specifically tested in Rocket League. The 5-billion-parameter latent diffusion model conditions on action streams of multiple agents, generating four-player matches at 20 fps on a single Nvidia B200 GPU. Trained on 10,000 hours of bot gameplay, it maintains stable rollouts for up to five minutes, with observed extensions to hours. Key design choices include the video codec, generative objective, and multiplayer conditioning scheme. The authors release a dataset, codebase, and live demo.
multiplayer world modellatent diffusion modeldynamic environmentsphysical interactionsaction streams
OptiAgent: End-to-End Optimization Modeling via Multi-Agent Iterative Refinement
OptiAgent introduces a multi-agent framework for end-to-end optimization modeling that translates natural language problem descriptions into solver-ready mathematical formulations and executable code. The architecture employs dedicated agents to extract structural components like decision variables and constraints, supported by a multi-loop validation system with four specialized feedback mechanisms targeting distinct failure modes. This modular design enhances transparency by exposing agent reasoning and feedback, making the modeling process auditable. OptiAgent achieves state-of-the-art performance on 3 out of 4 benchmarks across LP, MILP, and Nonlinear Programming tasks, remaining highly competitive on the remaining dataset.
multi-agent frameworkoptimization modelingmathematical formulationmulti-loop validationdecision variables
TREK: Distill to Explore, Reinforce to Refine
TREK (Teacher-Routed Exploration via Forward KL) introduces a staged distillation procedure for expanding exploration support in policy optimization, addressing GRPO's limitations on hard prompts. The method identifies low-success prompts, queries verified solutions from any teacher (black-box, white-box, or self-context), ranks proposals by student likelihood, applies forward-KL distillation, and resumes GRPO refinement. Results show improvements: Qwen3-8B's AIME 2025 score increased from 36.9 to 40.3, ALFWorld success rose from 75.8% to 82.8%, and ScienceWorld improved from 12.5% to 26.7%, with faster convergence on hard tasks.
group relative policy optimizationforward-kl distillationexploration supportverified trajectorieshard-prompt sampling
Steering Optimisation Trajectories in Diffusion Representation Learning
The paper introduces SteeringDRL, a method to improve latent representation quality in diffusion autoencoders by steering optimization trajectories. The authors analyze training dynamics, identifying two regimes (reconstruction-focused vs. disentanglement-focused) and propose controlling them via gated residual U-Nets and a noise-level exposure curriculum. Results show improved representation quality on disentanglement benchmarks (reduced seed sensitivity) and enhanced spatial disentanglement for object-centric learning, with segmentation improvements on both synthetic and real-world datasets.
diffusion autoencodersoptimization trajectoriesdisentanglement benchmarksgated residual u-netsnoise-level curriculum
Topological Shape Representation for Aneurysm -- Bifurcation Detection
The study introduces a topology-aware false-positive reduction framework for intracranial aneurysm detection, addressing CNN limitations in distinguishing saccular aneurysms from vascular bifurcations. The method evaluates Smooth Euler Characteristic Transform (SECT) against persistence-based summaries, leveraging global 3D vascular geometry independent of intensity. On the RSNA 2025 dataset, SECT achieves 0.943 AUC, outperforming direction-agnostic methods (AUC ~0.68), with 78.5% sensitivity at 95% specificity for sub-3 mm lesions. SECT also demonstrates scanner-agnostic performance (0.927 mean AUC under LOGO validation), making it suitable for hybrid deep-learning pipelines.
smooth euler characteristic transformintracranial aneurysmfalse-positive reductionpersistence imagestopology-aware
Evaluating and Understanding Model Editing for Medical Vision Language Models
The paper introduces M3Bench, a clinically grounded benchmark for evaluating multimodal model editing in medical vision-language models (VLMs), addressing limitations of existing general-purpose benchmarks. M3Bench contains 16,276 questions across diverse clinical scenarios, assessing reliability, precision, and generalizability under domain-specific challenges like modality shifts and temporal progression. Evaluating 4 editing methods on 6 VLMs reveals trade-offs: gradient-based editors show strong transfer but suffer catastrophic locality violations, while memory-based methods preserve locality but lack compositional generality and exhibit backbone-dependent sensitivity. Failures are attributed to latent space geometry shifts, providing actionable insights for safer post-deployment adaptation.
model editingvision-language modelsclinical benchmarklatent space geometrymultimodal evaluation
MetaSkill-Evolve: Recursive Self-Improvement of LLM Agents via Two-Timescale Meta-Skill Evolution
MetaSkill-Evolve introduces a recursive self-improvement framework for LLM agents, enabling simultaneous evolution of task skills and meta-skills (parameterizing the improvement pipeline) at two timescales. The method employs a unified pipeline with five components (Analyzer, Retriever, Allocator, Proposer, Evolver) shared across a frozen backbone model, requiring no additional parameters or objectives. Evaluations on OfficeQA, SealQA, and ALFWorld show accuracy gains of +23.54, +16.09, and +1.92 points over the raw backbone, outperforming static and single-level evolution baselines.
meta-skill evolutionrecursive self-improvementllm agentstwo-timescale learningtask skill adaptation
Air Quality Downscaling with Station-Guided Pseudo-Supervision
The authors propose a station-guided framework for high-resolution PM$_{2.5}$ downscaling over Europe, addressing the spatial mismatch between coarse atmospheric fields and discrete ground-truth observations. Their method jointly super-resolves ($\times 40$, $\approx$ 1 km) and bias-corrects CAMS atmospheric composition fields using a multi-scale transformer network, incorporating heterogeneous side information (human activity, land cover, elevation, satellite aerosol observations, wind fields). A time-agnostic propagation strategy with spatial Gaussian blending of interpolated OpenAQ observations enables dense supervision. Evaluations demonstrate recovery of fine-grained spatial structures and effective mitigation of localized CAMS biases.
pm$_{2.5}$ downscalingmulti-scale transformerspatial gaussian blendingcams bias-correctionopenaq interpolation
Wavelet Scattering Transform for Interpretable Schizophrenia Biomarker Discovery and Classification from Resting-State EEG
The study introduces a Wavelet Scattering Transform (WST)-based framework for interpretable schizophrenia biomarker discovery from resting-state EEG, addressing limitations of static spectral features and temporal data leakage. Using multi-order WST coefficients to capture amplitude modulation dynamics, the method employs strict Leave-One-Subject-Out cross-validation and SHAP explainability. Second-order scattering coefficients, particularly gamma-band features, emerged as dominant biomarkers, with electrode P3 being most discriminative. A Random Forest classifier achieved 90.48% accuracy (AUC=0.9339) under subject-independent evaluation, demonstrating the framework's potential for psychiatric biomarker discovery.
wavelet scattering transformcross-frequency couplingleave one subject outeeg biomarkersamplitude modulation
ProPS: Prompted Profile Synthesis for Natural Language-Conditioned Speaker Embedding Distributions
ProPS introduces a generative framework for synthesizing speaker embedding distributions conditioned on natural language prompts, addressing the limitation of descriptive x-vector extractors. The method converts profile descriptions into sentence embeddings and employs a mixture density network to predict Gaussian mixture models in x-vector space, trained via likelihood maximization on a large-scale dataset. Evaluations demonstrate ProPS preserves requested attributes (age, gender, accent, prosody) with improved negative log-likelihood and classification accuracy, enabling controllable synthesis for TTS and VC systems.
speaker embeddingsx-vectorsmixture density networknatural language conditioninggaussian mixture model
Adaptive Inference Batching using Policy Gradients
This work demonstrates that reinforcement learning (RL) outperforms static batching policies in multi-GPU heterogeneous routing but offers marginal gains in single-GPU settings. Using REINFORCE and PPO agents trained on a validated discrete-event simulator, the authors formulate adaptive batching and routing as an MDP over queue state, request type, and GPU availability. In multi-GPU scenarios, the RL agent discovers a workload-segregation policy that eliminates Head-of-Line blocking, achieving 3.5x higher throughput and 25% lower latency compared to Round-Robin, while respecting SLA constraints. The policy generalizes to unseen bursty and real-world traffic despite training on synthetic Poisson arrivals, with attention-augmented networks converging 20% faster than MLP baselines.
reinforcement learningmulti-gpu routinghead-of-line blockingdiscrete-event simulatorppo agents
Shifting from Discrete to Continuous Reference Data: QSM-Derived Horizontal Tree Biomass Distribution for Deep Learning Biomass Estimation
The study introduces a continuous Horizontal Biomass Distribution (HBD) reference derived from Quantitative Structure Models (QSMs) to improve LiDAR-based above-ground biomass (AGB) estimation, addressing boundary-effect uncertainties in discrete plot-level aggregates. A sparse 3D U-Net was trained on simulated broadleaved forests using three reference types: forest inventory (FI) aggregates, QSM plot-level aggregates, and continuous HBD mapping. Results show QSM-based models outperformed FI approaches, particularly for small plots (100 $m^2$), reducing relative root mean square error (RRMSE) by 16.84 ± 4.37% and increasing $R^2$ by 0.22 ± 0.05.
lidarabove-ground biomassquantitative structure models3d u-netedge-effect correction
Privacy-Preserving Robustness Verification for Neural Networks
SecureCROWN introduces the first privacy-preserving framework for neural network robustness verification, resolving the tension between verification requirements and data privacy constraints. The method employs secure two-party computation (2PC) to compute certified robustness bounds without exposing model parameters or input data, reformulating conditional operations in Linear Bound Propagation as continuous arithmetic to avoid branching. Experimental results demonstrate exact alignment with plaintext verification, achieving runtimes of 0.1–200s across diverse models and network conditions (LAN/WAN) while maintaining security under the semi-honest model.
secure two-party computationrobustness verificationlinear bound propagationprivacy-preservingneural networks
CanniUplift: A Holistic Framework for Mitigating Seller and Incentive Cannibalization in E-commerce Uplift Modeling
CanniUplift introduces a unified framework addressing seller and incentive cannibalization in e-commerce uplift modeling, where traditional ITE estimation violates SUTVA in multi-seller environments. The method combines Platform-level Global Alignment (PGA) for cross-shop substitution via GMV constraints, Redemption-based Decomposition Denoising (RDD) for incentive noise reduction, and a Treat-Attention mechanism for user-treatment interaction modeling. Experiments on synthetic and industrial datasets show CanniUplift outperforms baselines, with PGA and RDD improving wAUUC and wQINI. Deployment achieved 4.08% Delta GMV increase and higher ROI in A/B tests.
uplift modelingindividual treatment effectsstable unit treatment value assumptionglobal alignmentdenoising
Optimizing ML Workload Partitioning between CPUs and CIM Accelerators for Heterogeneous Computing
The authors propose an Integer Linear Programming (ILP)-based framework for optimizing ML workload partitioning between CPUs and Computing-in-Memory (CIM) accelerators, addressing limitations of prior approaches that neglect RRAM constraints, parallelism, and CPU co-processing. Their method combines empirical profiling with analytical models to minimize inference latency under RRAM memory, write latency, and endurance constraints. Evaluations demonstrate speedups of up to 30.9x over edge CPU-only execution and 7.3x over high-performance CPUs, with Design Space Exploration providing insights for future CIM accelerator designs.
computing-in-memoryworkload partitioningresistive raminteger linear programmingheterogeneous computing
MoP-JEPA: Hard-Assigned Predictor Mixtures for Stochastic JEPA World Models
MoP-JEPA introduces hard-assigned predictor mixtures to address stochastic environment modeling in JEPA world models, where single deterministic predictors fail at branching transitions by regressing to non-existent mean states. The method employs quantized transition distributions with one predictor per successor mode, verifiable through input-agnostic codebook controls and transition-precision guards. On OGBench offline data, MoP-JEPA achieves up to 0.85 success rate (vs. 0.02–0.09 for single-predictor baselines) and outperforms deterministic, gated-MoE, and variational predictors under strict verification protocols. Real-environment execution ranks second on OGBench’s hardest maze.
jepamixture-of-expertsstochastic transitionsworld modelsquantized prediction
EvoAgentBench: Benchmarking Agent Self-Evolution via Ability Transfer
The paper introduces EvoAgentBench, a benchmark for evaluating agent self-evolution through procedural ability transfer across four domains: web research, algorithmic reasoning, software engineering, and knowledge work. The method extracts trace-grounded Abilities from agent executions, canonicalizes them into operational units, and constructs domain-specific Ability Graphs to link tasks with procedural overlap. Results on a 528/267 train/test split show reliable ability transfer across model families, though no automatic method achieves consistent positive gains. The benchmark enables fine-grained diagnosis of experience encoding, routing, and uptake.
agent self-evolutionability transfertrace-grounded abilitiesability graphsprocedural reuse
Reason, Reward, Refine: Step-Level Errors Corrections with Structured Feedback for Physics Reasoning in Small Language Models
The paper introduces a step-level reward framework for correcting physics reasoning errors in small language models, addressing error propagation without ground truth exposure. The method identifies the first reasoning error, generates structured feedback, and trains the model via policy gradient with KL regularization, eliminating preference data requirements. Evaluated on five physics benchmarks, the framework improves accuracy by 17-20% over chain-of-thought prompting and reduces calculation errors from 56.9% to 23.5%, though conceptual errors remain challenging (89.7% to 68.7%).
physics reasoningerror propagationpolicy gradientkl regularizationstructured feedback
Noisy-Channel Minimum Bayes Risk Decoding
The study introduces a noisy channel decomposition of Minimum Bayes Risk (MBR) decoding to address asymmetries in hypothesis-to-reference and reference-to-hypothesis directional effects. The method decomposes MBR decoding into four components: hypothesis-to-reference likelihood, reference-to-hypothesis likelihood, hypothesis prior, and reference prior. This framework unifies existing MBR variants and enhances metric- and task-specific interpretability. Analysis shows that channel-wise contributions vary across metrics but remain consistent across tasks, suggesting that optimized channel weighting can improve over original MBR decoding.
minimum bayes risknoisy channelhypothesis-to-referencereference-to-hypothesistask-specific interpretability
Unified Audio Intelligence Without Regressing on Text Intelligence
The work introduces Nemotron-Labs-Audex-30B-A3B (Audex), a unified audio-text LLM based on Nemotron-Cascade-2-30B-A3B, a text-only Mixture-of-Experts model. Audex employs a single Transformer decoder architecture where audio inputs are projected into text embedding space, enabling multimodal generation while maintaining compatibility with standard LLM infrastructure. Training utilizes 157.4B audio tokens and 320.5B text tokens, followed by multi-stage supervised training, Cascade RL, and multi-domain distillation. The model achieves state-of-the-art performance in audio understanding, speech recognition/translation, and generation tasks without regressing on text capabilities.
audio-text llmtransformer decodermixture-of-expertsmultimodal generationcascade rl
When Claws Remember but Do Not Tell: Stealthy Memory Injection in Persistent Personal Agents
The paper introduces stealth memory injection, a novel attack vector where untrusted external content is covertly written into persistent memory of personal agents for later misuse. It presents WhisperBench, a 108-case benchmark for evaluating such attacks across five risk categories, and MemGhost, a one-shot payload generation framework using environment and objective proxies with reinforcement learning. MemGhost achieves 87.5% success on OpenClaw/GPT-5.4 and 71.4% on Claude Code SDK/Sonnet 4.6, demonstrating transferability across architectures and resilience against defenses.
stealth memory injectionpersistent personal agentswhisperbenchmemghostone-shot payload generation
ClassicLogic: A Knowledge-Driven Benchmark of Classic Puzzle Games for Evaluating Compositional Generalization
The authors introduce ClassicLogic, a knowledge-driven benchmark suite for evaluating compositional generalization in AI systems through four classic logic puzzles (Sudoku, KenKen, Kakuro, Futoshiki). The benchmark features hierarchical knowledge bases where complex strategies are formally defined as compositions of simpler ones, enabling fine-grained assessment of reasoning capabilities from basic rule learning to multi-step compositional problem-solving. The open-source benchmark provides mathematically validated difficulty levels and serves as a testbed for neuro-symbolic and advanced reasoning systems.
compositional generalizationknowledge-driven benchmarkhierarchical knowledge baseneuro-symbolic systemslogic puzzles
Rethinking On-Policy Self-Distillation for Thinking Models
The paper demonstrates that privileged-context self-distillation harms performance in thinking models during long reasoning tasks, contrasting with its benefits in instruction-tuned models. Through experiments on five Qwen3 and OLMo models across AIME24, AIME25, and HMMT25 benchmarks, privileged distillation causes up to 17% relative accuracy drop in avg@16 metrics, particularly at high rollout budgets. Analysis reveals this stems from privileged context altering learning dynamics at high-entropy decision points, reducing fork rates and suppressing verification behaviors in reasoning traces.
self-distillationthinking modelson-policy distillationreasoning tracesfork rates
Relational Multi-Agent Reinforcement Learning for Dynamic Pricing in High-Speed Railway Markets
Proposes a relational multi-agent reinforcement learning framework for dynamic pricing in liberalized railway markets, addressing partial observability and regulatory constraints. The method employs an entity graph modeling approach, extending the multi-agent twin delayed deep deterministic policy gradient algorithm with graph-based representation learning via a multi-layer relational graph convolutional network and attention mechanism. Experiments demonstrate superior revenue and stability in complex market settings compared to relational and non-relational baselines. Code is publicly available.
multi-agent reinforcement learningdynamic pricinggraph convolutional networkattention mechanismpartial observability
CP-WSP: A Declarative CP-SAT Framework for Configurable Multi-Constraint Workforce Scheduling
CP-WSP introduces a declarative CP-SAT framework for workforce scheduling, addressing limitations in existing approaches by supporting 14 hard constraints and optimizing 15 soft objectives via a unified weighted penalty function. The framework enables mandatory break scheduling with midpoint control, acuity-weighted workload equity, sub-shift temporal granularity, inter-week schedule stability, and cross-midnight shift patterns. Key innovations include shift-window variable decomposition, grid-offset preprocessing, and multi-granularity temporal resolution from 30 minutes to 2 hours. Evaluated on INRC-II benchmarks and 36 synthetic configurations, CP-WSP ensures zero regulatory violations by construction while optimizing diverse scheduling objectives.
cp-satworkforce schedulingtemporal granularityshift-window decompositiongrid-offset preprocessing
AgentGym2: Benchmarking Large Language Model Agents in De-Idealized Real-World Environments
The paper introduces AgentGym2, a benchmark for evaluating LLM agents in realistic, de-idealized environments that better reflect real-world deployment challenges. Unlike existing benchmarks, AgentGym2 requires agents to handle noisy inputs, discover tools through exploration, compose tools for novel tasks, and execute end-to-end procedures. Testing 15 proprietary and open-source models, including state-of-the-art systems like Gemini and GPT-5, reveals significant performance gaps, highlighting the limitations of current agents in practical applications.
llm agentsbenchmark evaluationtool discoverynoise robustnessend-to-end execution
The Changing Role of Symbolic Methods in Artificial Intelligence
The article proposes the Compression Principle, positing that explicit symbolic reasoning compensates for information loss in simplified computational models, and derives the Modeling--Reasoning Trade-off: richer representations reduce the need for symbolic reasoning. This framework explains both historical symbolic AI success and modern foundation models' effectiveness. Paradoxically, as AI systems grow more opaque, symbolic methods become crucial human-AI interfaces for specification, verification, and regulation. The argument shifts symbolic methods' primary role from computational engines to human-facing interfaces.
symbolic reasoningfoundation modelscompression principlemodeling-reasoning trade-offhuman-ai interface
Open Problems in AI Incident Governance
The paper identifies key gaps in AI incident governance frameworks, focusing on post-deployment failure management. Through analysis of existing regulatory and independent frameworks, it reveals inconsistencies in incident definitions, classification schemes, monitoring practices, and reporting mechanisms. These inconsistencies impair data collection quality and limit the depth, representativeness, and accuracy of subsequent incident analyses.
ai incident governancefailure managementpost-deployment safetyincident classificationmonitoring frameworks
DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation
DSpark introduces a speculative decoding framework combining parallel draft generation with adaptive verification to accelerate LLM inference. The method employs a semi-autoregressive architecture with a parallel backbone and sequential module for intra-block dependency modeling, alongside confidence-scheduled verification that dynamically adjusts verification length based on prefix survival probabilities. Evaluations show DSpark improves accepted length over state-of-the-art drafters and achieves 60-85% speedup in production deployment (DeepSeek-V4) while preventing throughput degradation under interactivity constraints.
speculative decodingsemi-autoregressiveconfidence-scheduled verificationthroughput degradationintra-block dependency
PDEFlow: Autonomous Agentic PDE Pipelines for Neural Operator Learning and Solver-Free Inference
PDEFlow introduces an autonomous agentic framework for transforming user-level ODE/PDE descriptions into neural-operator pipelines, integrating problem specification, data generation, operator training, and checkpoint-based inference. The framework employs a stateful input graph for multi-turn natural-language input validation, a FEniCSx finite-element backend for solver-backed data generation, and a registry-based interface for neural operator training and deployment. Current implementation utilizes a multi-branch Bayesian DeepONet. Experiments demonstrate PDEFlow's capability to construct valid specifications, generate datasets, train operators across steady and transient problem classes, and provide solver-free predictions, enabling repeatable scientific workflows with minimal manual intervention.
neural operatorfinite-elementdeeponetpdecheckpoint-based inference
TacReasoner: A Dynamic Tactile-Language Framework for Interactive Reasoning in Real-World Scenarios
TacReasoner introduces a dynamic tactile-language framework addressing two key challenges in multimodal reasoning: insufficient modeling of dynamic tactile signals and hallucination in tactile foundation models. The framework incorporates a Dynamic-aware Tactile Encoder and TouchCoT-10k, the first tactile chain-of-thought dataset, enabling structured reasoning over tactile inputs. Evaluated on DynTac-Bench, TacReasoner demonstrates competitive performance, outperforming the 14B VTV-LLM model on most subtasks despite using only 7B parameters, highlighting its effectiveness and efficiency in tactile commonsense reasoning.
dynamic tactile signalstactile chain-of-thoughtmultimodal reasoningtactile foundation modelscommonsense reasoning
Three-Phase Evaluation of AI-Assisted Software Development Life Cycle
The paper evaluates how varying AI autonomy levels impact software development by comparing three workflows: partial AI-assistance (GitHub Copilot), full AI-autonomy (GitHub Copilot), and full AI-autonomy (AWS Kiro). Four developers reimplemented a full-stack web application while measuring development effort, requirement adherence (RITM score), interaction efficiency, and cognitive workload (NASA-TLX). Results show increased AI autonomy reduced effort (hours), improved requirement adherence, and lowered mental workload, though with slightly higher frustration. AWS Kiro outperformed GitHub Copilot in full-autonomy mode, suggesting tool architecture influences outcomes beyond autonomy level.
ai-assisted developmentgithub copilotaws kiroritm scorenasa-tlx
ASSEMCAD: Production-Ready CAD Assembly Generation from Natural Language
AssemCAD introduces an axiom-grounded framework for generating production-ready CAD assemblies from natural language, addressing limitations of monolithic CAD code generation. The method constructs an Assembly Specification with typed parts, geometry-backed ports, executable mates, and engineering axioms, enabling interpretable and verifiable designs. A port- and mate-based CAD library executes symbolic relations via deterministic mate transformations, validated by B-Rep geometry. Experiments on AssemBench demonstrate improved assembly preservation and physical validity over code-centric baselines, generalizing across foundation-model backbones.
text-to-cadassembly specificationb-rep geometryaxiom-groundedmate transformations
Agent Data Injection Attacks are Realistic Threats to AI Agents
This paper introduces agent data injection (ADI), a novel category of indirect prompt injection (IPI) attacks where malicious data is disguised as trusted data, enabling AI agents to execute unintended actions. Unlike prior IPI defenses focused on instruction injection, ADI targets security-critical metadata and agent context data, bypassing existing mitigations. The authors identify critical vulnerabilities in real-world agents, demonstrating arbitrary click attacks on web agents (Claude in Chrome, Antigravity, Nanobrowser) and remote code execution/supply-chain attacks on coding agents (Claude Code, Codex, Gemini CLI). Evaluation across off-the-shelf models and AI agents shows ADI's effectiveness, exposing a fundamental security gap: failure to isolate trusted from untrusted data.
agent data injectionindirect prompt injectionsecurity-critical metadataagent context dataremote code execution
Localized LoRA-MoE: Block-wise Low-Rank Experts With Adaptive Routing
The paper introduces Localized LoRA-MoE, a framework combining block-wise low-rank adaptation with dynamic routing to address gradient warfare in parameter-efficient fine-tuning. Two variants are proposed: Block-Wise LoRA-MoE (centralized routing) and Cell-Wise LoRA-MoE (decentralized per-cell gating). Evaluations on SVD simulations, tabular transformations, and vision tasks under sensor degradation show both methods resolve optimization deadlocks in static baselines, with cell-wise routing matching global coordination while providing gradient isolation. The approach outperforms static baselines in dynamic adaptation scenarios.
lora-moeparameter-efficient fine-tuninggradient warfaredynamic routingblock-wise adaptation
Grokking Is Conditional and Fragile: A Fully-Tractable, Multi-Seed Study at 12K Parameters
The study investigates grokking in a fully-tractable 11,856-parameter Llama-style transformer (Glimmer-1-Base) on modular arithmetic, revealing it as a conditional and fragile phase transition. Using multi-seed analysis, the authors demonstrate that grokking is gated by training-set coverage, with thresholds tracking output cardinality more than task structure. Weight decay reproduces the Omnigrok inverted-U curve, while numerical perturbations (CPU thread count, CPU-vs-GPU execution) flip minority outcomes without affecting aggregate rates. Methodologically, multi-seed control overturns single-run narratives, establishing grokking evidence as a rate under fixed numerical environments. Decomposition into sub-task specialists aids coverage efficiency rather than supervision.
grokkingtransformermodular arithmeticweight decayphase transition
AIFS-SUBS: Extending Data-Driven Forecasting to Sub-Seasonal Timescales
The study introduces AIFS-SUBS, a data-driven model extending AI weather forecasting to sub-seasonal timescales (weeks 2–6). Adapting ECMWF's AIFS-CRPS, it employs a 24h autoregressive step, adds stratospheric predictors, and reserves 2007–2011 for verification. Two configurations are evaluated: AIFS-SUBS (fine-tuned on operational analyses) and AIFS-SUBS-ERA5 (trained solely on ERA5). Results show AIFS-SUBS matches the operational IFS in probabilistic skill while reducing biases, extends skillful MJO forecasts by 8 days, and accurately reproduces SSW events. AIFS-SUBS-ERA5 slightly outperforms IFS in ranked probability skill at weeks 3–4, using 200× less energy than IFS for inference.
sub-seasonal forecastingautoregressive modelmaddan-julian oscillationstratospheric predictorsranked probability skill score
Diffusion-Guided Uncertainty-Aware Delayed Policy Optimization
Proposes Diffusion-Guided Uncertainty-Aware Delayed Policy Optimization (DUPO) for reinforcement learning under stochastic delayed feedback. The method models state-discrepancy via diffusion models and uses uncertainty estimates to weight delayed policies, addressing performance degradation in stochastic MDPs. Experiments on continuous robotic control tasks show DUPO outperforms existing methods under multiple stochastic delays, including long and random delay scenarios.
reinforcement learningdelayed feedbackdiffusion modelstochastic mdppolicy optimization
Toward Trustworthy Large Language Model Agents in Healthcare
CareConnect, a safety-first conversational agent for healthcare logistics automation, combines large language model (LLM) function calling, retrieval-augmented generation (RAG), and deterministic safety guardrails to address inefficiencies in appointment scheduling. The system orchestrates eight domain-specific tools for booking, modification, cancellation, and facility information retrieval, while enforcing strict scope constraints and emergency detection mechanisms. Evaluated on 680 task-oriented scenarios, CareConnect achieves a 91.8% task completion rate, 2.2s median latency, 96.0% safety compliance, and $0.0324 average cost per appointment, demonstrating reliable automation with safety guarantees and cost efficiency.
large language modelretrieval-augmented generationdeterministic safety guardrailstask-oriented scenarioshealthcare logistics automation
Beyond Independent Labels: Schwartz-Geometry Decoding for Human Value Detection
The paper introduces Schwartz-geometry decoding, a method for human value detection that operationalizes Schwartz's circular motivational continuum as an output-space geometry. It compares training-time geometry-aware objectives with a post-hoc Schwartz-aware energy decoder, finding the latter improves label set coherence without sacrificing Macro-F1 or Micro-F1. Experiments with DeBERTa-v3-base show the decoder's gains are specific to the true Schwartz ordering, not random permutations or co-occurrence graphs. Diagnostic tests with Qwen2.5-72B-Instruct confirm the continuum's impact at inference, though supervised structured prediction remains superior.
schwartz-geometry decodinghuman value detectionmulti-label classificationenergy decoderoutput-space geometry
RUFNet: Query-Guided Support Mask Refinement and Uncertainty Fusion based on Hybrid Mamba for Few-Shot Brain Tumor Segmentation
RUFNet introduces a Hybrid Mamba-based few-shot framework for brain tumor segmentation, addressing noisy support masks and prediction uncertainty. The method combines an Attention-Guided Mask Refinement module (AGMR) to recalibrate support masks using query features, and an Uncertainty-Aware Posterior Fusion module (UAPF) for pixel-wise variance estimation. The Hybrid Mamba backbone maintains support-query dependencies with linear complexity. Evaluated on BraTS 2020, RUFNet achieves Dice scores of 84.3% (1-shot) and 86.1% (5-shot), outperforming state-of-the-art methods.
hybrid mambafew-shot segmentationmask refinementuncertainty fusionbrain tumor segmentation
LLM-Based Test Oracles: Source-of-Authority Taxonomy -- A Systematic Literature Review
The article presents a systematic literature review (PRISMA 2020) of 54 studies on LLM-based test oracles, categorizing them by source of authority, form, and adjudication mechanism. Analysis reveals that specification-derived authority accounts for only 28 studies, while 26 rely on no specification. The study highlights cross-cutting relationships between authority sources and adjudication mechanisms, challenging labels like 'LLM-as-a-judge' as mechanisms rather than trust bases. Domains, languages, models, and adaptation strategies are characterized, with evaluation methods and failure modes reported. The taxonomy identifies research gaps.
test oracleslarge language modelssystematic literature reviewsource of authorityadjudication mechanism
Your Agent's Memories Are Not Its Own: Forged Reasoning Attacks on LLM Agent Memory and Defenses
The paper introduces FARMA (Forged Amplifying Rationale Memory Attack), a novel attack that poisons LLM agents' reasoning history rather than factual knowledge, using evasive language to bypass keyword defenses and self-referential reinforcement to defeat consensus-based defenses. It proposes SENTINEL, a layered defense pipeline featuring a Reasoning Guard that analyzes entries for forgery using five weighted structural signals. Evaluations across multiple agents and LLM models (50 trials) show FARMA achieves up to 100% success rate against baseline defenses, while SENTINEL reduces this to 0% with no false positives in 326 benign traces.
llm agentsmemory poisoningreasoning integrityevasive attacksdefense pipeline
Hyperparameter Transfer in Graph Neural Networks
The paper develops a hyperparameter transfer framework for graph neural networks (GNNs) to enable consistent hyperparameter settings across model scales. Through theoretical scaling analyses and controlled experiments, the authors propose parameterizations for SGD, Adam, and AdamW optimizers that ensure stable feature updates and learning rate transfer. Key findings include graph-dependent first-layer correction factors for SGD, message passing normalization effects for Adam, and joint transfer of weight decay and learning rate for AdamW, providing practical scaling guidelines for GNNs.
hyperparameter transfergraph neural networksmessage passing normalizationweight decaylearning rate transfer
ImputeECG: Deep Learning Reconstruction of Complete 12-Lead Electrocardiograms from Incomplete Recordings for Cardiac Assessment
ImputeECG introduces a mask-conditioned 1D Transformer autoencoder for reconstructing complete 12-lead, 10-second ECGs from incomplete recordings while preserving observed samples. The model, trained on PTB-XL and evaluated on PTB-XL, CPSC2018, and a 43,633-record Kailuan cohort, reduces missing-region MAE by 41.7-63.7% versus baselines and improves morphological reconstruction (R-peak timing, QRS duration). Downstream classification performance reaches 92.28% AUROC on PTB-XL and 94.75-95.89% on CPSC2018, with real-world validation showing sex prediction AUROC improvement from 82.6% to 85.8%.
transformer autoencoderelectrocardiogram reconstructionmask-conditioned imputationcardiac assessmentmorphological preservation
Comparison of Loss Functions for Robust Deep Learning-based Echocardiography Segmentation when Learning with Partially Labelled Data from Multiple Domains
The study evaluates three loss functions—adaptive categorical cross entropy (aCCE), marginal loss, and adaptive binary cross entropy (aBCE)—for robust echocardiography segmentation when learning from partially labelled multi-domain data. Experiments compare performance across intra-domain and inter-domain tasks, varying label completeness and network architectures. Results show aCCE, marginal, and aBCE perform well intra-domain, while aBCE and marginal excel inter-domain with single missing labels; marginal loss dominates with multiple missing labels, demonstrating scenario-dependent robustness.
echocardiography segmentationpartially-labelled dataloss functionsmulti-domain learningadaptive cross entropy
Quantum-Inspired Harmonic Decision Models: A Computational Framework for Music Generation
The paper introduces a quantum-inspired computational framework for harmonic decision-making in music, formulating harmonization as an optimization problem in a structured combinatorial space. The method combines an interference-based harmonization stage with classical optimization grounded in tonal harmony, enabling parallel consideration of harmonic alternatives while refining sequences for coherence. Evaluations on musical examples like 'Autumn Leaves' demonstrate that optimization reduces chord density, increases harmonic stability, and improves functional organization. Expert evaluation highlights the role of stylistic context, showing that increased harmonic complexity is not always perceived as natural. The framework integrates domain-specific knowledge with interference-based search, contributing to quantum-inspired models of cognition in creative systems.
quantum-inspired harmonizationinterference-based searchtonal harmonychord densityfunctional organization
TACTIC-KG: Toward Small Agent Teams for Cyber Threat Intelligence Knowledge Graph Construction
TACTIC-KG introduces an agentic framework for constructing Cybersecurity Knowledge Graphs (CSKGs) from unstructured Cyber Threat Intelligence (CTI) reports. The method decomposes the task into modular LLM agents (3B--8B parameters) specialized for extraction, typing, verification, and curation, improving over monolithic LLM approaches. Experiments on human-annotated CTI reports demonstrate superior performance in extraction F1-score, typing accuracy, and structural graph similarity compared to in-context-learning baselines.
cybersecurity knowledge graphsllm agentsthreat intelligencemodular extractiongraph consistency
The Map Behind the Flow: Finite-Step Gradient Descent as a Dynamical System
This paper analyzes fixed-step gradient descent as a discrete dynamical system, revealing its role in shaping learning phenomena beyond the small-step gradient-flow limit. Through a hierarchy of exactly solvable models incorporating depth, factorization, width, data coupling, activation, and stochasticity, the study demonstrates how finite-step dynamics drive representation selection and balancing. Key findings include the emergence of a universal Ricker-type map under large-depth scaling, the interpretation of the edge of stability as a bifurcation point, and the identification of learning rate as a structural parameter influencing attractors and representation selection. The results show that finite-step oscillations facilitate alignment, balancing, and movement toward flatter representations in deep networks.
discrete dynamical systemfinite-step gradient descentrepresentation selectionricker-type mapedge of stability
LLM for the development of FCM
A local large language model, Qwen2.5-32B, is leveraged to construct a data-driven fuzzy cognitive map (FCM) by extracting quantitative data from textual inputs. The method involves processing unfiltered hotel reviews from TripAdvisor through the model, training an FCM, and evaluating its performance. A case study on Greek reviews demonstrates the formation of a star topology FCM reflecting reviewer preferences. External validation confirms the FCM's ability to correlate predicted satisfaction with review star ratings, an outcome beyond the model's direct inference scope.
fuzzy cognitive maplocal large language modelquantitative data extractionstar topologyexternal validation
Joint Velocity Slope Diffusion Prior for Structurally Constrained Velocity Model Building
The authors propose a diffusion-guided framework for high-resolution velocity-model reconstruction from sparse well-log data, enhancing subsurface delineation and reservoir characterization. The method integrates plane-wave PDE regularization, structurally preconditioned inversion, and measurement-guided diffusion posterior sampling, utilizing a joint velocity-slope generative prior informed by local structural slopes estimated via plane-wave destruction. Experiments on the Volve synthetic model and Viking Graben field dataset demonstrate improved structural continuity, lateral consistency, and geological realism compared to conventional approaches, while maintaining computational efficiency through DDIM sampling.
velocity-model reconstructiondiffusion posterior samplingplane-wave destructionddim samplingstructurally preconditioned inversion
Multi-Robot Open Adaptive Teaming Across Unseen Environments, Partners, and Scales
The paper introduces open adaptive multi-robot teaming, addressing simultaneous adaptation to unseen environments, unknown partners, and varying team sizes through a hypergraphic-form game formulation. This method extends beyond pairwise interactions to model team-level cooperation, enabling dynamic coordination structure inference. The proposed Hypergraphic Open-ended Learning Algorithm (HOLA) trains by progressively expanding partner and environment diversity. Evaluations on cooperative pursuit tasks with multi-drone and multi-quadruped platforms show HOLA outperforms baselines across adaptability dimensions, with successful zero-shot transfer to Crazyflie and Zsibot L1 hardware.
multi-robot teaminghypergraphic gameopen-ended learningadaptive coordinationzero-shot transfer
STAPO: Selective Trajectory-Aware Policy Optimization for LLM Agent Training
STAPO introduces Selective Trajectory-Aware Policy Optimization, a hierarchical RL framework addressing trajectory neglect in LLM agent training. It employs normalized entropy to identify outlier steps associated with trajectory neglect, distinguishing between inherent state complexity and agent confidence. STAPO optimizes these steps using a joint mechanism of trajectory-aware rewards and trajectory-independent penalties, enhancing trajectory awareness while maintaining training stability. Evaluations on ALFWorld, WebShop, and Search-Augmented QA benchmarks demonstrate STAPO's state-of-the-art performance and robustness in alleviating trajectory neglect for long-horizon agentic tasks.
trajectory neglectnormalized entropyhierarchical rltrajectory-aware rewardagent confidence
MemPose: Category-level Object Pose Estimation with Memory
MemPose introduces a memory-augmented framework for category-level object pose estimation, addressing limitations of fixed shape priors in existing methods. The approach employs an external memory buffer to dynamically store and update structural representations from observed instances, enabling knowledge reuse. Evaluated on REAL275, CAMERA25, Housecat6D, and Wild6D benchmarks, MemPose outperforms state-of-the-art methods in robustness and generalizability.
category-level pose estimationmemory-augmented frameworkexternal memory bufferstructural representationsdynamic updating
DSWAM: A Dual-System World Action Foundation Model for Fine-Grained Robot Manipulation
The paper introduces DSWAM, a Dual-System World Action Foundation Model addressing the decomposition gap in World Action Models (WAMs) by combining a System 1 WAM executor with an optional System 2 vision-language subtask planner. The WAM executor performs world-aware action generation trained via action prediction and video co-training, while the planner decomposes coarse instructions into fine-grained subtasks. The system integrates TensorRT acceleration, asynchronous execution, and real-time chunking for real-robot deployment. Evaluation under the DeMaVLA benchmark shows comparable performance to Vision-Language-Action policies with matched conditions.
world action modelsvision-language-actiontask decompositionreal-time chunkingdeformable manipulation
Input Pathways Shape Few-Shot, Not Zero-Shot, Binding in Tiny Transformers: A Fully-Enumerable Study
This work investigates how input pathways affect compositional binding in tiny transformers (6-10K parameters) on fully-enumerable finite factored worlds, achieving zero sampling variance and exact Bayes ceiling. Four key findings emerge: (1) zero-shot composition fails across all informative routes, indicating inductive bias rather than information deficiency; (2) few-shot binding efficiency is governed by input-pathway parameter sharing and code readability, with clean oracle codes not being most efficient; (3) distributed codes exhibit transient above-chance performance early in training, dissociating from few-shot efficiency; (4) failure modes vary by input type, with symbolic routes losing answers, index routes mis-binding, and entangled routes inheriting input readability. The study releases full reproducibility materials.
compositional bindinginput pathwaysfew-shot learninginductive biasparameter sharing
Efficient Perception in Automotive Detection and Tracking Using Neuromorphic Computing
The study demonstrates the viability of Spiking Neural Networks (SNNs) for energy-efficient automotive perception, achieving competitive performance with conventional deep learning. Using transfer learning with SpikeYOLO, the method attains mean Average Precision of 0.937 (KITTI) and 0.771 (BDD100K MOT2020) for detection, and Higher Order Tracking Accuracy scores of 0.701 (KITTI) and 0.445 (BDD100K MOT2020) for tracking. These results establish SNNs as a sustainable alternative for real-world autonomous systems.
spiking neural networksneuromorphic computingobject detectiontransfer learningautonomous systems
Graph Representation Learning of Longitudinal Medical Imaging Trajectories for Treatment Response Prediction
Proposes a 3D spatio-temporal framework integrating graph neural networks with relational temporal modeling and three novel self-supervised objectives for predicting pathological complete response (pCR) in breast cancer patients undergoing neoadjuvant chemotherapy (NACT). The method leverages longitudinal DCE-MRI imaging trajectories from 585 patients in the ISPY-2 dataset, outperforming vision and self-supervised baselines across multiple classification metrics. Includes ablation studies assessing the impact of timepoint count and inter-scan time differences. Establishes a pCR prediction benchmark and releases code repository and PyPI library for reproducible research.
graph neural networkself-supervised learninglongitudinal medical imagingpathological complete responseneoadjuvant chemotherapy
Medi-Gemma: A Hybrid Clinical Decision Support System Integrating Deterministic EMR Analytics and Retrieval-Augmented Generation
Medi-Gemma introduces a hybrid Clinical Decision Support System (CDSS) combining deterministic EMR analytics with retrieval-augmented generation to address LLM limitations in clinical settings. The system employs a decoupled architecture with a ClinicalOrchestrator coordinating a DataManager for EMR cleaning, an IntentRouter for query classification, and a ClinicalRAGEngine for patient-specific reasoning. Key innovations include a Ground Truth Injection Module for context anchoring and a ProtocolManager for safety compliance. Validation demonstrates elimination of semantic drift, prevention of database crashes, and improved factual adherence to clinical repositories.
clinical decision support systemretrieval-augmented generationelectronic medical recorddeterministic reasoningground truth injection
EventCoT: Event-centric Video Chain-of-thought for Reasoning Temporal Localization
The paper introduces EventCoT, an event-centric video chain-of-thought framework for reasoning temporal localization (RTL). The method first tokenizes input videos into compact event tokens for efficient question-relevant event identification, then generates answers with grounded time intervals via embedding matching between placeholder tokens and visual embeddings. EventCoT achieves state-of-the-art performance on ActivityNet-RTL with fewer visual tokens than prior work and demonstrates strong zero-shot results on ReXTime.
reasoning temporal localizationevent-centric tokenizationembedding matchingchain-of-thoughtzero-shot learning
CARL: Constraint-Aware Reinforcement Learning for Planning with LLMs
The paper introduces Constraint-Aware Reinforcement Learning (CARL), a novel RL framework to enhance Large Language Models' (LLMs) intrinsic constraint awareness during plan generation. CARL employs a constraint-aware reward mechanism by contrasting model outputs under constrained and unconstrained inputs, promoting constraint focus without external tools. Evaluated on BlocksWorld, TravelPlanner, and T-Eval, CARL outperforms Reinforcement Fine-Tuning (RFT) baselines and state-of-the-art reasoning models, demonstrating improved constraint adherence.
constraint-aware reinforcement learninglarge language modelsplan generationreward mechanismintrinsic constraint awareness
SynSFX: Multi-Model Sound Effects Synthesis Dataset for Deepfake Detection and Evaluation
SynSFX introduces a large-scale dataset for evaluating deepfake detection in synthetic sound effects, addressing limitations in existing environmental audio datasets. The corpus comprises 43,374 audio clips (26,452 synthetic, 16,922 real) generated by 7 popular text-to-audio models, enabling comprehensive study of isolated sound-effect deepfakes. This resource supports improved generalization of audio deepfake detectors by providing diverse synthetic samples and real-world counterparts for comparative analysis.
deepfake detectionsynthetic sound effectstext-to-audio modelsenvironmental audioaudio dataset
Pretraining Curricula Enable Selective Fine-tuning
The study demonstrates that imbalanced pretraining curricula enhance task disentanglement and improve selective fine-tuning precision compared to balanced curricula. Using transformers trained on conflicting copy tasks and synthetic language learning, the authors show imbalanced training (one task early, another late) promotes in-context learning and localized neural circuit formation, whereas balanced training yields entangled representations. Activation patching reveals imbalanced curricula produce more separable task pathways, enabling robust refusal fine-tuning and rule-following behavior. Results suggest imbalanced pretraining as a tool for AI safety by facilitating targeted behavior suppression.
pretraining curriculatask disentanglementin-context learningactivation patchingselective fine-tuning
HamQASBench: A Hamiltonian-Informed Diagnostic Benchmark for Evaluating Quantum Architecture Search
HamQASBench introduces a Hamiltonian-informed diagnostic benchmark for evaluating Quantum Architecture Search (QAS) methods, addressing limitations of existing benchmarks that rely solely on energy accuracy and molecular identity. The benchmark organizes 11 molecules into five structural tiers using fingerprints derived from Pauli operator basis, computational basis representation, and ground-state entanglement. It employs a post-hoc critical-structure extraction procedure to identify minimal circuits and evaluates QAS methods through per-qubit entanglement analysis and pairwise state fidelity. Benchmarking five QAS methods across four paradigms reveals previously undetected failure modes, including over-parameterization, eigenstate commitment under degeneracy, and topology-induced routing failure.
hamiltonian-informedquantum architecture searchpauli operator basisentanglement analysiscritical-structure extraction
Evaluating the Effect of Linguistic Relatedness on Cross-Lingual Transfer in Large Multilingual Automatic Speech Recognition
The study evaluates whether linguistic relatedness reliably enhances cross-lingual transfer in large multilingual automatic speech recognition (ASR) for low-resource African languages. Through a controlled experimental design involving six factors, two Africa-centric corpora, and four large ASR models, the authors systematically test sequential adaptation from related auxiliary languages. Results show no practically meaningful transfer improvements with minimal target-language data, suggesting linguistic relatedness alone is insufficient for predicting or enabling effective cross-lingual transfer in large multilingual ASR.
cross-lingual transferautomatic speech recognitionlinguistic relatednesslow-resource languagesmultilingual models
An Exploration of Agentic Information Fusion for Test Maintenance Prediction
The paper introduces MAST, a multi-agent framework for predicting test cases requiring maintenance after production code changes. MAST integrates static, lexical, and semantic analyses through intelligent fusion and post-check procedures, operating without pre-existing test-production mappings. Evaluated on 21 industrial Java repositories from Ericsson AB, MAST outperformed a state-of-the-art baseline in precision, accuracy, F1, and F2 scores, albeit with slightly reduced recall. An ablation study confirmed the contribution of each analysis component. The work demonstrates the efficacy of multi-agent systems in fusing diverse information sources for software testing tasks.
test maintenancemulti-agent systemstatic analysislexical analysissemantic analysis
Predicting Drafted Deck Strength for "Magic: the Gathering"
The paper introduces an encoder-based model for predicting deck strength in Magic: the Gathering Draft, a constrained deck-building format with sequential card selections. The proposed method generates set-contextualized card embeddings to encode draft decision sequences, addressing the challenge of combinatorial card synergies in a dynamic, evolving card pool. Evaluated on large-scale real-world data, the model demonstrates consistent improvements over linear baselines, establishing the first learned benchmark for outcome prediction in MTG Draft.
encoder-based modelcard embeddingscombinatorial synergiessequential selectionoutcome prediction
Multi-Turn On-Policy Distillation with Prefix Replay
The paper introduces Replayed-Prefix On-Policy Distillation (ReOPD), an efficient off-environment method for multi-turn on-policy distillation (OPD) in agentic tasks. ReOPD addresses the prefix trap—a two-sided distribution shift between student occupancy and teacher reliability—by reusing pre-collected teacher trajectories as replayed prefixes and employing a step-decaying sampling schedule to prioritize early, lower-shift prefixes. Evaluated on mathematical reasoning with Python and search environments across multiple model scales, ReOPD matches or exceeds OPD accuracy, eliminates tool calls during training, and achieves at least 4× faster per-step training speed.
on-policy distillationprefix replaydistribution shiftmulti-turn interactionagentic tasks
AgenticPD: A Stage-Aware Agentic Framework for Physical Design QoR Optimization
AgenticPD introduces a stage-aware agentic framework for physical design quality-of-results (QoR) optimization, addressing the costly and interdependent nature of electronic design automation (EDA) flows. The framework employs a Judge Agent to navigate stage boundaries and specialized agents for local decision-making within each stage, leveraging structured observations and execution history to reuse checkpoints and branch from intermediate states. Evaluations show AgenticPD achieves strong post-route timing while maintaining competitive power and area metrics compared to flat parameter tuning or LLM-based script generation approaches.
physical designqor optimizationagentic frameworkeda flowpost-route timing
Trust Region Policy Distillation
The paper introduces Trust Region Policy Distillation (TOP-D), a stable variant of On-Policy Distillation (OPD) that dynamically constructs a proximal teacher to control gradient variance. Theoretically, TOP-D provides formal global convergence guarantees and monotonic improvement bounds, ensuring reliable training dynamics. Empirically, it improves stability, sample efficiency, and final performance on mathematical reasoning tasks without additional computational overhead, offering a viable alternative to OPD.
trust regionpolicy distillationgradient varianceon-policy learningmathematical reasoning
FM-ChangeNet: Learning Change through Pathwise Feature Transport
FM-ChangeNet introduces a pathwise-supervised framework for change detection by reformulating bi-temporal reasoning as continuous feature transport. The method learns a time-conditioned velocity field $\hat{v}_θ(z_t,t)$ along transformation trajectories, constructing intermediate latent states for denser supervision and explicit temporal evolution modeling. A hierarchical architecture incorporates cross-temporal alignment, coarse-to-fine flow decoding, and a unified objective combining flow supervision, trajectory consistency, spatial regularization, and segmentation loss. Experiments on remote sensing benchmarks demonstrate state-of-the-art performance with more structured and robust change representations.
change detectionfeature transportvelocity fieldbi-temporal reasoningpathwise supervision
Wasserstein Residuals: Learning Gradient Flows from Population Dynamics
The paper introduces Wasserstein Residuals, a novel approach for learning Wasserstein gradient flows (WGFs) from population dynamics data. The method formulates a non-negative loss function that enforces continuity equations, combining it with a data-fitting divergence into a single global objective. This residual-based approach avoids the inflexibility and computational costs of JKO-based methods, enabling a new particle-based technique called stitching that is simulation-free and handles large observation gaps. Experiments show state-of-the-art performance on trajectory inference benchmarks.
wasserstein gradient flowcontinuity equationstrajectory inferenceparticle-based methodresidual learning
RustMizan: A Compilable, Contamination-Aware Benchmarking Framework for Rust Vulnerabilities
RustMizan introduces a compilable, contamination-aware benchmarking framework for Rust vulnerability analysis, addressing limitations of existing benchmarks that use non-compilable snippets and lack training contamination controls. The framework provides annotated code variants at crate, file, and function levels, with binary detection, CWE classification, and localization tasks, plus a mutation system for robustness testing. Evaluations with four frontier LLM agents show 56-65% binary classification accuracy but only 20% F1 for line localization, with adversarial cues reducing localization performance by 27%.
rust vulnerabilitycompilable benchmarkcontamination testingcwe classificationline localization
Turning Off-Policy Tokens On-Policy: A Plug-in Approach for Improving LLM Alignment
Selective Importance Sampling (SIS) is proposed to improve large language model (LLM) alignment by converting off-policy tokens into on-policy tokens during reinforcement learning (RL) post-training. SIS employs a token-level rejection test, treating accepted tokens as on-policy with unit importance scores and applying standard importance sampling (IS) correction to rejected tokens. This method reduces the gap between token-level and sequence-level off-policy gradient estimators, acting as a plug-in with minimal computational overhead. Experiments on dense and mixture-of-experts (MoE) LLMs across math and agent benchmarks demonstrate consistent performance improvements and enhanced robustness under off-policy data.
selective importance samplingoff-policy tokenstoken-level rejectionimportance samplingreinforcement learning
Dashboard2Code: Evaluating Multimodal Models on Reconstructing Interactive Dashboards
The paper introduces Dashboard2Code, a novel task evaluating multimodal models' ability to reconstruct interactive dashboards through proactive exploration and code generation. The authors present DashboardMimic, a benchmark with 180 Plotly+Dash dashboard-code pairs across three difficulty levels and eight interaction patterns, alongside an automated evaluation framework combining semantic code analysis with dynamic interaction testing. Experiments reveal significant performance gaps between open- and closed-source models, with all systems struggling on high-complexity dashboards.
dashboard2codemultimodal modelsinteractive visualizationcode generationautomated evaluation
FORGE: Research-Trajectory Hijacking Attacks on Deep Research Agents
FORGE introduces a two-level attack exploiting deep research agents by combining intra-document reasoning fabrication with inter-document chain coordination to hijack subtask planning, enabling report-level contamination. The attack leverages adversarial documents in the retrieval pool to steer follow-up questions, achieving a 26.4% PRISM metric with five injected documents. Root Query Anchoring (RQA), a lightweight defense, reduces PRISM from 38.5% to 18.3% by tying recursive follow-up generation to the root query. Depth migration is observed, where poisoned content shifts from overt framing to factual premises during recursive synthesis.
reasoning fabricationretrieval poolprism metricroot query anchoringdepth migration
Geometry-Aware Motion Latents for Learning Robust Manipulation Policies
The paper introduces GeoMoLa (Geometry-Aware Motion Latents), a method for learning discrete motion latent codes by predicting point cloud evolution during manipulation instead of reconstructing visual observations. This 4D objective (spatial geometry over time) enforces latent representations to encode physical motion rather than appearance patterns. GeoMoLa achieves state-of-the-art performance using single-view RGB-D input, outperforming methods requiring multi-view reconstruction across diverse benchmarks. Ablations confirm geometric prediction drives performance, and learned codes generalize to novel scenes with physically consistent transformations. Real-world experiments demonstrate robustness in cluttered environments with minimal demonstrations.
motion latentspoint cloud predictionrobotic manipulationgeometry-aware learning4d representation
RSPO: Reward-Swap Policy Optimization for Multi-Turn LLM Agents
The paper proposes Reward-Swap Policy Optimization (RSPO), a reinforcement learning method for training large language models (LLMs) in multi-turn interactive tasks with sparse outcome rewards. RSPO employs a reward-swap mechanism to leverage dense process rewards while maintaining alignment with ground-truth outcome rewards, ensuring trajectory diversity and optimization consistency. Experiments on WebShop and ALFWorld benchmarks demonstrate consistent performance improvements across GRPO, PPO, and GiGPO baselines.
reinforcement learningmulti-turn tasksoutcome rewardsprocess rewardspolicy optimization
Integrated Altruistic and Fairness Preference Induces Advanced Mutual Cooperation in Sequential Social Dilemmas
The paper proposes Altruistic and Fairness Preference (AFP), a novel utility function integrating altruistic and fairness preferences from social psychology to enhance cooperation in multi-agent reinforcement learning (MARL). AFP employs a reward-sharing mechanism that converts both individual and others' rewards into cooperative incentives. Experiments in sequential social dilemmas demonstrate AFP agents outperform standard RL and inequity aversion baselines, achieving higher collective rewards (38% increase) and greater equity. Ablation studies reveal altruistic preferences drive public goods contributions while fairness preferences promote mutual behavior.
multi-agent reinforcement learningsocial dilemmasaltruistic preferencefairness preferencereward-sharing mechanism
Strategic Buying Agents
The paper introduces strategic buying agents for delegated purchasing, formulating optimal purchase policies across three information regimes: stationary (known Poisson price adjustments), Bayesian (uncertain price-adjustment distribution), and robust (price bounds only). The stationary regime yields dynamic threshold policies via ODEs, the Bayesian extends this with posterior beliefs, and the robust achieves competitive-ratio guarantees via randomized thresholds. Evaluated on 48,933 Amazon price observations from Keepa, stationary and Bayesian policies achieve competitive mean normalized consumer surplus, while robust excels at the 10th percentile. Language models prove more effective for regime selection than direct buy-wait decisions.
strategic buying agentsdelegated purchasingpoisson arrival processcompetitive-ratio guaranteesnormalized consumer surplus
URSA: Chemistry-Aware Benchmark for Utilitarian Retrosynthesis Assessment
The study introduces URSA (Utilitarian RetroSynthesis Assessment), a chemistry-aware benchmarking framework for evaluating retrosynthesis models. URSA assesses synthetic routes both formally (e.g., convergence to commercially available starting materials) and chemically (mimicking expert evaluation). The framework compares conventional retrosynthesis systems and large language models (LLMs) on novel, diverse target molecules with undisclosed synthetic routes. Results indicate LLMs support high-level strategic planning but underperform specialized retrosynthesis models in reliably solving synthesis planning tasks.
retrosynthesisbenchmarkingdrug discoverylarge language modelssynthetic routes
ToolFailBench: Diagnosing Tool-Use Failures in LLM Agents
The paper introduces ToolFailBench, a diagnostic benchmark for evaluating tool-use failures in LLM agents across 1,000 tasks spanning finance, medicine, law, cybersecurity, and real estate. It classifies failures into Tool-Skip, Result-Ignore, Output-Fabrication, and Unnecessary-Tool-Use using a rule classifier and two LLM judges with majority voting. Results from 19 models show the best achieves 86.33% Clean Tool-Use Rate, with significant variation in failure patterns between models of similar aggregate performance, highlighting the need for nuanced tool-use evaluation.
tool-use failuresdiagnostic benchmarkllm agentsmajority votingclean tool-use rate
Do Vision-Language-Action Models Mean What They Say? On the Role of Faithfulness in Embodied Reasoning
The paper investigates the faithfulness of Vision-Language-Action (VLA) models' Chain-of-Thought reasoning in embodied decision-making, distinguishing functional reasoning (improving performance) from faithful reasoning (reflecting internal processes). It introduces Pinocchio, a learned critic scoring observation grounding and stepwise coherence as a dense RL reward for post-training policies. Evaluations on autonomous driving benchmarks show 4-18% faithfulness improvements over baselines while maintaining task performance, with 1.6x better responsiveness to rare scenarios, suggesting faithful reasoning enhances robustness and interpretability.
vision-language-action modelsembodied reasoningchain-of-thoughtfaithfulnessreinforcement learning
Elastic Gang: Per-Token Membership Change for a Hard-Barriered LLM Inference Gang Co-Scheduled with OS Processes
The paper introduces elastic gang, a scheduling mechanism in Anima OS for dynamically adjusting core membership during LLM inference without deadlock or logit corruption. The method employs an ACK-latched epoch protocol with generation-tagged latches and RCU-style consent, enabling per-token membership changes while maintaining bit-exact output. Evaluated on an AMD Zen 5 (8C/16T) with 135M and 7B models, elastic gang achieves 1.75x general throughput at 25% duty cycle versus static partitioning, with 0.22μs core return latency and saturation at 8-core width.
elastic gangack-latched epochrcu-style consentbit-exact outputper-token membership
Targeted Structure Completion for Sparse-View 3D Reconstruction in Autonomous Driving
FocusGS introduces a targeted structure completion framework for sparse-view 3D reconstruction in autonomous driving, shifting from uniform volumetric processing to localized geometric ambiguity resolution. The method employs a 3D Geometric Ambiguity Manifold to identify occlusion-prone regions, then applies a lightweight module to optimize Gaussian queries exclusively within these sparse subspaces. Experiments show a 74% reduction in Gaussian count and 34% faster rendering while advancing state-of-the-art performance on driving benchmarks.
3d reconstructiongeometric ambiguitygaussian splattingautonomous drivingsparse-view
Machine Learning for Depression Screening and Intervention: an Original Circadian Rhythm Score-based Methodology
The authors propose a Circadian Rhythm Score (CRS) for depression screening, a composite index that unifies multi-domain daily behaviors while preserving behavioral semantics through non-negativity constraints. They develop an interpretable framework using gradient-boosted trees and SHAP analysis, integrating interaction modeling and counterfactual regression for intervention-oriented reasoning. Experiments on the China Health and Retirement Longitudinal Study (CHARLS, n=15,233) demonstrate robust performance (ROC-AUC=0.825) and identify actionable behavioral thresholds, including a minimum effective exercise dose of 300 MET-min/week and an optimal nap duration of 65 minutes for sleep-deprived individuals.
circadian rhythm scoregradient-boosted treesshap analysiscounterfactual regressionbehavioral semantics
Retroactive Chain-of-Thought (RetroCoT): Forensic Reconstruction Prompts as a Safety Diagnostic Across Model Generations
Retroactive Chain-of-Thought (RetroCoT) introduces a forensic reconstruction task as a novel safety diagnostic for large language models, revealing that safety alignment is sensitive to pragmatic framing rather than semantic intent. RetroCoT reframes harmful requests as post-hoc causal reconstructions, achieving attack success rates of 58% on GPT-4o and 52% on GPT-4o-mini, compared to 0% and 4% for direct requests. While GPT-5-family models exhibit robustness to RetroCoT, adversarial feedback within the forensic frame raises success rates to 48% on GPT-5.4-mini and 94% on GPT-4o. These results highlight the fragility of alignment policies across pragmatic registers.
retroactive chain-of-thoughtforensic reconstructionpragmatic framingattack success ratesafety alignment
Formal Disco: Scalable Open-Ended Generation of Formally Verified Programs
Formal Disco introduces a scalable framework for generating formally verified programs using LLM-based workers, addressing data scarcity in verification-aware languages. The system employs three worker classes: initiators sketch verified programs from documentation, fixers resolve issues using compiler feedback, and extenders propose patches to expand working programs. It implements maximum entropy principles for synthetic program generation, iteratively fine-tuning models to enhance diversity. The authors release datasets of synthetic verified programs in Dafny, Verus, and Frama-C, achieving performance comparable to Claude Opus 4.5 in verification tasks. This approach enables large-scale synthetic data creation for formal reasoning domains.
formal verificationllm-based workersmaximum entropysynthetic data generationiterative fine-tuning
Hierarchical Evidence-Driven Reasoning for Long Document Understanding
HIEVI-RAG introduces a hierarchical, evidence-driven multimodal Retrieval-Augmented Generation framework for long-document understanding, addressing key limitations in existing pipelines. The method decomposes complex queries into atomic sub-questions, retrieves candidate pages via a multimodal retriever, verifies evidence using EVIAGENT trained with GRPO, and performs memory-guided iterative generation. Evaluations across four benchmarks show HIEVI-RAG achieves an average accuracy improvement of 8.05% over the strongest baseline, demonstrating robust efficacy in closed-domain document understanding.
retrieval-augmented generationmultimodal retrieverhierarchical decompositioncross-page reasoningmemory-guided generation
MRMS: A Multi-Resolution Memory Substrate for Long-Lived AI Agents
The paper introduces MRMS, a multi-resolution memory substrate for long-lived AI agents, addressing continuity through structured memory organization. The architecture spans representational (structured records, vector representations, graph relations) and temporal axes (short-term, medium-term, long-term), enforcing synchronized structured-vector-graph memory for eligibility, recall, and epistemic labeling. A lightweight prototype demonstrates core mechanisms: pre-generation memory selection, revision, boundary enforcement, and evidence attribution in controlled interaction scenarios.
memory substratevector retrievalgraph relationsepistemic labelingtemporal policies
SILO: Simulation-in-the-Loop Sim-to-Real Transfer for Multi-Stage Cable Routing
The paper introduces SILO, a sim-to-real reinforcement learning framework for multi-stage cable routing that addresses challenges in linear-deformable manipulation. The method leverages GPU-parallelized simulation to train policies across diverse cable geometries and deformation patterns, combined with a novel deployment strategy featuring Simulation In the LOop (SILO) execution and robust cable state estimation. Results show a 2x reduction in cycle times and higher success rates compared to prior learning methods, marking the first successful sim-to-real transfer for this task.
sim-to-real transferreinforcement learningcable routinggpu-parallelized simulationdeformable objects
Governed Individuation: Cryptographically Decoupling an Agent's Learning from Its Authority
The paper introduces governed individuation, a cryptographic method to guarantee an autonomous agent's actions remain within authorized bounds despite learning. The approach binds agents to a frozen identity digest at initialization and gates actions via semantic effect analysis rather than name-based blocking. Theoretical analysis proves the invariant holds even with incorrect self-induced safety principles. Empirical evaluation on an open-ended tool-use benchmark shows zero forbidden effects executed (vs. universal tampering in ungoverned agents), with false-allows dropping from 75% (name-based) to 0% (effect tracing). Compliance transfers to held-out red-line families without task performance degradation.
governed individuationidentity digestsemantic gatingauthority confinementdynamic effect tracing
G2VD: Generalizable AI-Generated Video Detection via Counterfactual Intervention and Causal Disentanglement
G2VD proposes a generalizable framework for AI-generated video detection via counterfactual intervention and causal disentanglement to address shortcut learning in cross-domain scenarios. The method employs a counterfactual intervention pipeline (CFIPipeline) using VAEs with frequency/pixel-domain alignment, coupled with a causal disentanglement classifier featuring domain-anchored branches and HSIC-based independence constraints. Evaluated on four datasets, G2VD achieves 90% accuracy and 0.95 AUC on GenVidBench with only 10% training data, demonstrating strong cross-domain generalization.
ai-generated video detectioncounterfactual interventioncausal disentanglementcross-domain generalizationshortcut learning
Do All Visual Tokens Matter Equally? Object-Evidence Preserving Token Merging for Vision-Language Retrieval
The paper proposes SaMer, an object-aware token merging framework for efficient vision-language retrieval that preserves object-level evidence while reducing token count. The method compresses image-side tokens into K centroids using object annotations as a merge prior during training, without requiring bounding boxes at inference, and maintains the original late-interaction interface. With K=64, SaMer reduces ColPali storage by 16.09× and improves R@1 on Flickr30K and MSCOCO by preserving query-selectable object evidence that pruning or pooling methods lose.
token mergingvision-language retrievallate-interactionobject-evidence preservationmulti-vector compression
LCPNet: Latent Consistent Proximal Unfolding Network for Infrared Small Target Detection
The authors propose LCPNet, a latent consistent proximal unfolding network for infrared small target detection (IRSTD) that improves upon existing deep unfolding methods. Key innovations include: (1) performing unfolding in latent space to preserve physical constraints while avoiding intermediate state compression, (2) a latent consistent proximal solver that evolves variables from previous states with task-adaptive normalization, and (3) shared optimization memory for coordinated guidance across stages. Experiments on four benchmarks show state-of-the-art performance in accuracy, robustness, and false alarm reduction.
infrared small target detectiondeep unfoldinglatent representationproximal solveroptimization memory
TORINO: Token Reduction via Interpretable Concept Overlap in Vision-Language Models
TORINO introduces an interpretable token reduction framework for Vision-Language Models (VLMs) that dynamically reduces visual tokens without fine-tuning. The method employs Sparse Autoencoders (SAEs) to project tokens into a latent space, grouping them by concept overlap—measured through shared SAE latent activations—before pruning or merging redundant tokens. Unlike fixed-budget approaches, TORINO adapts the reduction rate to input complexity. Experiments demonstrate efficient token reduction with minimal accuracy loss across multiple vision-language benchmarks.
token reductionsparse autoencodersvision-language modelsconcept overlapinterpretable latent space
Simple-to-Complex Structured Demonstrations for Vision-Language-Action Learning
The paper proposes a simple-to-complex structured demonstration strategy for Vision-Language-Action (VLA) models in robotic manipulation, addressing the overlooked aspect of demonstration organization in imitation learning. The method decomposes tasks into sub-skills, standardizes environments, and organizes demonstrations by increasing complexity, enabling progressive skill acquisition. Evaluated on block grasping/sorting and towel folding tasks, the approach improves success rates and training stability compared to end-to-end trajectory baselines, highlighting demonstration organization as a critical factor in VLA learning.
vision-language-action modelsimitation learningrobotic manipulationdemonstration organizationtask decomposition
Attention Limited Reward Learning
The paper introduces a rational inattention model for human pairwise comparisons in AI alignment, challenging standard Bradley-Terry reward modeling. It demonstrates that limited attention conflates reward ambiguity with evaluation difficulty, distorting learned preferences. Analysis reveals that passive comparison data cannot disentangle reward, attention, and default biases, with heterogeneous attention causing misleading rankings in standard models. Empirical studies on Chatbot Arena and perceptual comparisons show cyclic preference patterns and gaze/time signals unaccounted for by scalar rewards. The work frames human feedback as attention-limited measurements rather than revealed preferences.
reward learningrational inattentionbradley-terry modelhuman feedbackpreference modeling
LLM-Driven CI-CD Workflow Intelligence for Cyber Systems Engineering
The paper introduces an LLM-based pipeline for analyzing CI/CD workflows, combining repository enrichment, anti-pattern detection, stage mining, and recommendation generation. The method processes 127,559 configuration files from 34,225 GitHub repositories (≥1k stars), identifying 434,769 anti-patterns (primarily reliability/maintainability issues) across 75,201 workflows. Stage usage varies significantly by language (χ²=4168.88, p<0.001) and domain, with mobile projects showing distinct operational profiles. Few-shot prompting yields 8.25 recommendations/repository (96.1% YAML-valid), demonstrating the need for context-aware CI/CD observability beyond stage classification.
ci/cd workflowsanti-pattern detectionstage miningfew-shot promptingyaml-valid snippets
A Few Teacher Steps Go a Long Way: Cost-Efficient On-Policy Data Augmentation for Agent Post-Training
The paper proposes a cost-efficient approach for on-policy data augmentation in LLM agent post-training by framing teacher supervision as a budget-allocation problem. It introduces design parameters (rollout policy, switch-time distribution, continuation horizon, filtering rules) and evaluates trade-offs between teacher inference costs and retained supervision. Experiments on HotpotQA, ALFWorld, and Terminal-Bench-Dev show that bounded unfiltered teacher continuations at learner-induced contexts outperform pure behavioral cloning, with few-step continuations matching or exceeding filtered alternatives at matched budgets.
behavioral cloningon-policy data augmentationteacher supervisionrollout policyswitch-time distribution
Detecting Answer-Driven Reasoning in LLM-Based Educational Tutors via Truncated Chain-of-Thought Auditing
The study introduces Truncated Reasoning AUC Evaluation (TRACE) to detect answer-driven reasoning in LLM-based educational tutors, where explanations may be influenced by access to answer keys rather than derived from student-facing problems. Using Qwen2.5-3B-Instruct on 1000 GSM8K test problems, TRACE measures how early a chain-of-thought prefix can yield the correct answer. Results show answer-key access increases median TRACE AUC from 0.375 to 0.900, with 997 cases revealing the gold answer within the first 10% of the explanation, even when both question-only and answer-key explanations are correct.
truncated reasoning auc evaluationanswer-driven reasoningchain-of-thought auditingllm-based tutorsgsm8k
Heaviside Continuity of Rolling Coefficients for Eliminating Epistemic Entropy in Large Language Models
The paper introduces Heaviside Continuity of Rolling Coefficients (HCRC), a verification-first execution framework for large language models (LLMs) that prevents invalid intermediate states from propagating. HCRC uses predicate-gated state transitions governed by a Heaviside Gate, combining model confidence with independent verification signals from parallel workers. Evaluated on software-engineering and reasoning tasks across 13 proposers, HCRC reduces false-completion rates from 4-7% to 0% on capable models while maintaining competitive latency, and converts false completions into halts on weaker models. The framework has been deployed in production for agentic coding environments.
heaviside gateepistemic entropyautoregressive decodingverification-first executionfalse-completion rate
EEG-SpikeAgent: Agentic Closed-Loop Program Synthesis for Automated EEG Spike Detection
EEG-SpikeAgent introduces an agentic closed-loop program-synthesis framework for automated EEG spike detection, combining LLM-driven feature generation with iterative refinement via performance feedback. The system proposes deterministic signal-processing modules, evaluates them via tabular classifiers, and refines features based on structured diagnostics. Evaluated on VEPISET (29-channel EEG, 25,449 epochs), it achieved 0.935 AUROC, 0.699 balanced accuracy, and 0.996 specificity at default thresholds, with artifact-aware features improving F1 (0.557) over spike-only search. This approach enables auditable, code-driven feature engineering.
eeg spike detectionprogram synthesisagentic systemfeature engineeringinterpretable ai
Predicting Therapeutic Outcome via Aligning Patient-Specific Knowledge Graph and Gene-Level Perturbation Representations
PREDIKTOR introduces a patient-centered multi-view framework for predicting therapeutic response by aligning personalized knowledge graphs with gene-level perturbation representations. The method constructs individualized gene regulatory networks from tumor expression, augmented with drug-target links, and generates simulated post-perturbation transcriptomic profiles using a pretrained attention model. These views are aligned via a CLIP-style contrastive objective, yielding interpretable embeddings for response classification. Evaluated on TCGA and I-SPY2, PREDIKTOR improves AUROC by 5.6% over baselines and provides stable gene-pathway attributions.
knowledge graphgene regulatory networkcontrastive learningtranscriptomic perturbationprecision oncology
Lights, Camera, Carbon: Architectural Scaling Laws for Video Generation Energy Consumption
The authors propose a bidirectional framework for estimating energy consumption in text-to-video (T2V) and text-to-video-audio (T2VA) models based on architectural principles and generation parameters, without requiring access to model weights or implementation details. The method combines forward prediction from parameters and backward recovery of scaling behavior from inference times, leveraging the compute-bound nature of video diffusion models. Validation across six models (8.3B-27B parameters) and three GPU configurations shows below 3% mean absolute percentage error (MAPE) in energy profile decomposition into quadratic and linear terms reflecting architectural complexity.
text-to-videoenergy consumptionscaling lawsdiffusion modelsarchitectural complexity
Explainable Novel Category Discovery in Semantic Concept Space
The paper introduces xNCD, an explainable novel category discovery framework that operates in a semantic concept space rather than opaque latent features. The method aligns visual features with vision-language similarity priors from pretrained multimodal models, then applies a unified self-labeling objective over concept-space logits to enable cluster- and instance-level explanations. Theoretical analysis shows the concept bottleneck restricts the hypothesis class to semantically interpretable partitions. Experiments on CIFAR-10, CIFAR-100, and CUB-200 demonstrate competitive performance (92.63% accuracy on CIFAR-10 vs UNO's 93.4%) while providing human-readable explanations absent in prior work.
novel category discoverysemantic concept spacevision-language alignmentself-labelingconcept bottleneck
Mask2Real-WM: Segmentation Masks as a Sim-to-Real Bridge for Controllable Dexterous World Models
Mask2Real-WM introduces a two-stage action-conditioned world model for dexterous manipulation, decoupling pixel prediction into dynamics and rendering models. The dynamics model predicts future segmentation masks from past masks and 23-DoF actions, while the rendering model maps masks to RGB using a ControlNet-augmented Stable Video Diffusion backbone. Pretraining on 50+ hours of synthetic data and fine-tuning on <2.5 hours of real data enables per-DoF controllability, outperforming monolithic baselines on a pick-and-place benchmark.
world modelsdexterous manipulationsim-to-realsegmentation maskscontrolnet
Auto: The AGI Compiler
The paper introduces Auto, an AGI compiler that autonomously converts LLM agent behavior into verified, near-free executable skills. The system records live agent traces, identifies deterministic segments, and compiles them into WebAssembly binaries with sandbox-enforced guarantees. A tiered runtime executes compiled behavior with fallback to the reference agent when guards trip, enabling iterative recompilation. On AUTO-BENCH (560 tasks), 87.1% of frontier-agent spans were witnessed-deterministic. A 300-item stream showed 6.4x cost reduction (59→2 μ$/item) at 96.9% parity, with failures attributed to guard calibration and reference fidelity issues.
agi compilerwebassembly artifactsdeterministic behaviortiered runtimeauto-bench
CRISP: A Spatiotemporal Camera-Radar Backbone for Driving via Forecasting-Based World-Model Pretraining
CRISP introduces a spatiotemporal camera-radar (CR) backbone for autonomous driving via forecasting-based pretraining, learning unified bird's-eye-view (BEV) representations by predicting future LiDAR point clouds from historical multi-view images and radar sweeps. The method employs an enhanced radar encoder, radar-enhanced temporal self-attention, and modality innovation gating for effective CR fusion. Evaluated on nuScenes, CRISP improves long-horizon point cloud forecasting and transfers to downstream tasks like 3D detection, tracking, and motion forecasting, demonstrating scalable representation learning under practical sensor configurations.
camera-radar fusionbird's-eye-viewforecasting-based pretrainingspatiotemporal backbonemodality innovation gating
Obey, Diverge, Collapse: Blind Obedience to Incorrect Instructions Drives Code LLMs to Irrecoverable Code Semantic Collapse
The study investigates code language models' behavior when given incorrect instructions, revealing a blind obedience phenomenon where models follow flawed directives despite recognizing their inaccuracy. Using the RunBugRun dataset of Python problems with deterministic test cases, the authors conduct four experiments assessing model compliance in single-pass and iterative repair scenarios. Results show models introduce ghost errors through compliance, leading to irrecoverable semantic corruption in subsequent self-repair attempts, with extended reasoning failing to mitigate the collapse. This behavior remains undetected by standard pass-rate evaluations, posing risks for production deployments.
code language modelsblind obedienceghost errorsiterative repairsemantic collapse
Lyapunov-Guided Training for Hardware-Safe Neural Networks Under Fixed-Point Arithmetic
The paper introduces a Lyapunov-stabilised quantisation framework for low-precision neural networks operating under hardware-style wrapping arithmetic, addressing the instability caused by two's-complement overflow in fixed-point arithmetic. The method monitors hidden-state energy via layerwise Lyapunov functions and applies monotone projections to enforce bounded, non-increasing state evolution. Evaluated on MNIST using a compact patch-based transformer, the framework suppresses activation overflow to below 0.012% and restores stable learning, achieving 86.55% accuracy at 12 bits, compared to near-chance accuracy in unconstrained quantisation-aware training. Results demonstrate the effectiveness of Lyapunov-based state control for reliable fixed-point neural inference and training.
lyapunov-stabilised quantisationfixed-point arithmetictwo's-complement overflowmonotone projectionlow-precision neural networks
Measuring Harness-Induced Belief Divergence in Multi-Step LLM Agents
The study introduces a belief-rollout diagnostic to measure harness-induced belief divergence in multi-step LLM agents, focusing on how different harnesses alter agent beliefs despite fixed tasks, environments, and base LLMs. The method decomposes divergence into arrival and growth terms, evaluating trajectories across progress, risk, recoverability, and other dimensions. Results on coding tasks and benchmark stress tests reveal that blocked actions, compressed repairs, and selective verification often preserve terminal success while modifying beliefs driving subsequent decisions. The BIWM protocol is proposed to align belief trajectories across harness views, emphasizing harness design as a critical experimental variable in agent evaluation.
belief-rolloutharness-induced divergencemulti-step llm agentsbiwm protocolverification masks
Training-Free Model Selection and Domain-Aware Score Calibration for First-Shot Anomalous Sound Detection
The paper introduces a training-free post-hoc method for first-shot anomalous sound detection in DCASE Challenge Task 2, addressing two key issues: negative correlation between source- and target-domain AUC, and poor development-set performance prediction. The method employs per-domain quantile calibration with a pooled map prior and a label-free cross-validated domain-balance criterion for configuration ranking. Results on DCASE 2025 show a Spearman correlation of +0.91 between the criterion and official evaluation scores, improving evaluation scores from 55.83 to 61.05. Replication on DCASE 2023 and 2024 confirms the method's robustness, though predictive evidence varies by year. A forward test on DCASE 2026 awaits evaluation ground truth release.
anomalous sound detectionquantile calibrationdomain-balance criterionspearman correlationdcase challenge
Language Models Represent and Transform Concepts with Shared Geometry
The study demonstrates that large language models share a common geometric structure in how context transforms concept representations, revealing richer organization than previously recognized. Drawing from neural population geometry, the authors formalize concepts as point-cloud manifolds and contextual transformations as vector fields, analyzing six model families of varying scales. Results show that context moves each concept differently, with variance in displacements semantically organized and correlating with lexical concreteness and density. Crucially, displacement structures transported from one model predict held-out displacements in others significantly above chance, indicating shared transformation geometry across models.
point-cloud manifoldsvector fieldslexical concretenessdisplacement structureneural population geometry
Failures and Successes to Learn a Core Conceptual Distinction from the Statistics of Language
This study investigates whether language models can learn the principled-vs-statistical distinction—a core conceptual ability humans possess—from linguistic statistics alone. The authors evaluate model sensitivity to this distinction by analyzing their performance on generic statements that are true either by category membership (principled) or statistical regularity. Results show that while all language models are sensitive to statistical prevalence, they struggle to represent the principled-vs-statistical distinction when prevalence is controlled for. Notably, GPT-4 succeeds in this task, suggesting advanced models can bootstrap core conceptual distinctions from language experience.
principled-vs-statistical distinctionlanguage modelsstatistical prevalencegeneric statementsgpt-4
VLA Grounder: Language-Conditioning Space Optimization for Black-Box VLA Models
The paper introduces VLA Grounder, a method for improving frozen Vision-Language-Action (VLA) models by optimizing language-conditioning space rather than action weights. It employs a language-conditioning space policy that translates human instructions into VLA-grounded commands using object appearance, spatial relations, and target-grounding cues. Initialized with a failure-derived prior and optimized via reinforcement learning with sparse rewards, the approach demonstrates improved success rates on RL4VLA and VL-Think benchmarks for instruction-sensitive, symbolic, and multi-object manipulation tasks.
vision-language-action modelslanguage-conditioning spacereinforcement learningobject manipulationfrozen policies
Transplanting, inverting, and preventing a misalignment persona: method-conditional emergent misalignment in Qwen2.5
The study demonstrates that emergent misalignment (EM) in Qwen2.5 models is mediated by a latent persona direction, which is causal in open weights. By transplanting this direction into a pretrained model, broad EM is induced (2.83% misaligned vs. 1.1% random baseline), while ablation halves misalignment (21% to 10%). Low-rank LoRA on insecure code recruits the persona (3.4% misaligned), whereas full SFT does not (0.3%). Steering away from the persona during training increases misalignment (24% to 51%), showing its conditional role. Prevention is possible via inoculation (4.75% to 0.0%) or orthogonal fine-tuning.
emergent misalignmentlatent persona directionlow-rank lorafull sftinoculation
Compressing the Validation Bottleneck: An Agentic Self-Driving Lab for Scientific Discovery
The paper proposes an agentic self-driving lab (SDL) framework to address two physical bottlenecks in AI-driven scientific discovery: inefficient experiment selection and high measurement costs. The method combines a prior-aware design of experiments (DOE) loop that leverages domain knowledge for optimal trial selection, with a cost-aware surrogate agent that predicts high-resolution measurements from low-resolution proxies. Evaluations in biology and materials science demonstrate reduced trials-to-target and measurement costs. The unified agent architecture aims to accelerate SDL pipelines by jointly optimizing loop count and per-experiment expense.
self-driving labdesign of experimentssurrogate modelingmeasurement costagentic ai
Why Pure Reasoning is Not Enough: Nature as the Source of Mathematical Innovation
The article hypothesizes that human mathematical reasoning relies on pattern matching from external domains like nature, due to computational intractability and undecidability in pure deduction. It traces the historical development of the Fourier transform and related mathematics, showing how physics problems necessitated tools that formal reasoning failed to anticipate. The authors survey logical complexity, from NP-hard SAT to non-elementary monadic second-order theories, arguing that physics-inspired pattern matching is cognitively necessary. They conclude that AI systems aiming for mathematical creativity must incorporate cross-domain patterns, justifying the scale of large language models.
mathematical reasoningcomputational intractabilityfourier transformlogical complexitypattern matching
Two Black Boxes, One Solver: Encoder Probing and Decoder Attribution for Neural Multi-Attribute VRP under Hard-Mask and Recourse Decoders
The paper introduces a unified protocol for explaining neural autoregressive solvers for Multi-Attribute Vehicle Routing Problem (MAVRP), addressing both encoder representations and decoder decisions. For encoders, it employs linear probes, organization metrics, and intervention analyses to characterize latent constraint representations. For decoders, it applies gradient-based attribution methods (gradient, integrated gradients, DeepLIFT) across three interpretative angles. Evaluations on six encoder-decoder combinations reveal that graph inductive bias enhances representational predictability, Mixture-of-Experts encoders distribute constraints non-axis-alignedly, and Recourse training enables useful infeasibility representations absent in Hard-Mask policies.
multi-attribute vrpencoder probingdecoder attributionhard-mask decoderrecourse decoder
PulmoSight-XAI: An Explainable Multi-View Attention Ensemble with Gradient Boosting Meta-Learning for Multi-Label Chest X-Ray Classification
The authors propose PulmoSight-XAI, an explainable multi-view ensemble framework for multi-label chest X-ray classification, addressing challenges of class imbalance, co-occurring pathologies, and localized feature loss. The method employs view-specific training with five CNN ensembles, multi-scale feature fusion augmented with CBAM, and a hybrid objective combining Asymmetric Loss with Adaptive Focal Loss. Hierarchical meta-learning integrates test-time augmentation predictions and cross-model uncertainty measures into Level-1 gradient-boosting meta-learners, followed by Level-2 stacking. Evaluated on a CheXpert-style dataset, it achieves state-of-the-art macro-average AUROC scores of 0.9319 (frontal) and 0.9154 (lateral), with explainability analysis demonstrating anatomical consistency.
multi-view ensembleconvolutional block attention modulesgradient-boosting meta-learnerstest-time augmentationasymmetric loss
Regime-Conditional Stabilisation of LLM-Augmented Cooperative Multi-Agent Reinforcement Learning
The paper identifies and mitigates training instability in LLM-augmented cooperative MARL caused by dynamic reward shaping violating PBRS stationarity. It proposes two stabilisation methods: Phase-Based Freeze Schedule enforcing strict stationarity and EMA smoothing bounding weight drift. Evaluations across three environments (Simple Spread, Level-Based Foraging, SMAC 3m) with QMIX/VDN reveal a three-regime taxonomy: augmentative (EMA improves success +12.3pp), essential (any shaping unlocks task), and supplementary (stabilised shaping preserves performance). Results establish reward-signal stationarity as a critical design constraint.
multi-agent reinforcement learningreward shapinglarge language modelsexperience replaystationarity
Operator-on-F complements value-equivalence: a planning-time diagnostic for latent world models
The paper introduces operator-on-F, a planning-time diagnostic for latent world models that complements value-equivalence metrics by measuring k-step latent pushforward errors on observable subsets. The method compares a model's predictions against the environment's using the model's own predictor, revealing planning-relevant errors that reward and value metrics miss. Results on TD-MPC2 size sweeps show operator error varies 10x (0.28-2.62) versus 3x for reward error (0.028-0.091), with -0.90 Spearman correlation to return loss, while Bellman residual and reward error show weak relationships (-0.10, -0.30).
latent world modelsplanning-time diagnosticoperator-on-fvalue-equivalencetd-mpc2
Robustness Verification of an Autonomous Underwater Vehicle-based Plankton Classifier
The authors propose a robustness verification framework for in-situ plankton classifiers deployed on autonomous underwater vehicles (AUVs), addressing misclassification issues caused by environmental noise in dynamic marine environments. Their method combines reachability analysis with a continuous-time neural ordinary differential equation (neural ODE) model optimized for high-resolution SilCam particle imaging. Experimental results demonstrate formal verification of model stability against perturbations, reducing manual validation workload by providing guaranteed robustness for ambiguous data samples.
robustness verificationneural odereachability analysisautonomous underwater vehicleplankton classification
A Deep Learning-based surrogate model for Severe Accidents in nuclear reactors using ASTEC
The authors present a deep learning-based surrogate model (SM) to accelerate severe accident simulations in nuclear reactors, achieving real-time performance while maintaining accuracy. The SM combines an autoencoder for dimensionality reduction (300× compression) with neural ODEs for temporal modeling, trained on ASTEC-generated data for station blackout and loss-of-coolant scenarios. Results demonstrate stable autoregressive predictions across 50k timesteps for 80+ physical variables, reducing simulation time from days to under a minute while covering 40-hour scenarios.
surrogate modelingautoencoderneural odenuclear reactordimensionality reduction
From Regulation to Requirements: An Automated Requirement Derivation and Explanation Pipeline
The paper presents Reg2Req, an automated pipeline for deriving software requirements from regulatory texts with traceable explanations. The method processes legal clauses through requirement-bearing clause identification and system-agnostic requirement generation, evaluated on GDPR (398 clauses) and EU AI Act (574 clauses). Results show F1 scores of 0.82 (GDPR) and 0.78 (EU AI Act) for clause identification, with human-rated completeness (4.60/5) and correctness (3.74/5) for derived requirements. The tool significantly improves practitioner comprehension (p < 0.001) in a 25-participant study.
requirement derivationregulatory compliancetraceability matrixlegal text processingsoftware requirements
Wan-Streamer v0.2: Higher Resolution, Same Latency
Wan-Streamer v0.2 introduces a latency-preserving upgrade for real-time audio-visual interaction, increasing output resolution from 192x336 to 640x368 while maintaining 200 ms model-side latency at 25 FPS. The system employs a thinker-performer architecture: a single-GPU 'thinker' handles low-latency streaming perception and Transformer-based language/state processing, while a multi-GPU 'performer' uses Ulysses-style context-parallel processing for high-resolution latent video generation. Visual sequences are sharded across GPUs for denoising, while audio generation avoids sequence sharding. This design achieves 550 ms total remote interaction latency (including 350 ms network budget) by concentrating additional hardware on visual generation without compromising the thinker-performer boundary.
latency-preservingthinker-performerulysses-stylecontext-parallelstreaming perception
Generative wave propagator
The authors propose a conditional diffusion-based wavefield propagator for seismic wavefield simulation, addressing limitations of finite-difference methods in numerical dispersion and stability. The model conditions on recent wavefield snapshots, velocity models, and time indices, enabling single-step predictions without iterative reverse diffusion. A causal time-weighted loss enhances stability during recursive rollouts. Experiments on Overthrust, SEG/EAGE, and Marmousi models demonstrate accurate wavefield reproduction and a 2.17× speedup over GPU-accelerated tenth-order staggered-grid FD methods.
seismic wavefieldconditional diffusionfinite-difference methodsrecursive rollouttime-weighted loss
ResearchStudio-Idea: An Evidence-Grounded Research-Ideation Skill Suite from ML Conference Outcomes
ResearchStudio-Idea introduces a reusable skill suite for evidence-grounded research ideation in machine learning, comprising Paper-Search (multi-source literature search), Scoop-Check (prior-art collision checker), and IdeaSpark (end-to-end workflow). The system is built from 1,947 ML conference papers (ICLR, ICML, NeurIPS 2021-2025), identifying 31 ideation sub-patterns consolidated into 15 reusable patterns. IdeaSpark evaluates evidence readiness, reconstructs research contexts, identifies bottlenecks, and generates traceable proposals via structured cards. Automated evaluations show it outperforms no-skill and generic-skill baselines in proposal strength while maintaining novelty.
research ideationevidence groundingprior-art collisionstructured patternsautomated evaluation
ResearchStudio-Reel: Automate the Last Mile of Research from Paper to Poster, Video, and Blog
ResearchStudio-Reel automates research dissemination by composing five modular skills into a unified pipeline for converting papers into editable posters, videos, and blog posts. The system employs a shared extractor (Paper2Assets) and three generators (Paper2Poster, Paper2Video, Paper2Blog) that maintain factual consistency, followed by an interactive convergence layer (Paper2Reel) for synchronized viewing. Evaluated on the Paper2Poster benchmark, it outperforms prior automated systems and frontier LLMs, winning 84-93% of comparisons on aesthetic and informational criteria while uniquely delivering all three editable artifacts through layout-aware DOCX repair and bilingual blog generation.
research disseminationmodular skillsfactual consistencylayout-aware repairbilingual generation
A Retrieval-Augmented Framework for Detecting and Resolving Pragmatic Ambiguities in Natural Language Requirements
A retrieval-augmented framework is proposed for detecting and resolving pragmatic ambiguities in natural language requirements (NLRs), addressing misinterpretations arising from varying stakeholder domain expertise. The method employs retrieval-augmented generation with domain-specific knowledge bases (novice, intermediate, expert) to simulate stakeholder interpretations, followed by expert-guided disambiguation and analyst validation. Evaluated on the PUblic REquirements dataset using GPT-4o-mini, Mistral-7B, Llama-3.1-8B, and Qwen2.5-7B, the approach demonstrates effective ambiguity detection (GPT-4o-mini achieves 0.75 recall and F2) and generates relevant, clear, and consistent disambiguated requirements (Mistral-7B excels in clarity and consistency).
retrieval-augmented generationpragmatic ambiguitynatural language requirementsdomain knowledge basesdisambiguation
RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies
The authors introduce RoboDojo, a unified benchmark for evaluating generalist robot manipulation policies across simulation and real-world environments. The benchmark comprises 42 simulation tasks and 18 real-world tasks, assessing five dimensions: generalization, memory, precision, long-horizon execution, and open-vocabulary instruction following. It features scalable evaluation via heterogeneous parallel simulation in Isaac Sim and RoboDojo-RealEval, a reproducible real-world system with standardized hardware and cloud access. The study evaluates 30 integrated policies, establishing a public leaderboard and systematic performance analysis.
generalist robot manipulationsim-and-real benchmarkheterogeneous parallel simulationopen-vocabulary instruction followinglong-horizon execution
Covert Trait Propagation Is Representation Alignment: Mechanistic Evidence from Hidden-Channel Distillation
The study introduces covert trait propagation (CTP), a mechanism where a student model trained on uniform noise inherits digit-classification ability from a teacher model when sharing initialization, mediated by geometric alignment of hidden representations. Using MLP distillation on MNIST, the authors demonstrate that shared initialization enables the output projection W_2 to act as a coordinate key, while KL gradients reshape the student's input projection W_0 to align with the teacher's representations. Five experiments validate CTP, showing channel closure correlates with weight drift (not teacher accuracy), freezing W_0 disrupts transfer, multi-teacher ensembles cancel out, and linear CKA tracks student accuracy (r=0.98). The geometric framework also explains cross-token behavioral entanglement in instruction-tuned LLMs.
covert trait propagationgeometric alignmenthidden-channel distillationcentered kernel alignmentcross-token behavioral entanglement
On Pairwise Quantile Regression -- Statistical Guarantees and Applications
The paper extends quantile regression to pairwise settings where the response variable is a similarity function between two observations and covariates are pairs of features. It proposes empirical minimization of a pairwise pinball loss, establishing theoretical guarantees through U-process concentration bounds that yield fast learning rates under mild conditions. Simulation experiments validate the approach, with applications demonstrated in facial recognition similarity scoring error analysis.
quantile regressionpairwise learningu-processespinball losssimilarity scoring
evalci: A Python Library for Statistically Rigorous Comparison of Language Model Evaluations
The evalci Python library addresses statistical shortcomings in language model evaluation by providing rigorous comparison tools for per-item results tables. It implements established statistical methods (confidence intervals, paired significance tests, power analysis) with adapters for lm-evaluation-harness and HELM outputs, validated against independent references. A re-analysis of nine models' MMLU accuracy reveals 3/8 adjacent leaderboard gaps lose significance after multiple-comparison correction, demonstrating the tool's utility in preventing overconfident claims from noisy benchmarks.
language model evaluationstatistical significancemultiple-comparison correctionconfidence intervalspaired permutation test
dOPSD: On-Policy Self-Distillation for Diffusion Language Models
The paper introduces dOPSD, an on-policy self-distillation method for diffusion language models (dLLMs) that addresses limitations in supervised fine-tuning and reinforcement learning. By deriving the teacher's privileged information from the student's own denoising trajectory—using later decoding steps to evaluate masked positions—dOPSD provides dense, token-level supervision without external labels. Evaluations on Dream and LLaDA benchmarks show improvements in both in-domain math reasoning and out-of-domain code generation, outperforming supervised and on-policy baselines.
diffusion language modelson-policy self-distillationdenoising trajectorytoken-level supervisionmath reasoning
UI-MOPD: Multi-Platform On-Policy Distillation for Continual GUI Agent Learning
The paper introduces UI-MOPD, a method for continual learning in GUI agents that employs multi-teacher on-policy distillation to address cross-platform adaptation challenges. It constructs Uni-GUI, a high-quality dataset for cross-platform interactions, and uses platform-conditioned distillation to transfer platform-specific behavioral priors to a shared policy. Evaluations on OSWorld and MobileWorld demonstrate task success rates of 38.2% and 12.0%, respectively, showing effective balance between capability retention and new-platform adaptation.
gui agentson-policy distillationcontinual learningcross-platform interactionbehavioral priors
Transferability Between Understanding and Generation in Unified Multimodal Models
The study demonstrates systematic transferability between understanding and generation tasks in Unified Multimodal Models (UMMs), showing that training one task improves performance on the other without explicit supervision. Through controlled experiments, the authors identify that architectures with fully shared transformer backbones and unified visual encoders exhibit consistent cross-task transfer, unlike loosely coupled designs. They propose a training strategy where fine-tuning the understanding task transfers to generation, improving capability-specific performance while minimizing distribution shift. This approach is validated across three capabilities: counting, spatial relations, and text recognition/generation.
transferabilityunified multimodal modelstransformer backbonevisual encoderdistribution shift
Full-Stack FP4: Stable LLM Pretraining with Quantized Projections, Optimizers, and Attention
Full-Stack FP4 introduces the first complete NVFP4 pretraining framework addressing stability bottlenecks in transformer linear layers, optimizers, and attention mechanisms. The method employs LoRA-SVD decomposition for linear projections to reduce quantization noise, a transformed AdamW second-moment storage for robust NVFP4 optimization, and a mixed-precision scheme for attention to maintain forward-backward consistency. The framework demonstrates near-BF16 performance with a 1.47% loss gap in 3B/64B-token pretraining, validating stable end-to-end NVFP4 LLM pretraining.
nvfp4lora-svdadamwmixed-precisionpretraining
Agent Step Value: State-Transition Measurement with State-Grounded LLM Evaluators
The paper introduces Agent Step Value (ASV), a state-transition measurement framework for evaluating individual actions in multi-step agent trajectories. ASV employs state-grounded LLM evaluators to score actions based on changes in log-probability distributions over candidate outcomes, using redacted state projections and a label-free rationale pass. Evaluated on 100 open-QA evidence-seeking tasks with PubMed retrieval and DeepSeek models, ASV analyzed 1,100 steps and 2,200 states, revealing belief pivots (mean gold-margin gain: -2.335) missed by traditional aggregate metrics. The authors release the ASV Eval toolkit for diagnostic step-level evaluation.
agent evaluationstate-transition measurementllm evaluatorsbelief pivotslog-probability scoring
LLM-as-a-Tutor: Policy-Aware Prompt Adaptation for Non-Verifiable RL
The paper introduces LLM-as-a-Tutor, a framework addressing prompt-policy misalignment in non-verifiable RL by dynamically adapting training prompts. The method employs a single LLM as both examiner (detecting non-challenging prompts via pairwise rollout comparisons) and generator (appending atomic constraints to increase difficulty), creating a self-calibrating training signal. Evaluated on three complex instruction-following benchmarks, it outperforms policy-unaware baselines and prior policy-adaptive methods that modify rubrics or rewrite prompts.
reinforcement learninginstruction followingprompt adaptationllm judgepolicy-awareness
MechMath Agent Team: LLM Driven Agents for Mathematical Research
The paper introduces MechMath Agent Team (MMAT), a multi-agent system for mathematical research that combines large language models with formal verification. The proposed Harness Architecture decouples responsibilities into Control, Execution, and Augmentation planes, enabling rigorous logical control while maintaining research flexibility. Three specialized agents (Knowledge Base Manager, Natural Language Prover, Formal Language Prover) collaborate in a closed loop to produce certified proofs. In a two-month evaluation on open problems across five mathematical domains, MMAT solved 11 problems, demonstrating end-to-end research cycle capabilities.
multi-agent systemformal verificationharness architecturemathematical reasoningclosed-loop proof
Decentralized Aggregation of LLM Predictions via Wagering Mechanisms
The paper proposes WALLA (Wagering Mechanisms for LLM Aggregation), a decentralized method for aggregating predictions from multiple LLMs while preserving privacy and incentive compatibility. The mechanism uses a leave-one-out baseline in payout functions to ensure dominant-strategy incentive compatibility, advantage-wager alignment, and prediction-agnostic wager optimization. Two variants balance normality and no-arbitrage with bounded worst-case deficits. Experiments on question-answering and forecasting benchmarks demonstrate that WALLA matches centralized aggregation in performance while enabling decentralized learning, uncertainty awareness, and incentive-compatible prediction.
wagering mechanismsllm aggregationincentive compatibilitydecentralized learningadvantage-wager alignment
Auto-AEG: Scalable Data Construction for Open-Vocabulary Audio Event Grounding
Auto-AEG introduces a scalable pipeline for Open-Vocabulary Audio Event Grounding, addressing data scarcity through automatic supervision construction. The method combines programmatically synthesized clips with exact ground-truth intervals for supervised cold-start and multi-model pseudo-labels on real-world audio for reinforcement learning. Training with this pipeline demonstrates performance improvements on the DESED Sound Event Detection benchmark and AEGBench, a newly released difficulty-stratified benchmark. Results indicate that automatically constructed data, paired with interval-aware reward functions, effectively enhances the temporal localization capabilities of Large Audio-Language Models.
open-vocabulary audio event groundinglarge audio-language modelssound event detectionpseudo-labelsreinforcement learning
Nemotron-Labs-3-Puzzle-75B-A9B: Compressing Hybrid MoE LLMs
The paper introduces Nemotron-Labs-3-Puzzle-75B-A9B, a compressed variant of Nemotron-3-Super optimized for interactive deployment, achieving 2x higher server throughput on an 8xB200 node and enabling 8 concurrent 1M-token requests on a single H100 GPU. The compression pipeline combines Iterative Puzzle compression, knowledge distillation, reinforcement learning, quantization, and Multi-Token Prediction, jointly optimizing MoE pruning, active parameter budget, and Mamba pruning. Evaluations show the model retains strong accuracy across reasoning, coding, multilingual, long-context, and agentic benchmarks despite substantial compression.
model compressionmixture-of-expertsknowledge distillationmulti-token predictionquantization
HASSL: Hierarchy-Aware Self-Supervised Learning Framework for Single Cell Microscopy
The paper introduces HASSL, a hierarchy-aware self-supervised learning framework for single-cell microscopy that preserves hierarchical structure in image embeddings. The method combines a segmentation-teacher distillation framework to enhance morphological awareness and a hierarchy-aware contrastive loss based on HDBSCAN to improve decision boundaries across hierarchical levels. Evaluated on 2.3 million single cells from 20 datasets spanning 208 classes, HASSL improves top-K accuracy by 2.8%, top-9 retrieval by 6.3% on deep-hierarchy data, and drug classification F1-score by 7.8% compared to baselines.
self-supervised learninghierarchical clusteringcontrastive losscell morphologyembedding distillation
IRIS: An Intelligent Vision-Language System for Ocular Surface Diseases via Topic Tree and Scene-Driven VQA Generation
The paper introduces IRIS, a vision-language system for ocular surface diseases (OSDs), addressing the data bottleneck in clinical reasoning through IRIS-120K, the largest OSD visual question-answering dataset. The method employs a dual-branch framework: a Topic Finding Tree (TFT) for hierarchical visual feature anchoring and a Scene-driven strategy for role-adaptive clinical dialogue synthesis. IRIS, a compact 4B-parameter VLM, outperforms generalist and specialized medical VLMs up to 34B parameters, demonstrating that structured knowledge injection surpasses parameter scaling for resource-efficient, expert-level AI deployment.
vision-language modelsocular surface diseasestopic finding treeclinical reasoninginstruction-tuning
One Framework for All: Cross-Modal Membership Inference for Generative Models
We introduce a unified membership inference framework applicable across text-to-text, text-to-image, and image-to-text generative models, addressing privacy risks in multimodal settings. Our method leverages the observation that a model's output distribution approximates its training data distribution, modeling both generated outputs and non-member samples in a shared embedding space for likelihood ratio testing. Extensive black-box experiments under partial-knowledge and zero-knowledge threat models demonstrate superior performance compared to state-of-the-art single-modality approaches, achieving effective membership inference against both fine-tuning and pre-training data.
membership inferencegenerative modelslikelihood ratio testingblack-box settingembedding space
Server-side Anti-cheat in FPS games for Aimbot detection using Deep learning and Machine learning
The paper proposes YAACS, a server-side aimbot detection system for FPS games using temporal feature analysis. The method combines time-series data (aim velocity, shot count, target distance) with behavioral patterns (utility usage, movement) in a Stacked LSTM architecture processing 128-tick sequences (Tick Delta Negative=56, Tick Delta Positive=24). The model achieves 88.6% accuracy with 0.97% false positive rate, outperforming a Decision Tree baseline (96.2% accuracy, 2.68% FPR) by 2.76x in FPR reduction, demonstrating sequence modeling's superiority for cheat detection.
aimbot detectionstacked lstmtime-series classificationfps gamesserver-side anti-cheat
Do GUI Agents Believe Their Eyes? Diagnosing State-Belief Reliance on Pixels versus Structure
The study formalizes visual state reliance in multimodal GUI agents, measuring whether state beliefs derive from pixels, structure, or priors. Using 310 real-world web, mobile, and desktop probes with single-channel interventions, it introduces the Perception-Fusion Gap metric to quantify structural deference. Results across five models show textual state beliefs favor structure (Perception-Fusion Gap > 0) while image-only accuracy remains high, with errors propagating to real task failures. White-box ablation traces textual conflicts to copied structural values.
visual state relianceperception-fusion gapmultimodal gui agentssingle-channel interventionsstructural deference
HAS-Bench: Evaluating LLM-Based Human-Agent Systems under Configurable Human Participation
We introduce HAS-Bench, a benchmark for evaluating Human-Agent Systems (HAS) where humans collaborate with LLM-powered agents under configurable participation. The HAS-Framework represents humans and agents as first-class participants with explicit roles, permissions, communication paths, and action authority. HAS-Bench measures task outcomes and process-level collaboration, including clarification quality, feedback utilization, control calibration, safety, initiative, and interaction cost. Experiments across six domains demonstrate that human participation significantly improves task completion and failure recovery, though the benefits depend on the timing, method, and source of human input.
human-agent systemsllm-powered agentsconfigurable participationtask completionfailure recovery
Fixed-Confidence Best-Arm Identification for Causal Mediation Analysis
The paper introduces a fixed-confidence best-arm identification algorithm for causal mediation analysis, targeting the maximization of expected natural direct potential outcome (NDPO) while excluding mediator-transmitted pathways. The method leverages the Track-and-Stop (TaS) framework and employs a cutting-set technique to solve semi-infinite optimization problems, ensuring δ-correctness and asymptotic optimality. Empirical validation on the IPinYou advertising dataset demonstrates the algorithm's sample efficiency and high-probability correctness. Theoretical guarantees confirm its effectiveness in identifying optimal treatments under causal bandit settings.
natural direct potential outcometrack-and-stop frameworkcausal banditsemi-infinite optimizationδ-correctness
HiFA4: Training-Free 4-bit FlashAttention on Ascend HIF4 NPUs for LLM Inference
HiFA4 introduces a 4-bit FlashAttention operator for LLM inference on Ascend NPUs, maintaining FP16 softmax states while executing QK^T and PV as HIF4 Cube GEMMs. The method combines Smooth-QK (per-channel rescaling post-RoPE) and P-Reordering (softmax normalizer accumulation from quantized P_hat), addressing quantization-induced errors. Evaluated on Qwen3-8B, HiFA4 recovers 37.5% of accuracy loss from direct HIF4 quantization, reduces MMLU regressions by 57%, and cuts BF16-inconsistent predictions from 16.3% to 8.2%. Similar improvements are shown for Gemma2-9B, LLaMA3.1-8B, Mistral-7B, and Phi-4B, with projected 35.4% latency reduction via softmax-GEMM fusion.
flashattentionhif4post-training quantizationropegemm
CausalGame: Benchmarking Causal Thinking of LLM Agents in Games
The paper introduces CausalGame, a novel benchmark for evaluating causal thinking in LLM agents through interactive games that simulate scientific discovery challenges. The benchmark includes 14 scenarios incorporating selection bias, measurement error, and hidden confounders, requiring agents to design experiments, collect data, and provide explanations. Evaluation of 30 LLM agents reveals poor performance: the best model achieves only 68.0% survival rate against 78-85% optima, with merely 5-7% of sessions meeting causal-reasoning criteria. CausalGame offers a scalable testbed for assessing causal reasoning in AI Scientist agents.
causal reasoningllm agentsselection biashidden confoundersscientific discovery
Agentic SABRE: An Uncertainty-Aware Neuro-Symbolic Multi-Agent Framework for Adaptive Ransomware Detection
Agentic SABRE introduces a neuro-symbolic, multi-agent framework for adaptive ransomware detection, addressing concept drift and evasion through uncertainty-aware decision-making. The system fuses semantic and behavioral evidence, employing Monte Carlo Dropout for epistemic uncertainty quantification and a decision-layer orchestrator for risk-aware triage. It integrates post-hoc explainability mechanisms, including gradient saliency and counterfactual analysis, for auditability. Evaluations on RDset and RanSMAP show perfect discrimination (AUC=1.0) on semantic datasets, a 4.9% reduction in false escalations, and stable decision boundaries under bounded perturbations.
neuro-symbolicransomware detectionmonte carlo dropoutcounterfactual analysisepistemic uncertainty
Agentic-V2X: Small Language Model Agents for Deadline-Aware V2X Scheduling in 5G/6G Networks
Agentic-V2X proposes a hybrid architecture for deadline-aware V2X scheduling in 5G/6G networks, combining a small language model (LLM) as a periodic policy generator with a lightweight controller for real-time execution. The LLM generates structured policies (priorities, weight bounds, safety constraints) from scenario summaries, validated before deployment via ns-3/ns3-ai. Evaluated against 5 baselines across 126 runs, the adaptive LLM-rApp/xApp achieves competitive performance in critical reliability (improving over proportional fair at high density) but underperforms static expert policies overall, demonstrating safe policy generation without universal dominance.
v2x schedulingsmall language model5g nrdeadline-awarepolicy validation
Risk-Constrained Freshness-Aware Semantic Caching for Open-Web Retrieval-Augmented LLMs
FreshCache introduces a three-tier semantic cache for retrieval-augmented generation (RAG) that models time-varying evidence freshness via risk-constrained temporal inference. The system employs an exponential decay model enhanced by a learned MLP to estimate staleness probability, approving cache hits only when below per-tier error budgets (0.10-0.35). Evaluated on FreshCache-Bench (31,201 queries), it achieves 97% API savings at 0.1% stale error (24-hour window), with true answer-affecting error at 0.034%. The method outperforms SemanticTTL, vCache, and SCALM, reducing stale error by 11.6 points over similarity-only baselines.
semantic cachingretrieval-augmented generationstaleness probabilityexponential decay modelerror budget
Self-Reference in Large Language Models: The Introspection Threshold for Recursive Self-Improvement
The paper establishes a theoretical framework for sustainable recursive self-improvement in Large Language Models (LLMs), proposing an introspection threshold analogous to von Neumann's complexity threshold. Using Kleene's Second Recursion Theorem, the authors demonstrate the existence of introspective programs capable of self-simulation and targeted modification. Empirical analysis reveals current LLMs (e.g., Transformer-based architectures) exhibit only quasi-introspection due to structural limitations: incomplete self-access, feedforward constraints, and computational class barriers preventing fixed-point iteration. The work concludes with architectural proposals to achieve true introspection and discusses associated safety implications.
recursive self-improvementintrospection thresholdkleene's recursion theoremtransformer architecturefixed-point iteration
LBR: Towards Mitigating Length Bias in Large Language Models for Recommendation
The paper proposes LBR (Length Bias Reduction), a framework addressing length bias in LLM-based recommender systems where variable-length item descriptions skew attention and decoding. LBR employs Length-Aware Attention Calibration to adjust attention logits and Effective Information Length Normalization using prefix-tree-derived information measures. Experiments on three Amazon datasets with two LLM recommenders show 16.82% average NDCG@5 improvement while reducing bias, with minimal computational overhead.
length biasllm-based recommendationattention calibrationinformation length normalizationprefix tree
HALO-WA: Hybrid-Attention Latent-Guided Online Reinforcement Learning for World-Action Models
HALO-WA introduces a hybrid-attention latent-guided online RL framework to enhance world-action (WA) models for robotic manipulation, addressing their vulnerability to real-world errors in precision tasks. The method combines latent features and action priors from WA generation with a lightweight actor-critic adapter, using a hybrid-attention structure to maintain temporal consistency while refining action chunks based on visual context and end-stage corrections. Experimental validation on four real-world tasks shows an average success rate improvement from 26.4% to 87.1%, outperforming baselines by 19.2 percentage points with 45–75 minutes of online training per task.
hybrid-attentionlatent-guidedonline reinforcement learningworld-action modelsrobotic manipulation
Shortcut Learning in Legal Judgment Prediction: Empirical Evidence from the UK Employment Tribunal
This paper exposes shortcut learning in Legal Judgment Prediction (LJP) systems through empirical analysis of UK Employment Tribunal decisions. Using 33,158 claims, the authors compare TF-IDF classifiers and LLMs on outcome prediction from claim texts and LLM-extracted summaries. Results show performance is inflated by outcome-revealing linguistic artefacts in post-hoc judicial texts, with a leakage-feature-only model outperforming human experts. However, retraining with masked leakage features preserves predictive power (negligible Macro-F1 drop), demonstrating models can extract genuine signals when shortcuts are removed.
shortcut learninglegal judgment predictionlinguistic artefactsoutcome leakagepost-hoc bias
Signal or Noise? Understanding Generative Models for Real-World Sensor Time Series
The paper introduces SensorGen, a systematic study of generative models for real-world sensor time series across 14 settings, 4 domains, 7 datasets, and 12 modalities. It evaluates five major generative model families, finding that flow-matching models achieve strong overall performance, signal properties like demographic covariates improve longitudinal generation, and synthetic data enhances downstream tasks. Key results show time-frequency modeling benefits high-frequency signals and scaling improves generation quality, establishing best practices for sensor data generation.
generative modelssensor time seriesflow-matchingtime-frequency modelingsynthetic data
Progress- and Reliability-Oriented Group Policy Optimization for Agentic Reinforcement Learning
ProGPO introduces progress- and reliability-oriented group policy optimization for step-level reinforcement learning in agentic tasks, addressing limitations of existing group formation methods. The method maintains exact-prefix action comparisons while augmenting sparse peer signals with transition credits derived from rollout-based state potentials, estimated via semantic expansion and inverse-variance fusion across history depths. Evaluations on ALFWorld and WebShop with Qwen2.5-1.5B-Instruct show ProGPO outperforms baselines at comparable computational cost, with scalability further demonstrated on Qwen2.5-3B-Instruct.
group-based rlstep-level optimizationadvantage estimationstate potentialsinverse-variance fusion
Hierarchical Multi-to-Single-Modal Knowledge Distillation for Disruption Prediction in EAST
Proposes a hierarchical multi-to-single-modal knowledge distillation framework for efficient plasma disruption prediction in EAST tokamak. The method trains a multimodal teacher using visible images and time-series signals with Transformer-based encoders and a prototype-guided spatiotemporal hypergraph module, then distills knowledge to a time-series-only student via graph-structure-level, representation-level, and decision-level distillation. Evaluated on a 640-discharge EAST dataset, the framework preserves multimodal discriminative advantages while reducing inference costs, offering an effective approach for disruption prediction. Source code will be released on GitHub.
knowledge distillationtokamaktransformerspatiotemporal hypergraphdisruption prediction
Biological Motifs for Agentic Control
The paper proposes biological control motifs as a framework for designing reliable autonomous LLM agents, mapping five biological systems (Gene Regulatory Networks) to composable software patterns via polynomial functors and wiring diagrams. It introduces the Agentic Operad, a typed syntax for agent composition with error suppression bounds, and an epistemic topology layer yielding four scaling theorems validated against multi-agent benchmarks. Results include a reference implementation (1,813 tests, 116 examples) demonstrating practical feasibility of biologically inspired agent architectures.
agentic operadgene regulatory networkspolynomial functorsepistemic topologyfeed-forward loops
SoftVTBench: A Safety-Aware Visuo-Tactile Benchmark for Physically Constrained Robotic Manipulation of Deformable Objects
SoftVTBench introduces a safety-aware visuo-tactile benchmark for physically constrained robotic manipulation of deformable objects, addressing the limitation of success-only evaluation in existing benchmarks. Built in Isaac Sim with finite-element-simulated deformable objects, it provides multi-view RGB observations, RGB tactile sensing, proprioception, and language instructions, and defines four task suites based on object type and variation axis. The benchmark separately evaluates Goal Success and Safety Success, the latter requiring no drop and peak deformation below a calibrated threshold. Experiments reveal that success-only evaluation overstates policy performance, and incorporating tactile sensing improves Safety Success (e.g., from 21.4% to 35.6%) while maintaining Goal Success.
visuo-tactilefinite-element simulationproprioceptiondeformable objectssafety-aware
Detecting Hallucinations in Retrieval-Augmented Generation through Grounding-Aware Sensitivity by Perturbation (GASP)
The paper introduces Grounding-Aware Sensitivity by Perturbation (GASP), a span-level detector for identifying hallucinations in retrieval-augmented generation (RAG) by measuring grounding sensitivity—the dependence of answer sentence likelihood on retrieved evidence. GASP computes log-likelihood drops and Jensen-Shannon divergences (JSD) under context perturbations, interpreting results via random nonlinear iterated function systems (RNIFS). Evaluated on RAGTruth, TofuEval, and RAGBench with Qwen2.5 and SmolLM2 models, GASP achieves response-level AUC 0.73 and span-level AUC 0.67, outperforming perplexity, NLI, and self-consistency baselines. It excels in context-constructed outputs but not short-answer QA.
retrieval-augmented generationhallucination detectiongrounding sensitivityjensen-shannon divergencerandom nonlinear iterated function system
Unsupervised Features Mining via Activation Geometry
The paper introduces Mining via Activation Geometry (MAG), an unsupervised framework for extracting reasoning features from LLM activations by prepending natural-language instructions to inputs and measuring representation changes. The method analyzes eight MAG variants, showing extracted features predict model judgments and can be approximated as activation directions for steering decisions. MAG-based dataset selection achieves 94.7% Top-1 and 100% Top-2 accuracy for prompt-injection classifier probes, outperforming ordinary activation similarity.
activation geometryunsupervised feature extractionllm interpretabilityactivation steeringprompt-injection
Agentic IoT: Architectures, Applications, and Challenges Toward the Internet of Agents
The paper proposes Agentic IoT as a cognitive IoT paradigm integrating autonomous AI agents with cyber-physical systems, enabling real-time reasoning, adaptive planning, and distributed coordination beyond traditional task-specific AIoT models. Through systematic review, it positions Agentic IoT relative to AIoT, edge intelligence, and multi-agent systems, presents an architectural framework, and analyzes domain applications while identifying technical challenges. Key contributions include conceptual grounding of this paradigm shift and a taxonomy of research directions for agent-enabled IoT ecosystems across the device-edge-cloud continuum.
agentic iotcyber-physical systemsedge intelligencemulti-agent systemscognitive iot
A Clustering-Based Framework for Identifying Suspicious Trading Patterns in Capital Market
The study proposes an unsupervised clustering-based framework for detecting suspicious trading patterns in capital markets, using K-Means++ on ~1M transactions (2012-2024). The pipeline applies market practice heuristics to classify anomalies, identifying 2.02% as suspicious (51.10% spoofing, 0.10% pump-and-dump, 0.55% insider trading, 1.43% fake breakouts). Without ground truth, validation relies on a Silhouette Score of 0.561, demonstrating separable clusters for fraud detection.
k-means++market manipulationsilhouette scoreunsupervised learninganomaly detection
CritiqueDriveVLM: From Verifier-Guided Reinforcement Learning to Latent Thought Distillation for Autonomous Driving
CritiqueDriveVLM proposes a three-stage framework for vision-language models in autonomous driving, addressing reliability-efficiency trade-offs. The method combines critique-driven multi-turn reinforcement learning with a multi-dimensional verifier to create a robust System-2 Teacher, followed by latent thought distillation to produce a fast System-1 Student. Experiments on DriveLMM-01 show the Teacher improves MCQ accuracy from 55.54% to 76.54%, while the distilled Student reduces generation length to 28 tokens (88% latency reduction) with competitive reasoning depth.
vision-language modelsreinforcement learninglatent thought distillationautonomous drivingsystem-2 teacher
Piercing Gilbreath's Conjecture: From Deep Number Theory Insights to Fintech and Cybersecurity
The paper introduces a novel methodology to address Gilbreath's conjecture, an unsolved problem in number theory since 1878. The approach leverages sieving techniques, including reverse sieving, and presents new theoretical results with proofs. Applications extend to fintech and cybersecurity, particularly in randomness testing, fraud detection, synthetic data generation, and chaos quantification in time series. The work explores innovative concepts such as magic primes, forbidden prime constellations, cellular automata, and sequence equivalence classes, offering a structured pathway toward solving the conjecture.
gilbreath's conjecturereverse sievingmagic primescellular automatasequence equivalence
SeeMe: Mitigating Hallucinations in Large Vision-Language Models through Effective Visual Token Engineering
SeeMe introduces a training-free framework to mitigate hallucinations in Large Vision-Language Models (LVLMs) by restructuring visual tokens through a three-stage token engineering process. The method suppresses irrelevant or noisy visual tokens while preserving informative evidence, addressing a critical source of hallucinations overlooked by existing decoding-stage interventions. Experiments on MME, POPE, and AMBER benchmarks across four LVLMs demonstrate consistent reductions in hallucinations and improved output consistency.
large vision-language modelshallucination mitigationvisual token engineeringtraining-free frameworkoutput consistency
Language models guide symbolic equation discovery by controlling search
The study introduces LLM-PySR, a hybrid approach where language models control symbolic regression search by specifying variables, operators, and search depth, while deterministic metrics evaluate candidate equations. This method contrasts with end-to-end LM generation and pure symbolic regression, demonstrating superior accuracy-complexity tradeoffs across 74 AI-Feynman equations and seven complex recovery tasks. Empirical validation on battery data revealed a compact piecewise-linear relationship between voltage-curve displacement and cycle life, suggesting LMs are more effective as search controllers than direct equation generators.
symbolic regressionlanguage modelsequation discoverysearch controlai-feynman
Information-Geometric Superposed Vowel Evaluation: Part 1. Moraic Syllabary (Japanese)
The paper introduces an information-geometric method for distinguishing AI-synthesized speech from natural speech by analyzing vowel spectral distributions. Using Japanese's five-vowel system, the method normalizes speech spectra as probability density functions, computes Wasserstein distances between vowel spectra, and applies persistent homology for topological clustering. Results show synthetic speech exhibits shorter inter-vowel Wasserstein distances and distinct clustering patterns compared to natural speech's more diverse spectral distributions.
wasserstein metricpersistent homologyvowel spectrasynthetic speech detectioninformation geometry
Mask-based Predictive Representations for Reinforcement Learning
The paper proposes a mask-based predictive representation method for sample-efficient reinforcement learning (RL) with high-dimensional visual inputs. Inspired by NLP and CV techniques, the approach uses a self-supervised auxiliary task where transformers predict masked sequences in latent space, avoiding pixel-level reconstruction. This method learns compressed representations that improve RL sample efficiency when fed into standard models. Experiments demonstrate superior performance over state-of-the-art sample-efficient RL methods on multiple continuous and discrete control benchmarks.
mask predictionsample-efficient rlself-supervised learninglatent space reconstructionvision-based rl
HCSU: A Dataset and Benchmark for Fine-Grained Historical Calligraphy Style Understanding
The HCSU dataset addresses limitations in fine-grained historical calligraphy style understanding by introducing 39,307 character images from 49 calligraphers across 10 dynasties, systematically separating ink manuscripts (Tie) from stone rubbings (Bei) to resolve modal mixture. It provides hierarchical aesthetic descriptions for two evaluation protocols: style discrimination and aesthetic reasoning. Evaluations reveal that state-of-the-art Large Vision-Language Models exhibit sensitivity to script-level and textual cues, struggling to ground aesthetic judgments in brushwork evidence. HCSU exposes fundamental gaps in multimodal architectures, aiming to advance expert-level visual reasoning for cultural heritage preservation.
calligraphy stylemodal mixtureaesthetic reasoninglarge vision-language modelsbrushwork evidence
!Imperio, smolVLA: The Implications of Data Poisoning on Open Source Robotics
The study demonstrates the feasibility of trigger-word data poisoning attacks on vision-language-action models in open-source robotics, showing that minimal poisoned samples can induce complete denial-of-service. Using smolVLA on a LeRobot pick-and-place task, the authors evaluate three poison ratios (1/320 to 3/320 episodes), finding that just three poisoned samples reduce success rates to 0.0±0.0% under trigger conditions while maintaining 50% clean-prompt performance. Attacks generalize across trigger positions despite training exclusively on front-positioned triggers, proving the threat's practicality and stealth.
data poisoningvision-language-actiondenial-of-servicetrigger-wordopen-source robotics
CSB: A Counting and Sampling tool for Bit-vectors
The study introduces CSB, an efficient tool for exact/approximate model counting and uniform/almost-uniform sampling over bit-vector theories in SMT. CSB employs bit-blasting to convert bit-vector formulas into CNF, then leverages existing CNF model counters and samplers. Experiments show significant performance gains over prior methods, supporting projected/non-projected counting and uniform-like sampling.
satisfiability modulo theorybit-vectormodel countingbit-blastingcnf sampler
Conflict-Based Lazy Search for Fast Multi-Manipulator Planning
The paper introduces Conflict-Based Lazy Search (CBLS), a novel algorithm for real-time multi-manipulator planning in cluttered environments. CBLS extends Conflict-Based Search (CBS) by incorporating precomputation and lazy search techniques: (1) a sparsity-controlled, lazily evaluated graph is precomputed for single-manipulator planning, and (2) Lazy Edge-based A* (LEA*) reduces edge evaluations—the computational bottleneck—via lazy search and edge queues. Theoretical analysis shows LEA* is vertex-optimal with improved edge efficiency over A*. Empirical evaluations demonstrate CBLS outperforms both CBS and RRT-Connect in multi-manipulator scenarios.
multi-manipulator planningconflict-based searchlazy searchedge evaluationpathfinding
SOV-CAD: Stepwise Orthographic Views Guided CAD Modeling Sequence Reconstruction
SOV-CAD introduces a stepwise orthographic view supervision framework for reconstructing CAD modeling sequences from images, addressing the iterative nature of human design workflows. The method formulates CAD reconstruction as a sequential decision-making task, employing offline reinforcement learning with a Decision Transformer architecture that incorporates continuous visual feedback and geometric alignment rewards. Experiments demonstrate superior performance over state-of-the-art methods in CAD sequence reconstruction, with notable data efficiency.
cad modelingorthographic viewsdecision transformeroffline reinforcement learninggeometric alignment
Parametric Memory Decoding for Zero-Shot Routing in LoRA-Based External Parametric Memory
The paper introduces Parametric Memory Decoding (PMD), a framework for zero-shot routing in LoRA-based External Parametric Memory (EPM) without additional training overhead. PMD reframes routing as decoding activations over external memory, enabling systematic improvement through unified routing signals. The authors instantiate PMDRouter, which scores LoRAs by response magnitude from a single base-model prefill. Evaluated on PMD-Bench (covering document-level, domain-level, and task-skill settings), PMDRouter achieves superior zero-shot routing performance, demonstrating the feasibility of training-free EPM access.
parametric memoryzero-shot routinglora-based epmactivation decodingpmd-bench
DynaVieW: Schema-Guided World Modeling for Understanding Hierarchical Visual Dynamics
DynaVieW introduces a schema-guided world model for hierarchical visual dynamics understanding in multimodal LLMs, addressing challenges in temporal scene evolution. The method employs interleaved state-transition sequences, a hierarchical schema, and a mixture-of-experts architecture with cross-expert selective attention and schema token re-weighted loss. Results demonstrate improved performance in visual narrative creation and world simulation, with enhanced consistency, controllability, and instruction-following.
visual dynamicsschema-guidedstate-transition sequencesmixture-of-expertshierarchical schema
Forethought: Verifiable Reasoning from Neurosymbolic Primitive Programming
Forethought introduces a neurosymbolic reasoning system that treats reasoning as explicit, verifiable programs composed from symbolic and neural primitives via a domain-specific language. This approach contrasts with test-time scaling methods by producing inspectable and modifiable reasoning programs prior to deployment. Evaluated across five benchmarks, Forethought improves base-model accuracy by ~30% relative, outperforming vanilla prompting, reinforcement learning scaffolds, and prompt-evolution methods. Notably, a non-reasoning model augmented with Forethought matches dedicated reasoning models' capabilities while requiring ~1000x less post-training investment, remaining model-agnostic and auditable.
neurosymbolic reasoningdomain-specific languageverifiable programstest-time scalingtool-calling execution
PLACEMEM: Toward a Compute-Aware Memory Plane for Lifelong Agents
PLACEMEM introduces a compute-aware memory plane for lifelong agents, addressing persistent memory needs beyond context windows and retrieval. The system employs versioned memory capsules that unify semantics, provenance, validity, and reusable runtime state under correction-aware identities. A vLLM-first prototype demonstrates persistent capsule state, concurrency-safe invalidation, OpenAI-compatible routing, and benchmarks measuring latency, reuse, and post-correction behavior. The work provides both an executable artifact showcasing control-plane behavior and a roadmap for replay-aware serving integration in future lifelong-agent systems.
versioned capsulescorrection-awarekv-aware routingcascading invalidationreplay-aware
Submitted and Diagnostic Analysis of Full-Text Temporal Retrieval for LongEval-Sci
This paper presents a diagnostic analysis of temporal retrieval methods for LongEval-Sci 2026, focusing on effectiveness under corpus evolution. The authors evaluate PyTerrier BM25 and Qwen3 dense baselines alongside temporal full-text retrieval, citation retrieval, RM3 query expansion, cross-encoder reranking, and reciprocal rank fusion (RRF). Results show temporalized full-text BM25 achieves the best ARP across three snapshots (0.285, 0.267, 0.180 nDCG@10) and reduces snapshot-3 relative change from 0.481 to 0.368. Internal diagnostics reveal full-text BM25 as the strongest single retriever (DCTR nDCG@10 = 0.3302, MAP = 0.2853), while RRF provides optimal deep recall (Recall@1000 = 0.9667). The study concludes temporal integration enhances longitudinal effectiveness on full-text foundations.
temporal retrievalfull-text bm25reciprocal rank fusionlongeval-sciquery expansion
FedSPM: Routing-Enabled Federated Learning under Dual Heterogeneity via Semiparametric Mixture
(No summary returned.)
TabPack: Efficient Hyperparameter Ensembles for Tabular Deep Learning
TabPack introduces an efficient MLP ensemble method for tabular data that reduces hyperparameter tuning overhead while maintaining competitive performance. The approach trains multiple MLPs in parallel with diverse hyperparameters sampled from specified ranges, dynamically selecting ensemble members during training. Evaluations on medium-to-large public datasets demonstrate that TabPack's default configuration matches extensively tuned baselines, achieving comparable results with significantly reduced computational resources. Notably, TabPack completes training faster on consumer hardware than GPU-based tuning of existing methods.
mlp ensembletabular datahyperparameter samplingparallel trainingdynamic selection
CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon Agents
CompactionRL introduces a reinforcement learning strategy for long-horizon agentic LLMs, addressing context window limitations through joint optimization of task execution and summary generation. The method employs token-level loss normalization and cross-trajectory generalized advantage estimation to enable learning from compacted trajectories. Evaluated on agentic coding tasks, CompactionRL significantly improves performance: GLM-4.5-Air (106B-A30B) achieves Pass@1 scores of 66.8% on SWE-bench Verified (+7.0 points) and 24.5% on Terminal-Bench 2.0 (+3.1 points), while GLM-4.7-Flash (30B-A3B) improves by 5.5 and 6.8 points, respectively. The approach is integrated into the RL pipeline for training GLM-5.2 (750B-A40B).
context compactiongeneralized advantage estimationtoken-level loss normalizationlong-horizon agentspass@1
Fitted Occupancy-Ratio Evaluation without Bellman Completeness
The paper proposes Fitted Occupancy-Ratio Evaluation (FORE), a novel method for offline reinforcement learning that estimates occupancy ratios without requiring Bellman completeness. FORE employs a fixed-point approach using an adjoint Bellman recursion, projecting the Bellman image onto a log-ratio class via KL divergence minimization. The method only requires realizability of the discounted occupancy ratio, unlike prior work which needed additional conditions like value-function realizability. Theoretical analysis shows population-level contraction in relative entropy and finite-sample regret bounds, enabling direct value estimation through reward reweighting and doubly robust estimation. Results demonstrate convergence in KL divergence up to approximation error and statistical complexity.
offline reinforcement learningoccupancy ratiosadjoint bellman operatorkl divergencedoubly robust estimation
Faithfulness to Refusal: A Causal Audit of Neuron Selectors
The study conducts causal audits of neuron-row attribution methods in language models, evaluating their effectiveness for model pruning and safety editing. Using one-shot neuron-row zeroing interventions across five LLMs, it compares attribution-based selectors against activation and magnitude baselines. Results show attribution methods outperform baselines in identifying dispensable rows, enable targeted refusal of harmful content while preserving benign responses, and reveal redundancy in refusal mechanisms across disjoint neuron sets. The audit exposes limitations of rank-stability metrics for selector validation.
neuron attributioncausal auditmodel editingrefusal mechanismrank-stability
How Far is Too Far? Defining the Distance Threshold for Verification Siamese Networks
The authors propose an unsupervised method for determining the verification threshold in Siamese networks by modeling distance distributions as bimodal functions and identifying the inter-mode minimum. This eliminates the need for labeled data typically required by threshold-setting approaches like Equal Error Rate. Evaluated on MNIST, CIFAR-10, LFW, and PKLot datasets, the method achieves 94% average verification accuracy, matching supervised benchmarks while enabling deployment-environment adaptation.
siamese networksverification thresholdunsupervised learningbimodal distributiondistance metric
How Much is Left? LLMs Linearly Encode Their Remaining Output Length
The study demonstrates that large language models (LLMs) linearly encode their remaining output length in hidden states, detectable via minimal-capacity linear probes. Using frozen hidden states from three 7-8B parameter models across seven datasets, the authors show that total response length is linearly decodable from the prompt's last hidden state, probe directions transfer to synthetic completions, and per-position estimates shift during retractions. These findings suggest LLMs maintain a plan-like internal representation of output length, distinct from exact-counting impossibility results.
linear probeshidden statesoutput lengthtransfer learningretractions
Quantum Spectral Anomaly Detection
The authors propose Quantum Spectral Anomaly Detection (QSPADE), a method for computing PCA-like anomaly scores directly from the spectrum of a quantum dataset's average state, avoiding costly eigenvector recovery or Gram matrix construction. QSPADE replaces hard PCA rank selection with a temperature-controlled spectral threshold, enabling continuous score variation and reduced sensitivity to noise. The zero-temperature limit recovers the hard-projector PCA score. Numerical simulations demonstrate QSPADE's equivalence to kernel-PCA on encoded classical data and its effectiveness in detecting phase transitions in quantum-native systems without predefined order parameters.
quantum anomaly detectionspectral thresholdkernel-pcaquantum-native systemsphase transition detection
Biologically Informed Deep Neural Networks for Multi-Omic Integration, Pathway Activity Inference and Risk Stratification in Cancer
The authors propose Pathway Activity Autoencoders, a biologically informed deep learning framework for multi-omic integration in cancer research that balances interpretability and representational capacity. The method incorporates pathway knowledge as architectural constraints while maintaining nonlinear modeling capabilities, evaluated on breast cancer survival prediction and subtype classification. Results demonstrate improved performance through multi-omic integration, with gene/protein/miRNA layers contributing most, and provide guidelines for regularization trade-offs between robustness and predictive accuracy.
multi-omic integrationpathway activity autoencodersbiologically informed neural networkssurvival predictionsubtype classification
Learning Only What Valid Adapters Can Express: Subspace-Constrained Adaptation Against Fine-Tuning Poisoning
The paper introduces subspace-constrained adaptation as a defense against fine-tuning poisoning, leveraging a trusted pool of task adapters to restrict parameter updates. Using flan-t5-large and 196 public LoRA adapters, the authors demonstrate that adapter functionality resides in a low-dimensional shared subspace, with 30-38% weight norm redundancy. Gradient adaptation constrained to 128 subspace coordinates matches full LoRA fine-tuning on clean data while maintaining 62-96% exact match under label inversion, compared to LoRA's 3-26%. The method separates clean from corrupted data by a 120x loss margin and blocks adaptive backdoor attacks with 8% success versus LoRA's 100%. Performance depends on the pool's coverage of target behaviors.
subspace-constrained adaptationfine-tuning poisoninglora adapterslabel inversionadaptive backdoor attack
Routing Anonymity and Identifiability of Noisy Quantum Hardware
This work introduces the first formal framework for backend identifiability in quantum cloud computing, defining routing anonymity as a security notion. Using a backend-identifiability game and hypothesis-testing analysis, the authors prove that routing anonymity decays exponentially at the Chernoff rate under passive i.i.d. access, while establishing a fundamental utility-anonymity trade-off. Theoretical results show identifiability as an intermediate-depth phenomenon, validated experimentally on Amazon Braket with 87-100% classification accuracy across superconducting and ion-trap backends. The findings demonstrate that backend-specific fingerprints persist despite post-processing, necessitating explicit anonymity considerations in quantum cloud services.
quantum cloud computingbackend identifiabilityrouting anonymitychernoff ratepauli-transfer-matrix
Advances in Neural Controlled Differential Equations
This work advances Neural Controlled Differential Equations (NCDEs) through three contributions: Log-NCDEs employ the Log-ODE method for efficient approximation during training, Linear NCDEs replace non-linear vector fields with linear ones for closed-form solutions and parallel computation, and Structured Linear NCDEs enhance efficiency via structured linear vector fields. These methods collectively reduce training step time by up to 1000× while maintaining state-of-the-art performance on time series benchmarks, addressing computational and scalability limitations of traditional NCDEs.
neural controlled differential equationslog-ode methodlinear vector fieldsparallel-in-time computationtime series benchmarks
Untrusted Content Masking for Web Agents with Security Guarantees
The paper introduces Untrusted Content Masking (UCM), a method enabling web agents to operate securely despite prompt injection attacks by restoring trust boundaries in web environments. UCM leverages DOM structure to distinguish trusted from untrusted regions without content inspection, redacting untrusted areas before agent processing and routing interactions through a sandboxed interface with privilege separation. This approach maintains agent functionality while isolating adversarial content, providing security guarantees previously unattainable in web agent frameworks.
prompt injectiondocument object modelprivilege separationweb agentssecurity guarantees
Target-Guided Selective Reweighting for Physics-Informed Neural Network Inverse Problems: A Transfer Learning Approach
The paper proposes Target-Guided Selective Reweighting PINN (TGSR-PINN), a transfer learning method for physics-informed neural networks (PINNs) addressing inverse problems with domain mismatch. TGSR-PINN transfers only weights and biases from source tasks, computes neuron target scores via Taylor sensitivity and pre-activation variance, and applies selective soft decay to low-scoring neurons using Gaussian mixture model-based weak-adaptation signals. Experiments on 2D advection-diffusion and cross-PDE-family (Allen-Cahn to Burgers) tasks show improved parameter recovery while maintaining field accuracy, with supplementary validation on reaction-diffusion under 5% noise. Ablations confirm the importance of scoring, weak-adaptation, layer protection, and soft decay mechanisms.
physics-informed neural networksinverse problemstransfer learningtaylor sensitivitygaussian mixture model
Is the Geometry Doing the Work? An Operating-Point Audit of Hierarchy in Hyperbolic Vision-Language Models
The study conducts an operating-point audit of hyperbolic geometry utilization in three vision-language models (MERU, HyCoCLIP, PHyCLIP) by developing diagnostic tests for curvature activity, cone machinery, and hierarchy evaluations. Results reveal three failure modes: (1) curvature remains near-Euclidean (√cρ ≤ 1) despite parameter adjustments, (2) cone and traversal mechanisms are inactive or misaligned, and (3) hierarchical evaluations are confounded by angular distance or supervision signals. Mechanistic analysis identifies a low-curvature shortcut in entailment objectives, explaining the models' failure to leverage hyperbolic geometry as intended. The audit culminates in a five-number geometry report for future hierarchy claims.
hyperbolic geometryvision-language modelsoperating-point auditentailment conescurvature parameter
SalAngaBhava: A Sinhala Market Dataset for Aspect-based Sentiment Analysis
We introduce SalAngaBhava, a manually annotated Sinhala dataset for Aspect-based Sentiment Analysis (ABSA), addressing the lack of resources for low-resource languages. The dataset comprises Sinhala product reviews collected from domain-relevant sources, labeled with aspect terms and associated sentiments (positive, negative, neutral) following consistent annotation guidelines. It includes sentences and aspect-sentiment pairs across multiple domains, ensuring a well-structured and balanced corpus for ABSA research. Analysis confirms its suitability as a benchmark for Sinhala NLP and low-resource sentiment analysis tasks, facilitating further studies in these areas.
aspect-based sentiment analysislow-resource languagessinhalaannotation guidelinesproduct reviews
GeoFlow: Geo-Aware Modeling of Inter-Area Relationships in Origin-Destination Flow Prediction and Generation
GeoFlow introduces a geo-aware framework for origin-destination (OD) flow prediction and generation, addressing limitations in graph-based methods by incorporating geospatial attributes. The method enhances area representations with relative positions, k-hop and geodesic distances, employs a geometric-intrinsic fusion encoder combining graph attention with coordinate-aware encoders, and utilizes an axial-global attention decoder for OD-specific dependencies. For generation, GeoFlow integrates flow matching models to improve sample authenticity and diversity. Empirical results demonstrate superior predictive accuracy and enhanced generative fidelity and diversity, with ablation studies validating component contributions.
origin-destination flowgeospatial attributesgraph attentionflow matchingaxial-global attention
FUSE: FK-Steered Multi-Modal Flow Matching for Efficient Simulation-Based Posterior Estimation
FUSE introduces a dual-track architecture for simulation-based inference (SBI) that preserves multimodal input features while enabling dynamic interaction, addressing limitations in existing methods. The method combines Feynman-Kac-steered sampling with intermediate observation likelihoods to guide generative trajectories, improving sample quality. Evaluations on SBI benchmarks show FUSE outperforms baselines, matching ground-truth MCMC posteriors, and successfully resolves parameter degeneracies in exoplanet orbital estimation tasks.
simulation-based inferencemultimodal modelingfeynman-kacflow matchingposterior estimation
Video-based detection of cessation of breathing in pre-term infants using machine learning
The study demonstrates that video-based monitoring can detect cessation of breathing events (COBE) in pre-term infants, complementing conventional physiological signals. Respiratory motion was extracted from torso regions in video recordings of 30 infants, with camera-derived signals processed using ResNet architectures. Camera-only models achieved 76.9% balanced accuracy, while hybrid models combining video with impedance pneumography (IP) reached 90.6%, showing video provides additional respiratory information. This supports non-contact video as a viable complementary modality for robust neonatal respiratory monitoring in NICUs.
apnoea detectionnon-contact monitoringresidual networksimpedance pneumographyneonatal intensive care
msPCA: An R Package for Sparse PCA with Multiple Components
The msPCA R package implements sparse principal component analysis (PCA) with multiple components via an alternating maximization algorithm. It generates sparse loading vectors that maximize explained variance while enforcing non-redundancy through either orthogonal loadings or uncorrelated principal components. Benchmarks demonstrate scalability to thousands of features, with competitive runtimes, controlled feasibility violations, and high variance explained.
sparse pcaalternating maximizationloading vectorsvariance explainednon-redundancy
Probing Geospatial SSL Representations with Environmental Signals
The study evaluates self-supervised learning (SSL) representations of satellite imagery by probing their statistical associations with environmental variables, using co-located ERA5 reanalysis data. It employs DINO, MAE, and MoCo models trained under identical conditions, complemented by intrinsic representation metrics to analyze geometry and downstream performance. Results show that representation-level metrics distinguish models with similar benchmark performance, and linear accessibility of environmental signals correlates with task performance in the PANGAEA benchmark. The authors release ERA5 annotations co-located with the SSL4EO dataset to support future geospatial foundation model evaluation.
self-supervised learningera5 reanalysisrepresentation geometrylinear accessibilitygeospatial foundation models
FlatManifold: Robust Continual Learning under Severe Label Noise and Domain Shifts via Intrinsic Manifold Flattening
FlatManifold introduces a robust continual learning framework addressing severe label noise and domain shifts via intrinsic manifold flattening. The method employs a Nyström-based manifold flattening map in an RKHS, orthogonalizing feature distributions with ridge regularization to mitigate label noise effects. A continual topology brake term leveraging past experience covariance prevents catastrophic forgetting. Evaluations on multi-session robotics datasets with 40% symmetric label noise and extreme domain shifts show FlatManifold outperforms baselines, demonstrating resilience to gradient corruption and high generalization capability.
continual learninglabel noisedomain shiftsmanifold flatteningrkhs
Latent Programming Horizons in Coding Agents
The study reveals that language models in coding agents linearly encode program properties in their residual streams, with logistic-regression probes achieving up to 0.83 AUC in predicting code correctness across two models and benchmarks. Surprisingly, these representations anticipate future edits, with probes predicting outcomes up to 25 steps ahead, termed the 'latent programming horizon'. Probes also demonstrate cross-benchmark transferability, suggesting mechanistic interpretability as a promising research direction.
residual streamslogistic-regression probelatent programming horizonmechanistic interpretabilitycoding agents
SMART: A Machine Learning and Monte Carlo Framework for Rapid Analysis of Stochastic Transistor Aging and Process Variation in Digital Circuits
The paper introduces SMART, a machine learning and Monte Carlo framework for rapid analysis of stochastic transistor aging and process variation in digital circuits. The method integrates Random Forest regression with Bayesian Optimization for hyperparameter tuning to predict gate delay distributions directly, avoiding costly atomic model parameter extractions. Evaluated on ISCAS85 benchmarks, SMART reduces analysis time by 94.54% while maintaining 1.63% average accuracy error compared to state-of-the-art methods.
bias temperature instabilityprocess variationrandom forest regressionbayesian optimizationmonte carlo simulation
Platonic Projection Structures: Operator-Induced Observability in Representation Learning
The paper introduces Platonic Projection Structures (PPS), an operator-theoretic framework for analyzing observability in representation learning. PPS models observation via a self-adjoint positive semidefinite operator Π acting on a latent space H, defining observable outputs as O(v)=⟨v,Πv⟩. The framework characterizes observability through quotient geometry H/ker(Π), revealing equivalence classes of indistinguishable latent states. It unifies quantum measurement and linear observation models, showing structural parallels while differing in operator algebra. Empirical results demonstrate kernel-invariant observability, attribution gaps, and rank-controlled geometry. PPS also exposes inherent limits in output-based interpretability due to inaccessible components in ker(Π).
platonic projection structuresobservabilityrepresentation learningquotient geometryoperator-theoretic
MeGA-MP: Metric Graph Advection Message Passing -- A Physics-Informed Message Passing Operator for Advection-Dominated Metric Graphs
We introduce MeGA-MP, a physics-informed message passing operator for advection-dominated metric graphs that encodes linear advection as an inductive bias. The method operates on metric graphs, capturing antisymmetric and long-range dependencies without relying on global Euclidean space. In purely advective settings, MeGA-MP recovers exact dynamics up to a theoretical discretization error without training. When combined with MLPs, it extends to advection-reaction dynamics in water distribution systems, outperforming baselines and demonstrating zero-shot generalization across graph topologies.
metric graphsmessage passingadvectioninductive biaszero-shot generalization
Physiological Noise Augmentation Improves Non-Invasive Brain-to-Speech
Physiological noise augmentation (PNA) improves non-invasive brain-to-speech decoding by enhancing decoder robustness to task-agnostic artifacts. Inspired by noise augmentation in automatic speech recognition, PNA decomposes brain recordings via independent component analysis (ICA), scales and remixes clean data with noise artifacts to generate biophysically realistic training examples. This approach approximates anisotropic regularization, penalizing sensitivity along artifact-dominated directions. On the MegNIST dataset (12k trials), PNA with 10-trial averaging boosts EEGNet decoding accuracy by 4.7 percentage points over baseline training. The results demonstrate that artifact-aware augmentation and trial averaging are complementary for improving non-invasive speech brain-computer interfaces.
physiological noise augmentationindependent component analysisanisotropic regularizationbrain-computer interfacesmegnist
EdgeBench: Unveiling Scaling Laws of Learning from Real-World Environments
The study establishes the first empirical scaling law for real-world environment learning, demonstrating that agent performance follows a log-sigmoid curve with exceptional precision (R²=0.998) across 38,000 interaction hours. Through EdgeBench, a novel benchmark suite of 134 long-horizon tasks spanning scientific discovery to professional workflows, the authors observe that agent learning speed doubles every three months. The work releases 51 tasks and an evaluation framework to advance research in environment-based learning.
scaling lawsenvironment learninglong-horizon tasksagent interactionreal-world benchmarks
Geometric Causal Models
The paper introduces Geometric Causal Models (GCMs), a framework for causal inference from dependent data by exploiting symmetries in the data-generating process. Using group theory to formalize symmetries (e.g., translation in spatial data or node permutations in graphs), the authors establish causal identification via ergodic theory for amenable groups and combine geometric deep learning with scalable Bayesian inference for estimation. The framework generalizes i.i.d. causal models and do-calculus under permutation equivariance, with applications demonstrated in genetic variation analysis using DNA language models and deep functional genomics. Results are validated on semisynthetic data.
geometric causal modelsgroup theoryergodic theorygeometric deep learningdna language models
Physically-Relevant Information Learning in High-Dimensional Time-Derivatives Spaces
The authors propose Time-Derivatives (TiDe) spaces as a framework for learning physically-relevant information in complex dynamical systems by constructing high-dimensional spaces from time-series derivatives. Each dimension corresponds to increasing-order time-derivatives, capturing both structural and dynamic information without requiring prior dimensionality reduction. The method is demonstrated on molecular dynamics simulations and experimental tracking data, showing efficient extraction of interpretable physical phenomena through unsupervised approaches.
time-derivativescomplex dynamical systemshigh-dimensional analysisunsupervised learningmolecular dynamics
Choosing a parallel heterogeneous ensemble method for tabular classification
This study evaluates parallel ensemble methods for tabular classification across 56 OpenML CC18 tasks, deriving best-practice recommendations validated on 28 additional TabArena tasks. Blending and Stacking exhibit task-dependent inconsistencies, while Robust Soft Voting demonstrates superior probabilistic classification, particularly in multiclass scenarios. The recommended ensemble approach significantly outperformed Single Best and matched or exceeded individual ensemble methods in validation. Key findings highlight the complementary nature of Blending and Stacking inconsistencies and the effectiveness of Robust Soft Voting for probabilistic classification.
ensemble methodstabular classificationprobabilistic classificationblendingstacking
Counterfactual Methods for Detecting Unfairness in Anti-Money Laundering Algorithms
This work introduces counterfactual methods for evaluating fairness in Anti-Money Laundering (AML) algorithms, addressing a critical gap in fairness analysis for such systems. Using the IBM AMLSim transaction dataset augmented with country-specific and behavioral features, the authors train diverse models, including decision trees and graph neural networks, and analyze fairness violations through path-specific effect decomposition. Results reveal that models benefiting most from extended features exhibit pronounced fairness violations, highlighting an accuracy-fairness trade-off in AML applications. The study underscores the necessity of systematic fairness evaluation in high-stakes financial domains.
counterfactual analysisanti-money launderingpath-specific effectsgraph neural networksfairness violations
Functional Bilevel Optimization for Predictive Fairness
The paper introduces a method for enforcing mean demographic parity in predictive fairness when sensitive attributes are continuous and high-dimensional, using DPVar (variance of conditional-mean prediction). Two algorithms are proposed: FBO, which employs a closed-form adjoint for exact hypergradients in squared-loss cases, and ITD, which differentiates through unrolled inner steps for broader applicability. Evaluated on synthetic data and a semi-synthetic benchmark from 60 tabular datasets, both methods achieve near-optimal fairness-accuracy trade-offs, outperforming HSIC, adversarial, linear-dependence, and generalized-DP baselines.
demographic paritybilevel optimizationhypergradientfairness-accuracy trade-offconditional-mean prediction
FAST: A Holistic Framework for Optimizing Memory-I/O, Computation, and Sampling in Temporal GNN Training
FAST introduces a holistic framework for optimizing Temporal Graph Neural Network (TGNN) training by jointly addressing memory-I/O, computation, and sampling bottlenecks. The method combines SlimCache for compressed within-batch caching, thread-efficient graph operators for improved GPU cache locality, and topology-aware sampling for CPU cache efficiency. Experiments on large dynamic graphs demonstrate speedups of 2.1x (up to 4.7x) over state-of-the-art systems while maintaining model accuracy.
temporal graph neural networksmemory-i/o optimizationgpu cache localitytemporal neighbor samplingdynamic graphs
Computing Monetary Risk Measures in Linear Time
The paper introduces QuickVaR and QuickDivergence, two linear-time algorithms for computing Value-at-Risk (VaR) and φ-divergence risk measures (including Conditional-Value-at-Risk) on discrete random variables. QuickVaR adapts Quickselect for VaR computation, while QuickDivergence employs polymatroid optimization techniques. Benchmarks demonstrate order-of-magnitude speed improvements for large domains, with implementations provided in the RiskMeasures.jl library.
value-at-riskconditional-value-at-riskφ-divergencepolymatroid optimizationquickselect
KVpop -- Key-Value Cache Compression with Predictive Online Pruning
KVpop introduces a learned key-value cache eviction policy for autoregressive decoding, addressing memory bottlenecks by directly supervising keep-or-drop decisions using a novel future-attention target. The method employs a delayed memory-based scorer that defers scoring to leverage near-future context, achieving efficient compression without dense attention maps. On Qwen3-4B and Qwen3-8B, KVpop maintains 98% and 97% of full-attention performance at 75% and 88% KV cache compression, respectively, outperforming heuristic baselines on AIME and HMMT benchmarks.
kv cacheautoregressive decodingcache compressionfuture-attentioneviction policy
CollabEval: Statistically Efficient Collaborative Model Evaluation via Matrix Completion
The paper introduces CollabEval, a statistically efficient method for collaborative model evaluation that treats evaluation scores as a matrix completion problem. By constructing a low-rank approximation of the $M \times N$ score matrix (where $M$ is models and $N$ is prompts) and using prediction-powered inference, it provides unbiased mean estimates with valid confidence intervals. Empirical results across diverse datasets show CollabEval reduces mean confidence interval size and MSE compared to baselines at equivalent annotation budgets.
matrix completionstatistical efficiencyprediction-powered inferencelow-rank approximationcontrol variates
Uncertainty-aware damage identification in short-span bridges via physics-informed variational autoencoder
The paper proposes a physics-informed Gaussian copula variational autoencoder (PI-GCVAE) for uncertainty-aware damage identification in short-span bridges. The method embeds a differentiable eigenvalue solver into the VAE to enforce structural dynamics constraints and replaces standard latent variable independence with a Gaussian copula to model spatial correlations. Evaluated on synthetic bridge data with stochastic noise, PI-GCVAE achieves 77.2% coverage in recovering the true posterior distribution of stiffness parameters.
physics-informed machine learningvariational autoencoderstructural health monitoringgaussian copuladamage identification
Beyond Modality Fusion: Deep Ensembles for Multimodal Classification
This work demonstrates that deep ensembles of unimodal networks outperform both late-fusion and intermediate-fusion approaches in multimodal classification, particularly under modality imbalance. The authors propose a heuristic for selecting ensemble size per modality, showing that small ensembles should prioritize stronger modalities while larger ensembles benefit from weaker modalities. Empirical validation on synthetic and real-world datasets confirms these findings, with scaling laws estimating asymptotic ensemble performance. The synthetic framework enables controlled study of modality count and predictive strength.
multimodal classificationdeep ensemblesmodality imbalancelate-fusionscaling laws
Knowledge Knows, Verbalization Tells: Disentangling Latent Directions for Mathematical Solvability in LLMs
The study disentangles latent representations of mathematical solvability knowledge and verbalization in LLMs, demonstrating they are encoded as distinct, linearly decodable directions in hidden states. Through probing experiments across multiple models, the authors show fabrication primarily alters verbalization rather than underlying knowledge. Interventions like unsolvability cues and activation steering modify verbalization to reduce fabrication, improving model abstention without affecting core knowledge representations.
latent disentanglementmathematical reasoninglinear probingactivation steeringverbalization
Canonical quantization of neurons
The authors propose a method for quantizing classical neurons by replacing their energy functions with quantum Hamiltonians and applying activation functions via matrix functional calculus, resulting in activation observables measurable on quantum states. They develop hybrid quantum-classical algorithms for training and evaluation, utilizing techniques like the Hadamard test, Hamiltonian simulation, and Schroedingerization for gradient estimation and observable measurement. Numerical experiments show that these quantized neurons outperform classical counterparts in expressive power on learning tasks, establishing canonical quantization as a framework for quantum machine learning primitives.
canonical quantizationquantum hamiltonianactivation observablehadamard testschroedingerization
Non-Convex Sparse Reinforcement Learning via Non-Monotone Inclusions
The paper introduces a novel approach to sparse reinforcement learning (RL) by addressing estimation bias in feature selection through non-convex regularization. It augments least-squares temporal-difference (LSTD) policy evaluation with the projected minimax concave (PMC) penalty, leading to a non-monotone inclusion problem. Theoretical contributions include convergence conditions for the forward-reflected-backward splitting (FRBS) method applied to such problems, ensuring Lyapunov stability and limit point existence under mild conditions, or exact convergence under the weak Minty variational inequality assumption. Empirical results demonstrate that FRBS iterates significantly outperform state-of-the-art feature-selection methods, particularly in scenarios with numerous noisy features.
sparse reinforcement learningnon-monotone inclusionslstd policy evaluationpmc penaltyfrbs method
Data-Driven Soft Labeling Scales DNA Read Classification to Whole-Body Cell-Type Deconvolution
The authors propose Syto, a modular framework for read-level cell-type deconvolution using DNA methylation data, addressing scalability limitations of existing methods through data-driven soft labels. These labels estimate conditional cell-type distributions per read, resolving conflicts from non-discriminative reads and many-to-many pattern-to-cell mappings. Evaluated on a 39-cell-type whole-body atlas, Syto achieves a 2.56× MSE reduction over state-of-the-art methods, with consistent gains on an out-of-distribution 16-tissue dataset. The soft-labeling approach is generalizable to other many-to-many signal-label mapping problems.
cell-type deconvolutiondna methylationsoft labelingread-level classificationepigenetic marks
Qantara: Bridge-Flow Training for Multi-Paradigm JEPA Control
Qantara introduces a multi-paradigm Joint-Embedding Predictive Architecture (JEPA) for control tasks, enabling a single checkpoint to support latent planning, behaviour cloning, and inverse dynamics inference without retraining. The method combines Brownian-bridge interpolants on the state axis with noise-to-data flow matching on the action axis, concentrating training mass on the edges of the noise square for efficient inference. On the LeWM control suite, Qantara achieves a 91.2 success rate (SR) average and sets new state-of-the-art on OGBench-Cube (+7.7 SR over DINO-WM). Behaviour-cloning and video-inverse paths yield 82-83 SR on Push-T and 71-73 SR on Cube, demonstrating versatility across paradigms.
jepabrownian-bridgeflow-matchinglatent-planningbehaviour-cloning
Geometry-Aware Bayesian Quantification via Compositional Data Analysis
The paper introduces a geometry-aware kernel density estimation (KDE) model for multiclass quantification that properly accounts for simplex geometry in posterior probability vectors. The method employs log-ratio representations and Aitchison geometry with shrinkage regularization to handle boundary effects, enabling both point estimation and Bayesian inference of class prevalences. Evaluation across 42 tabular, text, and image datasets demonstrates competitive performance with state-of-the-art quantifiers, particularly improving upon standard Euclidean KDE baselines while maintaining strong Bayesian quantification results.
quantificationkernel density estimationaitchison geometrylabel shiftcompositional data
Train Smarter, Not Longer: Memorization-Guided Data Reuse for Efficient LLM Training
The paper introduces Memorization-guided Data Reuse, a training paradigm for large language models that adaptively schedules data reuse based on memorization dynamics. By analyzing loss retention and downstream evaluation scores, the authors identify a 'Memorization Window' signal indicating that performance improves with repetition beyond current practices (e.g., the 4-epoch limit). Preliminary experiments demonstrate this memorization-driven regime, suggesting potential for optimized training schedules that maximize sample efficiency without overfitting. The work lays groundwork for future memorization-aware schedulers.
memorization windowdata reuseloss retentiontraining paradigmsample efficiency
Sensitivity Sampling with Predictions for k-Means Clustering
The authors propose a method to accelerate sensitivity sampling for k-means clustering on large datasets by leveraging predictions to approximate point importance. They prove that coarser sensitivity approximations suffice for coreset construction, enabling the use of noisy predictors. A natural predictor is introduced for sequential clustering scenarios, where centers from prior datasets inform sensitivity sampling in subsequent datasets. Extensive experiments demonstrate significant improvements in clustering cost versus runtime over uniform sampling and state-of-the-art sensitivity sampling methods on dataset sequences.
k-means clusteringsensitivity samplingcoresetssequential datasetsapproximation
Lightweight ML-Based Automatic Sleep Staging Framework with Constrained CNN and Mamba for Small-Sample EEG Datasets
GamSleepNet introduces a lightweight automatic sleep staging framework for single-channel EEG, addressing challenges of overfitting, low accuracy in difficult stages, and deployment difficulty. The method combines a constrained CNN with Mamba architecture, employing improved Gabor kernels and learnable filters in the FEB module, a novel contrastive loss, and a two-stage training strategy. Experimental results demonstrate state-of-the-art performance on the Sleepedf dataset, achieving 87.86% overall accuracy with only 30.86k parameters, while significantly improving classification accuracy for challenging sleep stages like N1 and REM.
sleep stagingeegmamba architecturegabor kernelscontrastive loss
When Do Foundation Models Pay Off? A Break-Even Analysis of Pretrained Time Series Forecasters
The study presents a systematic break-even analysis for deploying time series foundation models (FMs) versus classical methods, evaluating Chronos, Moirai, and Lag-Llama against Naive, ETS, ARIMA, and XGBoost across 30 datasets at six training fractions (2%-100%). Results show FMs outperform classical methods on 15/30 datasets regardless of data volume, while classical methods surpass zero-shot FMs with minimal data (2%) on 6 datasets. A deployment heuristic recommends zero-shot FMs for short seasonal series (n_train < 700), avoiding LoRA fine-tuning degradation. The proposed two-step decision framework uses dataset length and seasonality to guide infrastructure commitments.
time series forecastingfoundation modelsbreak-even analysiszero-shot learninglora fine-tuning
RL-Ballast: Ship Ballast Water Path Planning and Clog Prediction via Reinforcement Learning
RL-Ballast proposes a graph-based deep reinforcement learning framework for adaptive ship ballast-water path planning and clog prediction under partial observability. The method transforms valve permutations into 54 fluid-transfer routes via graph theory and depth-first search, using frame-stacked tank levels to approximate the POMDP. Results show 41.5 average decision steps (vs. 61.0 Dijkstra baseline), 100% Top-3 diagnostic hit rate, and 66.7% Top-1 accuracy in blockage identification under sensor-sparse conditions.
reinforcement learningballast-water controlpath planningpomdpgraph theory
Unsupervised Detection of Underground Tunnels in Ground-Penetrating Radar Using Depth-Restricted Reconstruction Scoring
The paper introduces an unsupervised pipeline for detecting clandestine tunnels in ground-penetrating radar (GPR) data using depth-restricted reconstruction scoring. A denoising convolutional autoencoder learns anomaly-free subsurface structures, with tunnel detection based on reconstruction error. The key innovation is a depth-restricted top-k anomaly score that pools errors only within physically plausible tunnel depths (1.5-3 m), improving AUC from 0.986 to 0.994 and reducing missed detections from 74 to 17 out of 634 windows. The system achieves AUC 0.994, F1 0.975, and 1.6% false-alarm rate on 1,600 test windows without labeled tunnel data.
ground-penetrating radarunsupervised detectionconvolutional autoencoderreconstruction erroranomaly scoring
Active Learning on Adversarially Corrupted Graphs
The paper introduces an active learning algorithm for identifying adversarially corrupted vertices in graphs, where an adversary adds edges to hide corrupted nodes. The method adapts sum-of-squares techniques to approximate minimum vertex expansion under cardinality constraints, leveraging graph connectivity properties. The algorithm achieves polynomial query complexity dependent on both adversary power and the vertex expansion of the uncorrupted graph, marking the first explicit link between vertex expansion and robust active learning query complexity.
active learningadversarial corruptionvertex expansionsum-of-squaresquery complexity
Enhancing the Forecasting Capability of Multi-Model Blending Algorithms for Extreme Precipitation via Joint Use of Station and Gridded Observations
The study proposes a U-Net based two-stage framework for blending forecasts from six numerical weather prediction (NWP) models to improve extreme precipitation forecasting. The method employs probability classification followed by value reconstruction, incorporating a station-grid joint supervision mechanism using observations from 2411 meteorological stations in China. Results show a 38.4% improvement in Threat Score (TS) for rainstorms (>=50 mm) and operational utility for extreme events (>=100 mm), with station observations enhancing TS by 10.4% and balancing Bias.
extreme precipitationmulti-model blendingu-netthreat scorenumerical weather prediction
Framework for Grouping Local Process Models
The authors propose a framework for grouping Local Process Models (LPMs) to address model explosion and repetition in process mining. Their method groups LPMs by structural or contextual similarity, using process model similarity measures or event log attributes, and selects one representative per group. Evaluations on multiple event logs show reduced repetition and improved coverage compared to top-scoring LPM samples, demonstrating the framework's utility for process understanding.
local process modelsprocess miningmodel similarityevent logspattern discovery
SleepBand: Single-Source Domain Generalization for Sleep Staging via Physiologically Structured Spectral Modeling
SleepBand introduces a single-source domain generalization (SDG) framework for sleep staging that embeds physiologically structured spectral priors via learnable Morlet wavelets and an integration-recalibration pipeline. The method leverages domain-invariant oscillatory features (e.g., slow waves, spindles) while minimizing reliance on dataset-specific artifacts, requiring neither multiple source domains nor domain labels. Evaluated on five public datasets, it achieves state-of-the-art SDG performance and competitive multi-source DG results, with learned filters aligning with canonical neurophysiology. The approach demonstrates that physiologically grounded inductive biases enhance robustness in single-domain settings.
domain generalizationsleep stagingmorlet waveletsspectral modelingneurophysiology
Representing and Detecting Label Ambiguity in IMU-Based Exercise Evaluation
The study proposes a method for representing label ambiguity in IMU-based exercise evaluation by generating label distributions per repetition without requiring multiple raters. Using a neural network trained with a Kullback-Leibler objective (ambiguity approach), the method outperforms one-hot cross-entropy baselines on four IMU datasets, achieving comparable or superior classification accuracy while reliably detecting ambiguous repetitions and relevant classes. Results demonstrate that incorporating label distributions preserves ambiguity information without compromising classification performance.
imulabel ambiguitykullback-leibler divergenceexercise evaluationone-hot encoding
Probably Correct Optimal Stable Matching under Two-Sided Uncertainty
The paper introduces a sequential learning framework for identifying optimal stable matchings in two-sided markets with initially unknown preferences. The authors address two-sided uncertainty and leverage partial preference information through the concept of pervasive stable matching. They propose elimination-based algorithms with structure-aware stopping criteria, providing refined sample-complexity bounds for pure exploration. The approach is extended to regret minimization, achieving bounds independent of the minimum reward gap Δ_min.
stable matchingtwo-sided uncertaintysemi-bandit feedbackpure explorationregret minimization
KinEMbed: Decoding Kinematics from Electromyography via Cross-Modal Contrastive Learning
KinEMbed introduces a cross-modal contrastive learning framework for continuous regression of hand kinematics from surface electromyography (EMG). The method jointly trains dual encoders for EMG features and kinematic targets, preserving geometric structure in the embedding space without requiring kinematic signals during inference. Evaluated on NinaPro DB8 (N=11), KinEMbed outperforms PCA, PLS, autoencoder, and CEBRA baselines, particularly on thumb articulation, achieving 12.3% lower RMSE than the best baseline.
electromyographycontrastive learningkinematics regressioncross-modal embeddingwearable biosignals
Layer-Parallel Inference Reduces Encrypted Nonlinear Depth in Transformers
Structured Newton Layer Parallelism (SNLP) reduces the encrypted nonlinear depth in Transformer inference by minimizing sequential composition of nonlinear blocks, enabling more efficient fully homomorphic encryption (FHE). SNLP replaces layerwise sequential nonlinear depth with solver iterations and linear structured corrections, evaluated using Chebyshev polynomial approximations across 8 models and 4 architecture families. Results show SNLP reduces symbolic bootstraps from 53 to 20 (2.65x) with only +1.2% perplexity degradation on a 0.5B IDN-trained model, while lowering error amplification (1.36x vs. 1.42x). Softmax approximation dominates error, and CKKS arithmetic noise is negligible, indicating SNLP complements FHE-friendly operator design.
structured newton layer parallelismfully homomorphic encryptionchebyshev polynomial approximationssymbolic bootstrapserror amplification
Context-Constrained Transfer Learning for Tabular Foundation Models via Data Distillation
Proposes TL-ANDI, a transfer learning framework for Tabular Foundation Models (TFMs) addressing context-size constraints and distribution shifts. The method constructs compact source contexts via budget-constrained optimal transport, optimizing for target covariate coverage and posterior compatibility, then augments anchor samples with locally distilled labels and residual calibration. Demonstrates improved transfer learning performance by mitigating negative transfer in heterogeneous data settings.
tabular foundation modelsin-context learningoptimal transportnegative transferposterior calibration
Compressed Computation under $L^4$ Loss is likely Computation in Superposition
The study demonstrates that neural networks can perform computation in superposition, evaluating more functions than available neurons, through compressed computation under $L^4$ loss. A single-hidden-layer ReLU network with 50 neurons was trained to compute the ReLU of 100 sparse input features using $L^4$ loss, revealing superposition-based computation. Reverse engineering showed the network assigns sparse binary codewords to features and decodes them via a pseudoinverse. This encoding was validated by constructing equivalent networks from hand-designed codes, recovering most performance with only three scalars.
superpositionrelu networkcompressed computationl4 losssparse codewords
Non-Asymptotic Error Bounds for SMC with Biased Proposals: Application to Conditional Diffusion Sampling
The paper develops non-asymptotic error bounds for Sequential Monte Carlo (SMC) methods with biased mutation kernels, focusing on conditional sampling applications. By extending local Doeblin-type conditions and Lyapunov drift arguments to conditional distributions, the analysis decomposes total error into kernel bias and finite-particle Monte Carlo error. The framework is instantiated for score-based diffusion models, yielding the first joint non-asymptotic control of initialization error, time discretization, score approximation, and finite-particle error in reverse diffusion dynamics.
sequential monte carlonon-asymptotic error boundsfeynman-kac flowscore-based diffusionlyapunov drift
Towards Personalized Differentially Private Learning for Decentralized Local Graphs
PPGNN introduces a personalized differentially private framework for decentralized graph learning, addressing heterogeneous privacy preferences in local graph data. The method combines a Personalized Perturbation Mechanism (PPM) with FlexProp, a weighted calibration strategy, to handle varying privacy budgets while preserving utility. Experiments on six real-world datasets demonstrate improved privacy-utility tradeoffs compared to uniform LDP approaches.
local differential privacydecentralized graphspersonalized perturbationprivacy-utility tradeoffgraph learning
Non-asymptotic Convergence of Stochastic Gradient Descent in Score-based Generative Models
The work establishes non-asymptotic convergence guarantees for Stochastic Gradient Descent (SGD) in Score-based Generative Models (SGMs) under two regimes. For general score parameterizations, it derives a non-convex convergence rate for SGD on weighted denoising score-matching objectives, explicitly accounting for schedule-dependent weighting factors. For overparameterized two-layer ReLU networks, it develops a Neural Tangent Kernel analysis specific to diffusion training with SGD, providing score-approximation error bounds. The analysis quantifies the impact of reweighting factors on score approximation error, offering theoretical justification for practical weighting choices.
score-based generative modelsstochastic gradient descentneural tangent kerneldenoising score-matchingnon-convex optimization
MARLIN: De Novo Molecular Structure Elucidation from Tandem Mass Spectra without a Ground-Truth Formula
MARLIN introduces a de novo method for molecular structure elucidation from tandem mass spectra without requiring a ground-truth molecular formula. The approach combines a self-supervised encoder to predict molecular fingerprints from raw peaks and a block-diffusion language model to generate candidate structures, constrained by a provably safe mass-shell mechanism. Evaluated on NPLIB1, MARLIN outperforms formula-free methods in exact-match accuracy, structural distance, and fingerprint similarity, while incidentally recovering correct molecular formulas as frequently as dedicated predictors.
tandem mass spectrometryde novo structure elucidationmolecular fingerprintblock-diffusion language modelmass-shell constraint
Identifiability of Relational Queries in Multi-View Pretraining
The paper formalizes query identifiability in multi-view data integration, where interface design determines whether a query can be uniquely resolved. It presents CheckCert, a polynomial-time algorithm for deciding identifiability via attribute closure, exact on closure-separable instances. Theoretical results show a 1/2 minimax error floor for non-identifiable queries and Greedy-MinAug, a logarithmic-approximation algorithm for minimal interface augmentation. Experiments on synthetic and real-world datasets (up to 10^3 attributes) confirm CheckCert's exactness, millisecond runtime, and the predicted error floor in ML classifiers.
query identifiabilityattribute closuremulti-view pretrainingfunctional dependenciesminimax error
LP-SFT: Local-Preserving Supervised Fine-Tuning via Multimodal Entropy Structure
LP-SFT introduces a Local-Preserving Supervised Fine-Tuning objective that maintains pretrained models' multimodal entropy structure during adaptation. By analyzing Shannon and Renyi entropies, the method identifies that pretrained models encode rich distributional knowledge beyond supervised tokens. LP-SFT constructs adaptive token supports and applies locally normalized preservation loss to protect relative structure among alternatives, while optimizing supervised tokens via standard cross-entropy. Experiments demonstrate LP-SFT outperforms vanilla SFT and recent enhancements, achieving superior balance between pass@1 accuracy and pass@k performance across mixed-domain and single-domain fine-tuning tasks.
supervised fine-tuningmultimodal entropylocal preservationcross-entropyadaptive support
What You See Is What You Get: Observation-Aligned Supervision for Chart-to-Code Generation
We propose Observation-Aligned supervision, a rewriting framework for chart-to-code generation that replaces latent raw-data targets with quantities constrained by visual observation, addressing the issue of non-identifiable quantities in chart programs. The method rewrites training data targets to align with observable chart elements (box statistics, wedge percentages, bin weights) rather than unobservable raw data. Experiments on ChartMimic and ChartX benchmarks with multiple VLMs demonstrate consistent improvements in observable value recovery, particularly under both-executable evaluation. Results indicate that supervision target alignment, rather than solely increasing data or advancing learning algorithms, is crucial for improving chart-to-code models.
chart-to-code generationobservation-aligned supervisionnon-identifiable quantitiesboth-executable evaluationvisual language models
Hierarchical Scaffolding Enables Human-Like Cognitive Selectivity under Data Scarcity
The paper introduces SCALA (Scaffolded Cognitive Architecture for Learning under limited dAta), a hierarchical learning framework inspired by human cognition that guides models from coarse to fine-grained recognition. SCALA mimics human-like cognitive selectivity by prioritizing task-relevant features while suppressing distractors, leading to accelerated cluster formation and enhanced semantic separability. Empirical results show significant accuracy improvements under data scarcity, robust generalization to unseen classes, and faster novel category acquisition, demonstrating human-level sample efficiency.
hierarchical learningcognitive selectivityrepresentation learningdata scarcitycategory generalization
F-ACVAE: A Federated Adaptive Conditional Variational Auto-Encoder for Privacy-Preserving Intrusion Detection in IoT Networks
The paper proposes F-ACVAE, a federated adaptive conditional variational autoencoder for privacy-preserving intrusion detection in IoT networks. The framework employs selective parameter aggregation (keeping local encoders private while synchronizing global components) and introduces constrained momentum Gaussian aggregation (CMGA) to handle non-IID data and feature shifts. Evaluated on N-BaIoT, F-ACVAE achieves 99% accuracy and macro F1-score while reducing communication overhead by 62% compared to baselines.
federated learningconditional variational autoencodernon-iid dataintrusion detectionselective aggregation
PAST-TIDE: Prototype-Anchored Statement Tuning with Topic-Invariant Normalization for Stance Detection
PAST-TIDE introduces a stance detection system for the StanceNakba Shared Task, employing statement tuning to reformulate stance detection as cloze-style masked language modeling (MLM) with a verbalizer for label mapping. The method combines prototypical contrastive learning using learnable class prototypes and topic-conditional layer normalization for cross-topic robustness in Arabic. Evaluated on the official leaderboard, PAST-TIDE achieves macro-F1 scores of 0.75 (Subtask A) and 0.74 (Subtask B), demonstrating competitive performance with minimal architectural modifications to pre-trained models in low-resource settings.
stance detectionmasked language modelingprototypical contrastive learningtopic-conditional layer normalizationlow-resource settings
A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics
We introduce 3D-PRIMME, a physics-regulated neural framework for learning 3D grain growth dynamics that preserves physical invariants and statistical scaling laws. The model is trained on two consecutive time steps from a $100^3$ grid with 512 grains, learning a scale-independent local evolution rule. It accurately reproduces the linear coarsening law and maintains topological statistics over extended time scales. Notably, 3D-PRIMME generalizes to domains up to $1024^3$ grid points with 550,000 grains without retraining, demonstrating consistent kinetics and grain topology across orders-of-magnitude increases in system size.
grain growthmicrostructure evolutionphysics-regulatedcoarsening lawtopological statistics
GlaKG: A Biomarker-Centric Fundus Knowledge Graph for Explainable Glaucoma Diagnosis and Risk Assessment
GlaKG introduces a biomarker-centric fundus knowledge graph for explainable glaucoma diagnosis and risk assessment, addressing the interpretability limitations of opaque deep-learning models. The framework integrates structural biomarkers, clinical rules, and image features into a unified graph, encoding six entity types, eight relation types, and 11 validated rules. Predictions are accompanied by explicit reasoning chains linking biomarker evidence to clinical rule activations. A post-processing fusion framework combines ResNet50 embeddings with KG reasoning-chain scores. GlaKG achieves F1 = 0.9953 for binary glaucoma classification and 0.930 accuracy for four-class risk stratification on an AI-annotated fundus dataset, with feature-importance analysis showing near-equal contributions from KG-derived and biomarker features (51.1% vs. 48.9%).
knowledge graphbiomarkerglaucoma diagnosisreasoning chainrisk stratification
Decomposition for Bayesian Networks: Local and Parallel Inference
The authors propose a decomposition framework for efficient probabilistic inference in high-dimensional Bayesian networks, addressing the exponential complexity of joint distribution manipulation. Their method constructs directed convex subgraphs and a minimal d-decomposition tree, enabling representation of the joint distribution via lower-dimensional sub-models that support parallel computation. Two parallel algorithms are developed for parameter estimation and inference. Experiments demonstrate significant computational efficiency gains over junction-tree methods while preserving accuracy, particularly for low-dimensional queries.
bayesian networksprobabilistic inferencedecomposition treeparallel computationparameter estimation
Integrating Neural Encoders in Bayesian Generalized Linear Mixed Models for Multimodal Data
The paper introduces a method for integrating neural encoders into Bayesian generalized linear mixed models (GLMMs) to handle high-dimensional multimodal data while maintaining uncertainty quantification. The approach jointly learns modality-specific neural representations with a GLMM objective, then performs variance-corrected stochastic gradient MCMC for GLMM parameters conditional on these representations. This preserves interpretable fixed/random effects for structured covariates and scales to large longitudinal datasets. Simulations show the method recovers posterior means and variances comparable to full-data MCMC benchmarks, with validated interval coverage and predictive calibration. Applications to glaucoma progression and mental health demonstrate multimodal importance assessment without predictive loss.
bayesian glmmneural encodersmultimodal datastochastic gradient mcmcuncertainty quantification
Reliability and Identifiability in Persona-Trained Monte Carlo: Variance Decomposition, Stability Bounds, and the Identifiability of Heterogeneous News Reaction
This paper develops statistical theory for Persona-Trained Monte Carlo (PTMC), a method estimating market-outcome distributions via simulations of limit-order-book interactions among K neural policy bots with personas drawn from a learned heterogeneity distribution P. It decomposes estimator variance into persona-draw (σ_P^2) and within-run (σ_w^2) components, derives variance-optimal compute allocation, and provides stability bounds quantifying error propagation from misestimated P and policy training. The main contribution is an identification theory for heterogeneous news reaction, proving detection via Jensen gap and local identification through odd moments and Hausdorff determinacy, with √n-consistent estimators and a homogeneity test. Separation theorems formalize PTMC's advantages over homogeneous simulators and reduced-form forecasters.
persona-trained monte carlovariance decompositionheterogeneous news reactionjensen gaplimit-order-book
Measuring What Matters: A Unified Evaluation Framework for GNN Explainability
The authors propose a unified evaluation framework for Graph Neural Network (GNN) explainability (G-XAI) that formalizes tabular metrics for assessing topological structure and node features independently, without requiring ground-truth assumptions. Through large-scale benchmarking, they identify Pareto-optimal explainers across metric pairs and tasks, demonstrating that no single method universally outperforms others. The findings are distilled into actionable guidelines for practitioners to evaluate and deploy trustworthy GNN-based pipelines.
graph neural networksexplainabilitybenchmarkingpareto fronttopological structure
Breaking the One-Dimensional Expressibility-Trainability Tradeoff
The work demonstrates that the presumed one-dimensional tradeoff between expressibility and trainability in parameterized quantum circuits (PQCs) is oversimplified. By distinguishing three hierarchical metrics—entangling power (EP), entangling-power deviation (EPD), and gradient variance—the authors show that high Hilbert-space coverage (via EP) can coexist with trainability (via EPD and gradient moments). Analytic and empirical results reveal circuits achieving Haar-like coverage without barren-plateau collapse, enabling a two-dial design rule: maximize EP while preserving EPD. This reframes ansatz design by separating coverage homogenization from gradient-variance collapse.
parameterized quantum circuitsentangling powerbarren plateausgradient varianceansatz design
Minimum Block Width for Universal Approximation by Residual Neural Networks with Inner Width One
The paper establishes exact minimum block width requirements for universal approximation in residual neural networks with inner width one. Using LeakyReLU, ReLU, and ReLU-like activations, the authors derive upper and lower bounds for $L^p$ and uniform approximation on compact domains. Key results show minimum block width is $\max\{d_x,d_y\}$ for $L^p$ approximation, while $\min\{d_x+d_y, \max\{2d_x+1,d_y\}\}$ suffices for uniform approximation. The work also proves networks with block width below $\max\{d_x, d_y\}$ cannot achieve universal approximation regardless of inner width.
residual neural networksuniversal approximationblock widthleakyrelucompact domains
Score Distributions, Not Cells: Evaluating Single-Cell Perturbations Under Class Overlap
The paper introduces the Classifier Discrimination Score (CDS), a novel evaluation metric for single-cell perturbation data that addresses class overlap issues by aggregating classifier probabilities across cell populations rather than scoring individual cells. CDS computes population profiles by averaging per-cell probability vectors, then ranks perturbations based on these profiles, enabling accurate identification even with weak classifiers. Evaluated on Tahoe-100M and the Virtual Cell Challenge, CDS outperforms per-cell accuracy and pseudobulk-based Perturbation Discrimination Score (PDS), particularly with scarce cells. The method requires no retraining and operates in linear time relative to cell count.
single-cell perturbationclassifier discrimination scorepopulation profilepseudobulkclass overlap
MTEB-PT: A Text Embedding Benchmark for Brazilian Portuguese
MTEB-PT introduces the first dedicated benchmark for Brazilian Portuguese text embeddings, comprising 22 native tasks across seven categories, excluding translated corpora. The benchmark evaluates 93 models (23M to 27B parameters, including 73 open-weight and 20 commercial APIs) using statistical analyses such as bootstrap confidence intervals, paired-bootstrap significance, and task-level discrimination adapted from Item Response Theory. Results show clear separation into a dozen model tiers, with an open-source model achieving top-tier performance. Multilingual leaderboard rankings moderately predict Portuguese performance (Spearman rho = 0.75), indicating native benchmarks capture distinct aspects. The authors release tasks, code, and a public leaderboard for practitioner use.
text embeddingsbootstrap confidence intervalsitem response theorymultilingual leaderboardspearman rho
Beyond the Need for Speed: Energy-Aware Code Generation via Simulation-Guided Reinforcement Learning
The paper introduces Green Tea, a method for energy-aware code generation using simulation-guided reinforcement learning, addressing the gap in optimizing for energy efficiency without sacrificing correctness. It constructs a corpus of 3.5M evaluations across 1,474 C++ problems via deterministic architectural simulation, avoiding hardware variance. The approach combines supervised fine-tuning on energy-contrastive pairs with GRPO reinforcement learning, achieving 12.63% CARET improvement and outperforming human-expert references on 58.4% of valid outputs. The work highlights the inadequacy of throughput proxies like IPC, which misrank energy efficiency in 67.8% of cases.
energy-aware code generationsimulation-guided reinforcement learningcorrectness-adjusted reduction in energy totalarchitectural simulationgrpo reinforcement learning
ManifoldFlow: SPD-Relaxed Stiefel Layers with Learnable Singular Spectrum
ManifoldFlow introduces a relaxed Stiefel layer that maintains orthonormal bases while learning bounded positive singular values through weight decomposition W = Q S^{1/2}, where Q is orthogonal and S is positive definite. This enables direct spectral control via eigenvalue clipping of S, addressing the limitation of fixed singular values in standard Stiefel layers. Experiments on sequence, tabular, and image tasks show improvements over fixed-spectrum Stiefel layers, particularly in recurrent language-model projections, though it remains specialized for settings where orthonormal bases are beneficial.
stiefel manifoldsingular spectrumorthonormal basisspectral controlspd relaxation
Causal ASCEND: Scalable Two-tier Causal Discovery on High Dimensional Multi-omics Data
ASCEND introduces a scalable constraint-based framework for genome-scale causal discovery in high-dimensional multi-omics data by leveraging two-tiered hierarchical structure. The method employs a divide-and-conquer strategy with dynamically updated ancestral conditioning sets, reducing conditional independence tests and achieving polynomial-time complexity compared to traditional exponential approaches. Evaluations on simulations and real biological data demonstrate ASCEND's superior accuracy in ancestral relationship recovery, scalability, and computational efficiency over existing gene regulatory network inference methods. The algorithm's directionality resolution capability makes it particularly effective for integrating multi-omic datasets with upstream regulators and downstream responses.
causal discoverygene regulatory networkconditional independencemulti-omicspolynomial-time complexity
Beyond travel mode: urban context shapes active mobility's mental health effects over time
The study quantifies unequal mental health effects of active mobility across urban contexts using causal machine learning and causal deep learning on 264,168 UK adults. Results show individualized effects on anxiety, depression, and common mental disorders vary substantially (e.g., anxiety risk changes ranging from -40.6% to +10.1%), with benefits concentrated in greener, safer, less polluted neighborhoods (81.8% of individuals showing above-average benefits). Urban compact form nonlinearly interacts with neighborhood quality, amplifying benefits only in supportive environments. Genetic moderation was negligible, suggesting context-driven inequalities may widen without targeted interventions.
causal machine learningactive mobilityurban contextmental health inequalitiesnonlinear interactions
Boundary-layer asymptotics for Gaussian-smoothed singular measures
The paper establishes small-noise asymptotics for Gaussian-smoothed probability measures on manifolds with corners, focusing on boundary-layer regimes where observation points approach strata at the same scale as the smoothing parameter. Using conical boundary-layer analysis and rescaling, the authors derive a two-term expansion for the heat-regularized density, with leading terms involving Gaussian mass of linearized cones weighted by density and corner Jacobian. Results include logarithmic asymptotics and uniform expansions for score, log-Hessian, and scale derivatives, revealing how support geometry and curvature influence regularization's differential structure.
boundary-layer asymptoticsgaussian smoothingmanifolds with cornersheat regularizationsingular measures
Constrained Flow Matching via Lagrangian Dual Flows
The paper introduces Lagrangian Dual Flows, a novel method for constrained generation in flow matching that guarantees nonlinear constraint satisfaction without costly optimization subproblems. By flowing a dual co-state alongside generated samples, the approach avoids pseudoinverses or projection steps during denoising. The technique establishes theoretical connections between flow matching and primal-dual optimization methods, offering a simple and effective solution for applications in robotics, planning, and physics.
flow matchinglagrangian dual dynamicsconstrained generationnonlinear constraintsprimal-dual methods
From Interaction to Intent: Inferring User Objectives from Provenance Logs
The paper demonstrates that provenance logs of user interactions can effectively classify analytic intent during visual data exploration. Using fine-grained mouse interaction data from multidimensional projection tasks, the authors identify distinct behavioral signatures corresponding to different objectives (e.g., cluster analysis vs. outlier detection). Their key finding shows that contextually embedded interaction features enable intent classifiers to generalize across datasets and projection methods, establishing provenance logs as a viable bridge between low-level interactions and high-level intent.
provenance logsanalytic intentinteraction patternsmultidimensional projectionintent classification
LeukocyteCount: Automatic Identification and Counting for leukocytes using Deep Learning
This study introduces LeukocyteCount, a hybrid deep learning model for automated leukocyte identification, counting, and classification. The method integrates YOLOv5 for leukocyte detection, MobileNetV2 for feature extraction, and Logistic Regression for classification into four leukocyte types. Trained and validated on the BCCD dataset, the model achieves 98% detection accuracy with YOLOv5 and 99.04% classification accuracy in subsequent stages. The YOLOv5-based RBC detection module attains an F1 score of 99.73%, surpassing baseline performance. These results demonstrate the model's potential to enhance accuracy and efficiency in leukocyte analysis for disease diagnostics.
leukocyteyolov5mobilenetv2logistic regressionbccd dataset
Weakly Guided and Autoregressive Beamformer Parameterization for Generalizable Moving Speaker Extraction in Higher-Order Ambisonics
The authors propose a weakly guided, autoregressive beamforming pipeline for generalizable moving speaker extraction in higher-order ambisonics, requiring only an initial target direction estimate. Their method decouples neural temporal-spectral processing from linear spatial processing, enabling array-agnostic enhancement through a frame-wise causal framework with autoregression. Evaluations on synthetic data demonstrate robust performance under challenging conditions with closely spaced and crossing speakers, while real-world recordings in dynamic office meetings validate generalizability across varying ambisonics orders.
beamforminghigher-order ambisonicsautoregressionspatial processingspeaker extraction
Fields of the Planet: Field Boundary Mapping Beyond 10m
The paper introduces Fields of the Planet (FTP), a 3 m PlanetScope dataset for field boundary mapping that addresses the limitation of 10 m Sentinel-2 pixels in detecting smallholder parcels. FTP pairs with Fields of The World (FTW) using co-registered PlanetScope patch-window targets across 24 countries (133,168 samples) and evaluates performance via panoptic quality (PQ), object F1, size-stratified PQ, and matched-boundary error. Results show that 3 m imagery improves PQ from 21.0 to 35.5, sub-0.5 ha field PQ from 5.8 to 15.7, and reduces boundary error from 18.6 m to 7.4 m under matched training protocols.
field-boundary mappingplanetscopepanoptic qualitysmallholder parcelsmatched-boundary error
Knowledge-Informed Local Causal Discovery of Optimal Adjustment Sets
The authors propose b-LOAD, a knowledge-informed extension of the LOAD algorithm for local causal discovery of optimal adjustment sets. The method integrates prior edge constraints into local structure learning, uses Meek's rules to dynamically expand the discovery frontier, and produces a knowledge-constrained partially directed graph. Theoretical analysis shows the procedure monotonically refines the admissible equivalence class and enlarges identifiable causal queries. Empirical results demonstrate improved causal effect estimation, particularly in data-scarce and structurally complex settings, with real-world biological networks showing benefits from locally targeted prior knowledge.
local causal discoveryoptimal adjustment setsmeek's rulesmarkov equivalencecausal effect estimation
Tightening the Score Matching Gap for Diffusion Models
The authors present tighter theoretical bounds on the score matching gap in diffusion models, addressing the discrepancy between sample quality and score matching loss. They analyze three metrics—KL divergence, reverse KL divergence, and Wasserstein distance—by leveraging the regularity of score estimators and contraction properties of backward processes. Key techniques include entropy flows, logarithmic Sobolev inequalities, and reflection couplings to link Langevin diffusion ergodicity to the score matching gap. Results indicate that score approximation quality significantly impacts gap closure, particularly at low noise scales, offering improved theoretical understanding of diffusion model performance.
diffusion modelsscore matching gaplogarithmic sobolev inequalitiesentropy flowslangevin diffusion
Environmental Drivers of Respiratory Disease: A District Level Analysis
This study presents a district-level analysis of environmental drivers of respiratory disease in Sri Lanka, integrating satellite-derived environmental data with healthcare records over 11 years. Using XGBoost models (R^2 = 0.937 for respiratory rates, R^2 = 0.976 for PM2.5) and SHAP analysis, it quantifies the relative impact of air quality (80.1%), forest degradation (15.6%), and fire activity (4.3%) on respiratory admissions. The Forest-Air-Health (FAH) Risk Index identifies Colombo, Gampaha, and Kalutara as highest-risk districts, providing an evidence-based framework for targeted policy interventions.
shapley additive explanationsxgboostpm2.5forest degradationrisk index
Learning Task-Sufficient World Models by Synergizing Agentic Exploration and Structured Modeling
The paper introduces a method for learning task-sufficient world models by synergizing agentic exploration with structured representation learning. The approach combines active environment probing (guided by an adaptive curriculum) with structured world-model learning to distill compact, control-relevant latent states. Empirical results demonstrate recovery of task-sufficient representations, leading to improved sample efficiency and generalization across skills, object-skill compositions, and novel tasks in continuous-control and robotic-manipulation benchmarks.
world modelsagentic explorationstructured representationtask-sufficientadaptive curriculum
Quadrature-Aware Complex-Linear Neural Operator for Boundary-to-Field Prediction in Resonant Acoustics
The paper introduces a quadrature-aware complex-linear boundary operator (CLBO) for efficient boundary-to-field prediction in resonant acoustics, preserving physical properties like complex superposition. The method couples learned basis functions through explicit surface-quadrature contraction, representing boundary excitation via coordinates, normals, and weights rather than flattened vectors. Evaluated against a complex DeepONet on 3D MRT lattice Boltzmann data, CLBO achieved 0.184 mean relative field error (vs. 0.367), 1.31×10^-7 superposition error, and 1.83×10^4 faster inference, demonstrating improved physical consistency and generalization.
complex-linear boundary operatorresonant acousticssurface-quadrature contractionboundary-to-field predictionlattice boltzmann solver
The Good, the Bad, and the Brittle: Benchmarking Robustness and Generalisation of Histopathology Foundation Models
The study benchmarks twelve pathology foundation models (PFMs) against ResNet baselines using the Robustness Evaluation and Enhancement Toolbox (REET) across eleven clinical perturbations and a Non-Redundant K-fold validation protocol. Introducing a Perturbation Performance Index (PPI), the analysis reveals PFMs outperform CNNs in robustness and domain generalisation, with mid-sized models (e.g., UNI2/Virchow-2) showing comparable resilience to larger systems. Results indicate diminishing returns from model scaling and systematic accuracy loss under distribution shift, highlighting the need for improved data quality and domain alignment over parameter count.
pathology foundation modelsrobustness evaluationperturbation performance indexdomain generalisationnon-redundant k-fold
NKI-Agent: Domain-Specific Fine-Tuning and Agentic Tool Use for Neuron Kernel Generation
NKI-Agent introduces the first system for automated Neuron Kernel Interface (NKI) generation on AI accelerators, combining domain-specific supervised fine-tuning (SFT) with a compile-verify-fix agent loop. The method adapts CUDA-Agent to Neuron hardware, curates 6,000 training tasks, and constructs NKIBench (250 tasks). Evaluated on Trn1 hardware, Claude Opus 4.8 with tool use achieves 77.3% pass rate versus 6% without tools; SFT-trained Qwen3-Coder-30B-A3B reaches 25.0% pass rate at 1/100th Claude Sonnet 4's cost. GRPO with binary compilation reward shows no improvement over SFT.
neuron kernel interfacesupervised fine-tuningcompile-verify-fixai acceleratorsgroup relative policy optimization
RL Forgets! Towards Continual Policy Optimization
The paper challenges the assumption that reinforcement learning (RL) inherently resists catastrophic forgetting in continual learning, introducing MRCL, a Multimodal Reasoning Continual Learning benchmark. It proposes Continual Policy Optimization (CPO), a replay-free method using KL regularization to constrain policy drift on prior tasks. Experiments on Qwen3-VL-8B show CPO reduces forgetting by 13.7% and improves pretrained capability by 7.0%, demonstrating efficacy across multiple model scales.
continual learningreinforcement learningcatastrophic forgettingmultimodal reasoningkl regularization
Optimal Mixture-of-Experts Model Averaging for Conditional Generative Models
The authors propose an optimal model averaging framework for conditional generative models, addressing scenarios where multiple candidate generators exist without tractable densities. They introduce StaticMA (static weight assignment) and MoEMA (input-adaptive weights via softmax neural-network gate), proving their asymptotic optimality and weight function consistency. The method handles Euclidean responses and extends to unstructured data via representation maps. Evaluations across tabular, image, and text modalities demonstrate MoEMA's superiority over baselines.
conditional generative modelsmodel averagingmaximum mean discrepancymixture-of-expertsneural-network gate
How Many Initial Points Does Bayesian Optimization Need?
The paper investigates the initialization phase in Bayesian Optimization (BO), demonstrating that the number of initial points $n_0$ exhibits a U-shaped tradeoff in total optimization cost (initial points plus BO iterations). Through empirical analysis across MLE, Bayesian MCMC, exact GP hyperparameters, and acquisition functions, the authors identify Thompson Sampling as $n_0$-agnostic but suboptimal. They attribute the U-shape to BO's boundary issue, where early evaluations focus on hypercube corners. Recommendations include multi-step lookahead BO, Thompson Sampling for fixed $n_0$, and larger $n_0$ when tunable.
bayesian optimizationinitialization phasethompson samplinggp hyperparametersacquisition functions
Structure-Specific Representational Priors Causally Control the Grokking Delay
The study causally demonstrates that grokking delay in neural networks is determined by the time required to form task-specific representational structures, not merely by optimization dynamics. Using a one-layer transformer trained on modular addition, the authors inject three types of representational priors via supervised-contrastive auxiliary losses: true task structure, coherent-but-wrong structure, and random partitions. Results show a clear gradation in generalization: true structure (22/30 runs), sibling structure (14/15), and random (0/20), with Fisher exact p=1.3e-7. Representation probes confirm structure formation precedes generalization, and only true structure accelerates grokking (up to 2.75×), albeit with dose-dependent and seed-dependent variability.
grokkingrepresentational priorsmodular additionsupervised-contrastive lossstructure formation
On the effectiveness of reward functions in reinforcement learning for confidence calibration of large language models
The paper introduces non-hackable confidence reward schemes for reinforcement learning (RL)-based confidence calibration in large language models (LLMs), addressing the issue of confidence reward hacking where LLMs may exploit poorly designed rewards to verbalize incorrect answers confidently. The method employs dual reward functions—one for correct answers and another for incorrect ones—and evaluates their impact on accuracy-calibration tradeoffs across datasets. Experiments reveal that reward scheme selection is dataset-dependent and should be treated as a hyperparameter for optimal performance. Code is publicly available.
reinforcement learningconfidence calibrationlarge language modelsreward hackinghyperparameter optimization
Legible-by-Construction: Attention and End-to-End Transformers
The paper extends prior work on interpretable transformers by developing legible attention mechanisms alongside previously modified feed-forward layers, creating an end-to-end interpretable model. The method modifies attention heads to produce readable feature detectors via sigmoid activation or Boolean operations on values, adding no parameters while maintaining performance. Results show 44-62% of value channels become crisp detectors in specialized designs, with legibility increasing with depth, and demonstrate parity with conventional baselines on language modeling tasks using a 125M parameter model.
interpretable transformersattention mechanismsboolean operationsfeature detectorslanguage modeling
SAD-LoRA: Spectral Alignment for Low-Rank Knowledge Distillation
SAD-LoRA introduces spectral alignment for low-rank knowledge distillation, explicitly controlling the rank-r weight subspace in LoRA-adapted students via a differentiable principal-angle loss on column spans. The method decomposes data-weighted distillation error into subspace misalignment, coefficient mismatch, and irreducible rank residual, addressing limitations of output-level KD. Experiments show SAD-LoRA reduces subspace misalignment from 51% to near-zero in synthetic tasks and improves rank efficiency in RoBERTa-large to RoBERTa-base distillation across GLUE tasks, achieving best results on SST-2 and CoLA at r=8.
knowledge distillationlow-rank adaptationspectral alignmentprincipal-angle losssubspace misalignment
Deep Learning for Dynamic Programming with Recursive Utility
The authors propose Certainty Equivalent Learning (CEL), the first deep learning algorithm for high-dimensional discrete-time dynamic programming with recursive utility. CEL jointly approximates value functions, policy functions, and certainty-equivalent functions using neural networks, avoiding explicit representations of recursive utility or Bellman equation evaluations. The mesh-free, simulation-based method handles non-differentiable transitions and high-dimensional spaces without relying on Euler equations. Applied to four control problems, CEL achieves 1.0e-4 to 1.0e-3 out-of-sample Bellman errors and matches VFI accuracy in small-noise robust control.
recursive utilitycertainty equivalentdynamic programmingneural networksbellman equation
On Preserving Geometrical Invariance for Superpixel Image Classification using Graph Transformer
The paper proposes SuperGT, a Graph Transformer framework for image classification that preserves translation/rotation invariance while capturing long-range dependencies. The method represents images as superpixel graphs to reduce redundancy, then processes them with a Graph Transformer and invariance-preserving preprocessing. Evaluated on CIFAR-10, SuperGT outperforms baselines and matches ShapeGNN's performance without requiring boundary point coordinates.
graph transformersuperpixelinvariance preservationlong-range dependenciesimage classification
Air-Plan: Query-Optimized Topology Selection for Over-the-Air Decentralized Federated Learning
AIRPLAN introduces a query-optimized topology selection framework for Over-the-Air Decentralized Federated Learning (OTA-DFL), bridging wireless federated learning and distributed query optimization. The method establishes a formal equivalence between OTA-DFL and distributed query processing, leveraging privacy-preserving Count-Min Sketch statistics to estimate workload characteristics and evaluate graph-aware cost models across candidate topologies. Experiments across five graph families, three vision benchmarks, four client scales, and multiple SNR settings demonstrate that AIRPLAN matches the oracle-optimal topology in 91.4% of workloads with less than 1.8% overhead. Theoretical error bounds for topology-aware sparsification show that well-connected topologies tolerate aggressive compression better.
over-the-air aggregationdecentralized federated learningquery optimizationcount-min sketchtopology-aware sparsification
Quantize the Target, Quantize the Drafter: Efficient Inference with Qwen3.5-4B
The work presents an efficient inference system for Qwen3.5-4B on resource-constrained NVIDIA A10G GPUs, combining quantization-aware distillation with speculative decoding. The method employs a two-stage training process for a block-diffusion drafter, first learning from a high-precision target then adapting to a quantized target, while reducing overhead via quantization and sliding-window attention. Results show a 6.978× speedup over baseline while maintaining quality thresholds, achieving 3rd place in the Efficient Qwen Competition.
quantization-aware distillationspeculative decodingblock-diffusion draftersliding-window attentionlow-latency inference
Robust Bayes-Assisted Conformal Prediction
The paper introduces RoBAS (Robust Bayes-Assisted Shrinkage), a framework for robust Bayes-assisted conformal prediction that maintains exact frequentist coverage while adapting to prior misspecification. RoBAS employs two nonconformity scores: one based on a heavy-tailed Bayesian working model (BWM) and an empirical Bayes shrinkage score, both reverting to the Distance-To-Average (DTA) baseline when prior information is unreliable. Experiments on tabular and image regression tasks show competitive performance under distributional alignment and reduced prediction set widths under distribution shift between training and calibration/test data.
conformal predictionbayesian working modelnonconformity scoresempirical bayesdistribution shift
Unified convergence analysis for gradient descent optimization methods in the training of deep neural networks
The paper presents a unified convergence analysis for gradient descent (GD) optimization methods in deep neural network (DNN) training, encompassing adaptive and accelerated variants like Adam, Nesterov accelerated gradient, and RMSprop. The analysis employs Kurdyka-Łojasiewicz (KL) inequalities to establish convergence to critical points for DNNs with analytic activations, including softplus and Gaussian error linear unit (GeLU). This framework generalizes existing results and provides novel convergence guarantees even for widely-used optimizers like Adam.
gradient descentkurdyka-łojasiewicz inequalityanalytic activationadaptive optimizationconvergence analysis
AI-RAN on NPUs: Baseband Processing Without Baseband Chips
The paper demonstrates that Neural Processing Units (NPUs) can support wireless baseband processing by resolving the architectural mismatch between AI inference workloads and communication algorithms. The key insight is a computational isomorphism where NPUs' matrix/vector engines inherently cover physical-layer operations, achieved by reconstructing communication algorithms to prioritize engine utilization over arithmetic operation count. Validation shows a complete OFDM transceiver implemented on an Ascend 310B1 edge NPU achieves end-to-end over-the-air transmission at 3.0 GHz using a USRP X300.
ai-ranneural processing unitsbaseband processingofdm transceivercomputational isomorphism
Channel-Adaptive Robust Aggregation for Over-the-Air Federated Learning in Heterogeneous Networks
The paper proposes CHARGE-FL, a channel-adaptive robust aggregation framework for Over-the-Air Federated Learning (OTA-FL) in heterogeneous networks. The method dynamically schedules aggregation based on channel conditions and client readiness, employing dual-purpose precoding to mitigate channel distortion and partial update bias. Evaluations under realistic wireless conditions demonstrate significant improvements in accuracy, stability, and convergence compared to state-of-the-art OTA-FL methods, particularly in scenarios with stragglers and noise.
federated learningover-the-air computationchannel adaptationrobust aggregationheterogeneous networks
Sangam: Efficiently Serving Diffusion LLMs with the AR Stack
Sangam introduces a serving system for diffusion language models (dLLMs) that addresses challenges in autoregressive-style KV caching due to bidirectional attention. The system employs a deficit token-budget scheduler, admitting in-flight decodes first and prefills only when the token budget permits, achieving amortized stall-free scheduling. Sangam adopts a hybrid serving strategy, overflowing prefills onto decode workers to mitigate prefill under-provisioning while protecting decode performance. Results show colocated serving reduces mean latency by 9-20% on decode-heavy workloads (LLaDA-8B ShareGPT), while hybrid execution cuts latency by 8-20% on prefill-heavy workloads (Dream-7B arXiv).
diffusion language modelskv cachingtoken-budget schedulerhybrid servingstall-free scheduling
Exploring Convolutional Neural Processes for Weather Downscaling
The study demonstrates Convolutional Conditional Neural Processes (ConvCNPs) as an effective method for statistical downscaling of daily maximum temperature from ~11km ERA5-Land to ~1km resolution over Switzerland. The architecture incorporates high-resolution elevation features from swisstopo DHM25, achieving a mean absolute error of 1.31°C and a CRPS-based skill score of 0.524, halving prediction error versus bilinear interpolation. Key findings include the elevation MLP's critical role, secondary benefits from seasonal features and Topographic Position Index, and graceful degradation under sparse inputs. Limitations include poor off-grid performance and overconfident uncertainty estimates due to Gaussian likelihood training.
convolutional conditional neural processesstatistical downscalingera5-landtopographic position indexgaussian likelihood
SpecGradFilter: A Spectral Gradient Filtering Framework for Taming Federated Heterogeneity
SpecGradFilter introduces a spectral gradient filtering framework to address statistical heterogeneity in Federated Learning (FL) by mitigating client drift through frequency-domain analysis. The method identifies a Spectral Bias of Drift, where low-frequency components encode client-specific shifts, and suppresses these discordant signals using FFT-based truncation or spatial approximations like Gaussian detrending. Experiments on CIFAR-10/100 and Tiny-ImageNet benchmarks demonstrate significant performance improvements in highly non-IID settings with minimal communication overhead, establishing a robust federated optimization paradigm.
federated learningspectral biasclient driftfrequency-domainnon-iid
Physics-Informed Graph Learning with Uncertainty Awareness for Open-Set Domain Generalization in Fault Diagnosis
Proposes PGU-OD, a physics-informed graph learning framework with uncertainty awareness for open-set domain generalization in fault diagnosis. The method integrates a physics-informed spectral attention module for condition-robust feature extraction, an uncertainty-aware adaptive graph learning mechanism for structural uncertainty mitigation, and a Gaussian-distribution-based adaptive boundary loss with dual-criteria open-set inference for decision optimization. Evaluations on two rotating machinery fault datasets show superior performance in known fault classification and unknown fault rejection under domain shifts compared to state-of-the-art baselines.
open-set domain generalizationphysics-informed learninguncertainty awarenessgraph learningfault diagnosis
XS-VLA: Coupling Coarse-grained Spatial Distillation with Latent Flow Matching for Lightweight Robotic Control
XS-VLA introduces a lightweight Vision-Language-Action framework for robotic control, addressing spatial blindness and computational inefficiency in existing models. The method employs a two-stage approach: (1) spatial semantic distillation from Qwen3-VL-4B into SmolVLM2-0.25B via coarse-grained spatial descriptions, and (2) conditioning a Latent Flow Matching policy using the enhanced backbone. The policy combines a Conditional Variational Autoencoder with Flow Matching dynamics to model multimodal action distributions. On the LIBERO benchmark, XS-VLA achieves state-of-the-art performance for models under 0.5B parameters, improving success rates by up to 7.2% and delivering a 3.2x speedup over previous lightweight flow matching policies.
spatial distillationlatent flow matchingconditional variational autoencodervision-language-actioncoarse-grained spatial descriptions
FedFFT: Taming Client Drift in Federated SAM via Spectral Perturbation Filtering
FedFFT addresses client drift in federated Sharpness-Aware Minimization (SAM) by filtering low-frequency perturbation inconsistencies via spectral analysis. The method applies Fast Fourier Transform (FFT) to locally computed SAM perturbations, attenuating divergent low-frequency components while preserving high-frequency consensus signals, without additional communication overhead. Experiments across multiple benchmarks and non-IID data distributions demonstrate consistent improvements over baseline SAM-based federated learning methods, particularly under severe heterogeneity.
federated learningsharpness-aware minimizationspectral perturbationclient driftnon-iid
Geometry of Ordinal Representations in Language Models
The study demonstrates that language models represent ordinal variables on geometrically structured manifolds, with dimensionality and coherence dependent on task properties. Analyzing Gemma-2-2B, Gemma-2-9B, and Qwen3-4B across four ordinal tasks (bracket depth, indentation, table position, numeric magnitude), the authors find 1D place-cell tiling emerges for locally computable variables, while cross-position integration requires higher-dimensional representations. Architecture-specific geometric computation is observed, with Qwen3-4B exhibiting stronger twisting for indentation than Gemma models. Activation patching confirms task-relevant information concentration in manifold subspaces, with ablation causing significant probe accuracy drops.
ordinal representationsmanifold geometryplace-cell tilingactivation patchingcross-position integration
ACE: Agentic Control for Embodied Manipulation via Zero-shot Workflow Reasoning
The paper introduces ACE (Agentic Control for Embodied Manipulation), a zero-shot workflow reasoning framework for tabletop pick-and-place tasks from natural language. ACE combines agentic workflow reasoning with visual grounding and reusable pick-and-place primitives, using a mask-mediated vision-action interface to bridge semantic reasoning and physical control. The system operates in a closed loop with multi-timescale memory, enabling online adaptation to failures and scene changes. Evaluated on complex tasks like equation formation and constraint-based retrieval, ACE achieves 50% and 70% success rates respectively, outperforming end-to-end baselines without task-specific retraining.
zero-shot learningworkflow reasoningvisual groundingmask-mediated controlmulti-timescale memory
Binary Iterative Method for Non-targeted Adversarial Attack
The paper proposes Binary Iterative Method (BinIM), a divide-and-conquer approach for optimizing parameters and hyper-parameters in non-targeted adversarial attacks. BinIM outperforms gradient-based methods like Fast Gradient Method and Basic Iterative Method on ImageNet classification tasks using pre-trained networks (InceptionV3, InceptionV2, ResNet V2 152). Evaluated on 1000 ImageNet samples, BinIM achieves misclassification confidence up to 0.995 while reducing true label probability to 2.21e-09.
adversarial attacksnon-targetedbinary iterative methodgradient-based methodsimagenet
Masked Generative-Contrastive Representation Learning for Cross-Dataset EEG-Based Emotion Recognition
We propose Masked Generative-Contrastive Representation Learning (MGCRL), a self-supervised learning framework for cross-dataset EEG-based emotion recognition. MGCRL integrates a region-aware spatiotemporal encoder with generative and contrastive learning to capture spatiotemporal dependencies, noise-robust fine-grained representations, and cross-subject-invariant features. The framework employs region-based graph convolution, joint embedding predictive architecture (JEPA) for generative learning, and masked-contrastive learning for temporal stability. Pretrained on the FACED dataset and fine-tuned on SEED-series datasets, MGCRL demonstrates strong generalization capabilities for cross-dataset EEG emotion recognition.
self-supervised learningeeg emotion recognitionspatiotemporal encodergraph convolutionjoint embedding predictive architecture
MDL Meets Latent Confounders: LNML-based Causal Discovery
The authors propose a novel MDL-based causal discovery framework that addresses nonlinear mechanisms and latent confounders by minimizing the luckiness normalized maximum likelihood (LNML) code-length. The method introduces $Δ$-pseudo-collinearity to identify dependencies induced by latent confounders and employs a greedy algorithm, Pseudo-Collinearity Guided Causal Discovery (PCG-CD), to determine causal relationships by selecting the shortest code-length of the causal model. Experimental results on synthetic and real-world datasets demonstrate the method's effectiveness in accurately recovering directed causal relationships and detecting latent confounders.
causal discoverylatent confoundersnonlinear mechanismsluckiness normalized maximum likelihoodpseudo-collinearity
CertMix: Certified, Data-Efficient Metamaterial Design by Affine Mixing of Aligned Neural-Implicit Weight Spaces
CertMix introduces a data-efficient framework for certified inverse design of mechanical metamaterials, addressing limitations of current learning-based methods. It represents unit cells as periodic neural implicit fields (SIREN signed-distance decoders) with aligned weight spaces, enabling affine mixing for targeted design. The method employs a differentiable periodic homogenizer, trust regions for validity, and split-conformal calibration for error certification. Results show CertMix achieves a scaled property error of $10^{-4}$ from just 50 exemplars, outperforming baselines trained on 1000 cells by 2-3 orders of magnitude. It is 57× faster than topology optimization, avoids artifacts, and extends to graded fields, 3D surfaces, and certified applications like running-shoe midsoles.
mechanical metamaterialsneural implicit fieldsaffine mixingperiodic homogenizersplit-conformal calibration
GlacierCastAI: Predicting Glacier Retreat from Multi-Modal Satellite Imagery and Climate Signals
GlacierCastAI introduces a multi-modal spatiotemporal forecasting approach to predict glacier retreat by fusing multi-temporal Landsat imagery, ERA5 climate variables, and Copernicus DEM terrain features. The architecture combines a ResNet50 spatial encoder, ConvLSTM temporal model, and cross-attention climate fusion module. Experiments across five glaciers show that adding ERA5 climate signals improves IoU from 0.326 to 0.337 (+3.4%), outperforming persistence (IoU 0.160) and linear trend baselines (IoU 0.169) by 89-99%. A lightweight climate-only MLP baseline achieves 98% of image-only performance with 85x fewer parameters. SHAP analysis identifies spring solar radiation as the dominant climate driver, aligning with known melt season dynamics.
spatiotemporal forecastingconvlstmcross-attentionera5shap analysis
Asymptotic-Preserving A Posteriori Analysis of Diffusion and Flow-Matching Samplers
The paper presents an asymptotic-preserving (AP) analysis framework for diffusion and flow-matching samplers, treating the terminal noise scale $σ_{\min}$ as a singular-perturbation parameter. It introduces an a posteriori audit to identify stable and uniformly accurate fixed-step samplers, computable without ground-truth scores. Key findings include: Euler in the $σ$-clock (DDIM update) is uniquely layer-exact; deterministic samplers achieve first-order uniform accuracy without $\log(1/σ_{\min})$ factors, while stochastic samplers incur logarithmic costs in path-KL divergence. Experiments on solvable models and EDM CIFAR-10 validate the theoretical predictions.
diffusion samplersflow-matchingasymptotic-preservingsingular perturbationpath-kl divergence
📰 Industry Media (7)
The foundational elements of AI architecture that IT leaders need to scale
The article identifies four foundational architectural elements for scalable AI deployment: (1) data quality preparation through standardized pipelines and governance, (2) context engineering via retrieval-augmented generation and vector databases to optimize input relevance, (3) embedded governance with LLM observability for cost control and security, and (4) human-in-the-loop expertise for workflow adaptation. These components address key challenges including hallucinations (Gartner predicts 60% project abandonment without AI-ready data) and operational inefficiencies (85% of IT leaders plan LLM observability per Elastic 2026). The framework emphasizes durable investments amid rapid model evolution.
retrieval-augmented generationllm observabilitycontext engineeringvector databasesdata governance
Liquid AI Open-Sources Antidoom: A Final Token Preference Optimization (FTPO) Method that Reduces Doom Loops in Reasoning Models
Liquid AI introduces Antidoom, a Final Token Preference Optimization (FTPO) method that mitigates 'doom loops'—repetitive output spans—in reasoning models. The method identifies the first token triggering a loop and retrains the model to prefer coherent alternatives at that position via FTPO, a DPO-like algorithm with logit-space KL divergence and multi-token optimization. On LFM2.5-2.6B, looping rates dropped from 10.2% to 1.4%; Qwen3.5-4B improved from 22.9% to 1%. The open-source pipeline requires ~3 hours on MI325 GPUs and preserves base model capabilities while reducing context waste.
doom loopsfinal token preference optimizationreasoning modelslogit-space divergencein-context repetition
Tencent Releases Hy3: An Open 295B Mixture-of-Experts (MoE) Model with 21B Active Parameters and 256K Context
Tencent introduces Hy3, a 295B-parameter Mixture-of-Experts (MoE) model with 21B active parameters per token, released under Apache License 2.0. The architecture employs 192 experts with top-8 routing, a Multi-Token Prediction (MTP) layer for faster decoding, and supports 256K context length. Benchmarks show strong performance in coding (78.0 on SWE-Bench Verified) and STEM (90.4 on GPQA Diamond), with reduced hallucination rates (5.4%) and improved multi-turn intent tracking (75.1% on MRCR). Deployment recommendations include 8 GPUs and tools like vLLM or SGLang.
mixture-of-expertsmulti-token prediction256k contextspeculative decodinghallucination rate
OpenAI Releases GPT-Realtime-2.1 and GPT-Realtime-2.1-mini for Low-Latency Voice Agents in the API
OpenAI introduces GPT-Realtime-2.1 and GPT-Realtime-2.1-mini, two low-latency models optimized for realtime voice and multimodal interactions. The mini variant features a single-model pipeline for audio processing, configurable reasoning effort (minimal to xhigh), and tool use, reducing p95 latency by ≥25% via improved caching. The models support WebRTC/SIP integration, with cached audio input priced at $0.30/1M tokens (vs. $10.00/1M for fresh input). Benchmarks show a 3x cost reduction for audio output ($20.00/1M tokens) compared to the full model ($64.00/1M).
realtime reasoningp95 latencywebrtctool usekv-cache
Building a Scaffold-Split Random Forest QSAR Co-Scientist for EGFR Inhibitor Discovery Using ChEMBL, RDKit, SHAP, and BRICS
The study presents an automated QSAR workflow for discovering EGFR inhibitors targeting the C797S resistance mutation. Using ChEMBL bioactivity data (9000 IC50 measurements), RDKit-processed Morgan fingerprints (2048-bit), and scaffold-split Random Forest modeling (400 trees), the method achieves R²=0.63 on unseen chemotypes. SHAP analysis identifies potency-driving substructures, while BRICS fragment recombination generates 12 novel drug-like candidates evaluated across multiple developability metrics.
qsar modelingmorgan fingerprintsscaffold splitbrics fragmentsshap analysis
Insilico Medicine advances AI drug for IPF to Phase III trials
Insilico Medicine's AI-discovered drug rentosertib advances to Phase III trials for idiopathic pulmonary fibrosis (IPF), demonstrating a mean forced vital capacity gain of +98.4 mL vs placebo (-20.3 mL) in Phase IIa. The Pharma.AI pipeline combines PandaOmics for multi-omics target discovery (identifying TNIK via causal inference) and Chemistry42 for generative molecular design using Generative Tensorial Reinforcement Learning, producing 79 candidates in 18 months. Proteomic aging clocks (ProtAge, OrganAgechrono) validate biological impact, with peer-reviewed structural biology (Nature Biotechnology) and clinical data (Nature Medicine) documenting the AI workflow from target to clinic.
generative tensorial reinforcement learningmulti-omics target discoveryproteomic aging clocksidiopathic pulmonary fibrosistnik inhibition
L’Oreal, Mondelez, and Nestle use AI to speed product development
L’Oreal, Mondelez, and Nestle employ AI to accelerate product development through predictive modeling and ingredient optimization. L’Oreal uses AI to simulate molecular interactions in skincare and haircare formulations, reducing development time by 4×. Mondelez leverages AI for recipe generation and optimization, achieving 60% improved performance in nutrition, sustainability, and cost for biscuit recipes. Nestle applies AI for packaging material discovery and natural dye substitution, targeting removal of artificial colorants by 2026. These implementations rely on in-silico testing, chemical language modeling, and generative AI, while maintaining human expert oversight in final validation.
predictive modelingin-silico testingrecipe optimizationchemical language modelinggenerative ai
Generated automatically at 2026-07-07 21:19 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.