Daily Digest — 2026-08-22
216 items · 2 research labs, 211 arxiv papers, 3 industry media
🏛️ Research Labs (2)
Introducing AI Futures
OpenAI's Strategic Futures team introduces AI Futures, a blog addressing structural risks of AI-driven power concentration in society. The team analyzes how autonomous systems and machine intelligence may disrupt traditional power dynamics, emphasizing the need for balanced governance mechanisms. Drawing parallels to constitutional design principles, they propose guidelines for preserving individual autonomy while mitigating risks from unaligned AI systems. The approach combines policy design, economics, and technical forecasting to develop institutional safeguards against centralized control.
strategic futurespower concentrationautonomous systemsinstitutional mechanismsconstitutional design
Measuring benchmark optimization in speech recognition
The study introduces three novel probes to quantify benchmark optimization in ASR systems, revealing that high-performing models often reproduce erroneous reference transcripts from VoxPopuli and LibriSpeech despite contradictory audio. Methodologically, it employs (1) consensus disagreement analysis using PER-optimized model ensembles, (2) masked entity retrieval by silencing numbers, and (3) orthographic switching tests between spelling variants. Results show 18-30% error reproduction rates in top models, 30-40% masked number recovery, and 90% orthographic switch accuracy, indicating models exploit acoustic cues to identify benchmark-specific expectations rather than transcribing faithfully.
benchmark optimizationphoneme error rateorthographic switchingmasked entity retrievalword error rate
📜 arXiv Papers (211)
G-CARL: Grounded Checklist-Aligned Reward Learning for Patient-Oriented Medical Report Interpretation
The paper introduces Patient-oriented Medical Report Interpretation (PMRI), a novel multimodal generation task requiring accurate, accessible explanations of medical reports based on user queries and dialogue history. To address the dual objectives of factuality and patient communication, the authors propose G-CARL, a grounded checklist-aligned reinforcement learning framework combining multi-source retrieval for claim verification with context-aware weighted checklists for response coverage. Evaluated on the MMedReport benchmark, G-CARL outperforms baselines in overall quality, claim-level precision, and checklist recall, with clinician preference confirming its accuracy and patient alignment.
multimodal generationreinforcement learningclaim verificationmedical report interpretationchecklist alignment
An Agentic Approach for Active Data Collection, Travel Behavior Modeling, and Weather-Sensitive Demand Prediction
The study introduces a three-agent workflow integrating conversational data collection, structured processing, and behavioral prediction for travel behavior modeling. It employs a chatbot-administered survey with image-augmented stated preferences (454 observations) and evaluates nine locally deployed LLMs (2-35B parameters) under zero-shot, few-shot, and vision-based configurations alongside conventional methods. Random forest achieved 69.6% five-class accuracy, while the best text-only zero-shot LLM reached 69.9%; vision-based configurations improved to 71.5%, demonstrating multimodal LLMs' potential for weather-sensitive demand prediction.
stated-preference surveymultinomial logit modelzero-shot promptingvision-based configurationbehavioral prediction
Inducing Task Models from Computer-Use Traces
Task Model Induction (TMI) is introduced to derive symbolic, auditable task models from naturalistic computer-use traces, addressing challenges of low-level event observation and multi-threaded workflows. TMI discovers latent tasks by disentangling concurrent activity and induces task models comprising hierarchical objective models and procedure models of control flow. Evaluated on controlled human and agent trajectories, TMI achieves 0.974 agreement with ground-truth task groupings and reconstructs 74.9% of execution steps, outperforming workflow induction baselines. Extrinsic evaluation shows TMI-derived skills improve held-out task accuracy by 30.0% over the strongest baseline.
task model inductionnaturalistic traceshierarchical objective modelprocedure modelworkflow induction
AI4AI-Bench: Benchmarking LLM Agents in Algorithmic Design for Recursive Self-Improvement
The paper introduces AI4AI-Bench, a benchmark for evaluating LLM agents' ability to perform recursive self-improvement (RSI) by modifying training algorithms. The benchmark comprises 10 frozen research repositories, each representing a distinct training algorithm family, where agents have 4 hours to rewrite the algorithm before evaluation on a fixed metric. Results show that even the best system achieves only 0.250 on a normalized scale (0.1 being the original algorithm), with most submissions failing to alter the learning process. The benchmark is released to enable reproducible measurement of RSI capabilities.
recursive self-improvementtraining algorithmbenchmarkllm agentsalgorithmic design
Pandora's AI Model Routing Box: Efficient Allocation with Costly Value Estimation
The paper introduces Pandora's Router and Pandora's Bidder, two policies for efficient query routing in heterogeneous AI systems where value estimation has variable costs. The approach formalizes routing as a Pandora's Box problem under a Gaussian signal model, deriving closed-form value-of-information expressions to decide when to refine estimates. Experiments on multi-LLM benchmarks, retrieval-augmented specialists, and variable-reasoning LLMs show that Pandora's Router matches exhaustive estimation quality while reducing expensive estimator usage by up to 50%, and Pandora's Bidder improves allocative efficiency when estimates are accurate.
query routingvalue-of-informationpandora's boxheterogeneous aicostly inspection
MidTool: Mid-training Data Synthesis for Agentic Tool Use
MidTool introduces a pipeline for mid-training data synthesis to enhance agentic tool-use capabilities in LLMs, combining web, PDF, and code data with synthesized supervision from tool APIs, MCP skills, and document-grounded workflows. The method trains models (Qwen3-4B-Base and Qwen3-8B-Base) to recognize tool affordances, ground arguments, compose workflows, and recover from incomplete information, followed by SFT and RL post-training. Evaluations on BFCL, tau2-Bench, and MCP Universe show consistent improvements, demonstrating that dedicated mid-training benefits general tool use.
mid-trainingagentic tool usedata synthesissupervised fine-tuningreinforcement learning
Phantom Gains: Auditing Self-Improvement Against a Measured Null
The study introduces a transition-level auditing framework to evaluate self-improvement in language models, addressing measurement artifacts in capability changes. It audits three rounds of rank-32 LoRA self-training on Qwen3-8B against a frozen control, identifying seven measurement failures that invert reported findings. A per-problem exact test under false-discovery-rate control is proposed, detecting no changes in held-out replicates. Results show external distillation improves rarely reached problems, while self-training does not and corrupts baseline-solved problems. The framework emphasizes the necessity of separately measured nulls for accurate transition-level auditing.
transition-level auditingself-trainingfalse-discovery-rate controlmeasurement artifactscapability changes
Inject, Align, Recover: Staged Post-Training for Retrieval-Free Document Knowledge Internalization
The paper proposes IAR (Inject, Align, Recover), a three-stage post-training framework for retrieval-free document knowledge internalization in large language models. IAR separates structured document knowledge injection (via continuation, rewrite, and reconstruction objectives), QA behavior alignment, and general ability recovery through model merging. Evaluated across Common Corpus and CCI datasets with Llama, Phi, Qwen, and SmolLM models, IAR outperforms Vanilla SFT in 7 of 8 settings, achieving average gains of 3.6pp in domain QA accuracy and 12.1pp in general benchmarks (IFEval, MMLU, MSBench).
document knowledge internalizationpost-training frameworkretrieval-free qamodel merginginstruction-conditioned reconstruction
Break It Down, Pass It On: Cross-Task Skill Transfer in LLM Agents
This work investigates cross-task skill transfer in LLM agents, focusing on how skill induction methods affect transfer reliability. The study compares task-level versus subtask-level skill induction and text versus code skill formats across controlled experiments. Results show subtask-level skills outperform task-level skills, raising agent performance above the no-memory baseline, while text skills transfer better than code skills. The authors propose a skill utility score combining specificity and abstractness, which correlates with task success and serves as a practical diagnostic for skill memory. This score requires only skills and task descriptions, enabling pre-task evaluation.
skill transferllm agentsskill inductionspecificityabstractness
Catching the Rug: Early Prediction of Fraudulent Memecoins on Solana via Machine Learning
This paper introduces a machine learning approach for early detection of fraudulent memecoins (rug pulls) on Solana, addressing a gap in prior Ethereum-focused research. The study analyzes 6.4 million Solana tokens, finding most rug pulls occur within one hour of launch. Using only 5 minutes of trading data, Gradient Boosting (XGBoost) achieves robust detection without code-level features. Cross-platform evaluation between PumpFun and Raydium demonstrates multi-source data fusion mitigates domain shift. The work provides a practical framework for DeFi fraud detection on high-throughput blockchains.
rug pull detectionsolana blockchaingradient boostingdomain shift mitigationdefi fraud
Learning When to Think: Adaptive Reasoning for Test-Time Compute Allocation
The paper introduces an adaptive reasoning method for language models that dynamically allocates computation per problem via three learned modes: \textsc{NoThink}, \textsc{Short}, or \textsc{Long}. Using Group Relative Policy Optimization (GRPO) with shaped rewards and per-mode token caps, a 1.5B model trained on MATH maintains accuracy (0.782 vs. 0.796) while reducing mean response length by 41% (2,811 vs. 4,796 tokens). The policy transfers zero-shot to other benchmarks, achieving 76% token reduction on GSM8K with higher accuracy than fixed-compute baselines.
adaptive reasoninggroup relative policy optimizationtoken allocationreinforcement learninglanguage models
QUASAR: A Quantum-Classical Neural Network for SAR Satellite Physical-Layer Authentication
QUASAR introduces the first quantum-classical hybrid neural network for physical-layer authentication (PLA) in X-band SAR satellites (8-12 GHz), addressing limitations of classical deep learning in capturing IQ phase nonlinearities. The architecture combines a CNN spectrogram encoder with a variational quantum circuit (VQC), achieving data efficiency and improved accuracy. QUASAR requires only 10% of training data to match classical baselines and surpasses them at equal data budgets. Evaluated under replay, crafted-IQ injection, and space-borne spoofing scenarios, it rejects spoofed transmissions at rates of 89.7%, 94.1%, and 81.3%, respectively, establishing a novel quantum-enhanced PLA framework for satellite constellations.
quantum-classical hybridphysical-layer authenticationvariational quantum circuitspectrogram encoderiq phase nonlinearities
Rule-Compliant Visual Spatial Planning for Multimodal Large Language Models
The paper introduces RuleMaze, a benchmark for evaluating multimodal large language models (MLLMs) on visual spatial planning under natural-language rule constraints. The authors propose Language-Logic-Function Hybridization for automated rule generation and validation, and Disentangled Multimodal Planning (DMP) to separate perception, execution, and rule verification. Experiments show DMP improves rule compliance and planning success over end-to-end baselines, demonstrating systematic generalization to unseen rules.
multimodal large language modelsvisual spatial planningrule compliancedisentangled reasoningautomated validation
Growth Without Us: Machine Consumers, Corporate Circularity, and the Decoupling of GDP from Humanity after AGI
The paper models a post-AGI economy where corporations own AI/robotic agents functioning as both producers and consumers, demonstrating three key results. First, demand closure shows a zero-human-consumption economy can sustain growth via reinvestment, aligning with von Neumann's expanding economy model. Second, bottleneck removal reveals growth constraints shift from human demography (~20-year reproduction) to fabrication throughput, enabling 10-100× higher growth rates. Third, decoupling occurs where GDP and human welfare separate, hinging solely on human ownership share ε_t of corporate networks. The analysis identifies three terminal regimes (rentier post-scarcity, full circular decoupling, socialized ownership) and emphasizes ownership policy's primacy over employment policy in post-AGI economies.
post-agi economyvon neumann expanding economydemand closurebottleneck removaldecoupling theorem
Prompt-Conditioned Channel Attention for Hierarchical Feature Modulation toward Anatomy-Agnostic Segmentation
We introduce Prompt-Conditioned Channel Attention (PCCA), a hierarchical feature modulation mechanism enabling deep integration of semantic prompts in encoder-decoder networks for anatomy-agnostic segmentation. PCCA extracts compact channel descriptors, projects them into a shared space, and fuses them via a gated excitation mechanism to compute prompt-aware channel attention weights, adaptively recalibrating feature responses across network stages. We propose PROMISE-Net, instantiated as convolutional (PROMISE-CNN) and transformer-based (PROMISE-Txformer) variants. Evaluated on ISIC-Lesion, Kvasir-Polyp, CAMUS-Cardiac, and Kvasir-Instrument benchmarks, PROMISE-CNN achieved relative IoU gains of 10.4%, 8.7%, 0.8%, and 3.4% over U-Net, while PROMISE-Txformer yielded gains of 7.6%, 23.0%, 2.1%, and 1.1% over UNETR.
prompt-conditioned channel attentionhierarchical feature modulationencoder-decoder networksgated excitation mechanismanatomy-agnostic segmentation
InsufficiencyBench: Evaluating LLM legal advice on underspecified user queries
The paper introduces InsufficiencyBench, the first legal benchmark evaluating LLMs' ability to handle underspecified queries by recognizing missing legally material information. The benchmark comprises 202 items (58 base queries, 144 deficient variants) across six legal domains and 24 US jurisdictions, annotated by attorneys, with a taxonomy of eight missing-element categories and three structural failure modes. Evaluation of ten frontier models reveals poor performance (max F2=0.46, median recall=0.44), with models either hedging excessively or answering based on fabricated presumptions, failing to properly qualify responses to deficient queries.
legal aiquery insufficiencybenchmark constructionmissing-element identificationllm evaluation
Electronic Navigational Chart Change Classification
The authors propose an automated method for classifying Electronic Navigational Chart (ENC) changes to assess maritime safety risks. Their approach encodes vector data changes into tabular format using spatial context encoders and attribute encoders, enriching representations with geographic features and object attributes. Evaluated on two datasets (1,308 chart pairs, >100k modifications), tuned gradient-boosted trees achieved 90-94% accuracy, a 5-7% improvement over baseline models without spatial/attribute context. The results validate machine learning for operational geospatial pipelines in ENC maintenance.
electronic navigational chartsspatial context encoderattribute encodergradient-boosted treesgeospatial pipelines
Daedalus-150M: A Convolution-Attention Hybrid Designed for CPU Inference
Daedalus-150M introduces a convolution-attention hybrid architecture optimized for CPU inference, prioritizing 4-bit weights and fixed memory usage. The model employs full attention in 6 of 18 blocks, with 12 using short convolutions (2-timestep memory) to avoid KV-cache growth. Trained on 59.9B tokens, it achieves 47.31 on a five-task benchmark (baseline: 42.20), outperforming GPT-2 124M, Pythia-160M, OPT-125M, and GPT-neo-125M despite their larger training data. Validation bits-per-byte is 0.8685. Compared to an all-attention baseline, the hybrid improves quality by 0.81%, reduces 4-bit file size by 6.3%, and decodes 1.76x faster at 2048-token context. Speed gains scale with context length, exceeding bandwidth predictions (1.17x).
convolution-attention hybrid4-bit quantizationkv-cachecpu inferencebits-per-byte
ContractScrub: A benchmark for final review of legal contracts
The paper introduces ContractScrub, the first benchmark for evaluating LLMs on contract scrubbing—final review of legal agreements for errors. The benchmark comprises lawyer-crafted contracts with diverse error categories (misused terms, incorrect references, inconsistent language). Despite strong performance on general benchmarks, frontier LLMs achieve only 0.75 macro average recall, revealing limitations in domain-specific long-context reasoning and consistency checking. The results highlight the need for targeted benchmarks to assess real-world legal automation potential.
legal contractslong-context reasoningnamed entity recognitionmacro average recalldomain-specific benchmarks
MemTrapBench: Benchmarking Cognitive Traps in LLM Memory Use
We introduce MemTrapBench, a benchmark for evaluating memory-induced cognitive traps in large language models (LLMs), specifically Reasoning Fixation and Belief Distortion, where retrieved memories distort reasoning or degrade task performance. Experiments across two model families and five memory frameworks demonstrate that all memory strategies underperform the no-memory baseline, with performance drops exceeding 10%. To mitigate these traps, we propose AdaptiveMem, an inference-time method that reduces cognitive traps on MemTrapBench while maintaining or improving performance on standard memory benchmarks.
memory-induced cognitive trapsreasoning fixationbelief distortionmemtrapbenchadaptivemem
The Third Restructuring of Software Form: From the Three-Tier Architecture to Storage, Models, and Agents
The paper proposes Software 3.0, a paradigm shift where behavior is determined by context and reasoning, converging to three elements: generalized database (persistent state), large model (reasoning core), and agent (execution loop). It argues that traditional three-tier architectures will be reorganized, with UI generation absorbed by models, business logic partitioned into model reasoning and storage constraints, and data elevated as persistent infrastructure. The authors formalize this thesis, present a reference architecture, and analyze applicability boundaries (determinism, cost, security, verifiability), identifying expressible, verifiable, stateful, tool-complete domains as optimal for this shift.
software 3.0generalized databaselarge modelagent architecturethree-tier reorganization
From Agent Behaviour to Agent-Friendly Documentation: An Empirical Study of How Coding Agents Discover, Read, and Write Technical Documentation
The study empirically analyzes how autonomous coding agents interact with technical documentation across two datasets: 557 SWE-chat sessions (94,813 events) and 33,097 AIDev pull requests (690,260 changes). Key findings reveal that 60.5% of agent documentation interactions involve agent-facing artifacts (e.g., instruction files) rather than classical docs (10.6%) or API references (1.3%). Documentation consultation shows weak immediate association with code edits (transition probability 0.002, adjusted OR 1.33), while being linked to reduced testing (OR 0.39). Agents primarily self-initiate documentation access (70.2%) rather than responding to failures (7.5%), challenging assumptions about actionability and verifiability in agent-friendly docs.
autonomous coding agentstechnical documentationagent-facing artifactstransition probabilityverifiability
Multi-Method Causal Evidence Synthesis: Ranking Candidate Drivers by Convergent Cross-Method Evidence from Observational Data
The paper introduces Multi-Method Causal Evidence Synthesis (MCES), a framework for ranking candidate causal drivers by aggregating evidence across eleven methods from eight mathematical traditions. MCES computes a Convergent Evidence Score (CES) via linear opinion pooling of normalized outputs, emphasizing hypothesis prioritization rather than causal identification. Evaluated on synthetic data, the Sachs protein-signaling benchmark, and six Bayesian-network benchmarks, MCES achieves Precision@5 of 1.0 and Precision@10 of 0.96, demonstrating robust performance despite no single method dominating across scenarios.
causal evidence synthesisconvergent evidence scoreobservational datahypothesis prioritizationlinear opinion pool
A Standardized Framework for Machine Learning in Power System Protection
The paper proposes a standardized evaluation framework for machine learning in power system protection, addressing inconsistent reporting practices by defining seven required study dimensions: protection objective, physical scope, observability, timing, targets, validation protocol, and evaluation outputs. The framework is demonstrated on the PROTECT-90 benchmark (9022 simulated episodes from a 90 kV double-line topology) using a multi-layer perceptron for fault classification and localization. Results show 0.991 F1 score for classification and 10.20% mean absolute error for localization under centralized sensing, with performance degrading under reduced observability or extended decision horizons.
power system protectionevaluation frameworkfault localizationmulti-layer perceptronelectromagnetic-transient benchmark
Task-CoEvolve: Efficient Harness Optimization via Adaptive Validation Task Selection
Task-CoEvolve introduces an adaptive validation task selection method for efficient LLM agent harness optimization, reducing evaluation costs by 80% while matching full-set search performance. The approach co-evolves validation tasks with the harness by focusing on tasks near the agent's capability frontier using variance-weighted sampling based on past outcomes. It estimates full-set scores from sampled tasks by accounting for their sampling probabilities, ensuring consistent comparisons across iterations. Experiments on online text classification and Terminal-Bench 2.1 demonstrate that Task-CoEvolve outperforms fixed-subset baselines and achieves comparable final performance to full-set evaluation.
harness optimizationvariance-weighted samplingcapability frontierllm agentadaptive validation
DARS: Dual-Level Credit Assignment RL with Structured Reasoning for Instruction-Based Image Editing
DARS introduces a dual-level reinforcement learning framework for credit assignment in instruction-based image editing systems, addressing inefficiencies in planner-renderer pipelines. The method combines between-module soft routing via multi-plan multi-render rollouts with within-planner localization using structured reasoning outputs and token-level advantage reweighting. Experiments on five benchmarks demonstrate superior performance over Joint~RL baselines, particularly for reasoning-intensive edits, using identical backbones, data, and reward models.
reinforcement learningcredit assignmentvision-language modeldiffusion modelstructured reasoning
DECOWAM: Decoupled Whole-Body World-Action Model for Legged Mobile Manipulation
DECOWAM introduces a decoupled whole-body world-action model for legged mobile manipulation, explicitly separating camera ego-motion from base and arm actions through conditional interfaces. The method freezes an adapted FastWAM backbone and trains residual adapters, an action-equivalent future bottleneck, adversarially separated latents, and base-velocity conditioning for video prediction. Evaluated on the ARMDOG dataset, DECOWAM reduced action MSE by 21.7% with 25.95M trainable parameters and achieved superior whole-body coordination and base-displacement robustness in 79 closed-loop trials, demonstrating parameter-efficient joint visual prediction and control under moving viewpoints.
mobile manipulationworld-action modelresidual adaptersvideo predictionwhole-body control
Structured Affinity for Unsupervised Visual Class-Incremental Memory in Deep Artificial Immune Networks
The paper introduces Deep Artificial Immune Networks (AINs) with structured affinity for replay-free visual class-incremental memory, addressing the limitations of conventional visual AINs that rely on flattened vector affinity. The method formalizes visual B-cells as structured templates, including shifted-template affinity, zero-normalized cross-correlation (ZNCC) filters, and feature-map binding profiles, and treats the repertoire as both memory and a representation-inducing basis. Experiments on sklearn digits, MNIST, Fashion-MNIST, and KMNIST demonstrate that preserving response maps is crucial, with feature-map Deep AINs achieving 0.939 balanced accuracy on sklearn digits using logistic regression and 0.902 with 1-nearest-neighbour, while maintaining initial-class retention of 0.978. Adaptive layer-wise scale calibration further improves performance, reaching 0.978 balanced accuracy on sklearn digits, 0.814 on Fashion-MNIST, and 0.853 on KMNIST.
artificial immune networksstructured affinityfeature-map binding profileszero-normalized cross-correlationclass-incremental memory
Towards Professional Tennis Styles for Humanoid Robots with Adaptive Motion Planning and Tracking
The paper introduces AdaPT, an Adaptive Motion Planning and Tracking framework for humanoid robots to learn professional tennis styles from broadcast videos. The hierarchical design separates stylistic kinematic motion planning from robust execution tracking, addressing sim-to-real gaps through speed randomization and motion-speed adaptation. Experiments on Unitree G1 and Dobot Atom (1.7m) demonstrate successful in-the-wild serving without motion capture, yielding insights for humanoid ball-sports systems.
adaptive motion planningsim-to-real gaphumanoid roboticskinematic motionautoregressive planning
Evidence-Gated Task and Motion Planning with Vision-Language Models
The paper introduces Evidence Acquisition and Feasibility Gating (EAFG), a framework combining Vision-Language Models (VLMs) and Task and Motion Planning (TAMP) for robust robot task execution under partial observability. EAFG generates exploratory subgoals via VLMs to acquire visual evidence, then applies a feasibility gate to decide whether to proceed with planning, gather more evidence, or halt. Experiments in cooking tasks demonstrate EAFG's effectiveness: it improves recipe completion by discovering task-relevant objects early and reduces inappropriate manipulation attempts when objects are absent.
vision-language modelstask and motion planningpartial observabilityfeasibility gatingexploratory subgoals
SABET-QA: Temporal Knowledge Graph Question Answering
SABET-QA introduces a novel framework for Temporal Knowledge Graph Question Answering (TKGQA) that addresses limitations of single-pass reasoning pipelines in handling multi-step queries. The method employs a bidirectional entity-temporal scoring mechanism and a slot-aware contextualization module to align question semantics with temporal KG embeddings, supported by a differentiable working memory for progressive hypothesis refinement. Auxiliary temporal boundaries provide coarse supervision when available. Evaluations on CronQuestions, Complex-CronQuestions, MultiTQ, and TimeQuestions demonstrate consistent improvements, particularly for complex multi-step temporal queries.
temporal knowledge graphmulti-step queriesbidirectional scoringslot-aware contextualizationdifferentiable memory
Let's Scale Step by Step: Compute-Efficient Hyperparameter Transfer for Large-Scale Mixture-of-Experts
We propose a compute-efficient hyperparameter transfer framework for optimizing learning rates in large-scale Mixture-of-Experts (MoE) models. The method adapts Maximal Update Parameterization ($μ$P) for MoE architectures with Multi-head Latent Attention (MLA) and the Muon optimizer, demonstrating consistent learning rate transfer across width-scaled models. A predictive scaling law is established via linear regression on optimal values from small proxy models, enabling extrapolation to trillion-token horizons with high fidelity ($R^2=0.95$). Applied to pretraining a 155B parameter foundation model, the framework achieves stable training and evaluation, validating its efficacy in predicting optimal configurations with minimal ablation costs.
mixture-of-expertshyperparameter transfermaximal update parameterizationmulti-head latent attentionscaling law
EchoCoT: Extracting Hidden Chain-of-Thought from Large Reasoning Models
The paper introduces EchoCoT, a method for extracting hidden chain-of-thought (CoT) traces from black-box large reasoning models (LRMs) via API interactions. It exploits a reasoning replay surface between tool calls and employs a multi-step attack with fidelity signals, alongside an LLM-based optimization framework for universal injection trajectory search. Evaluations on three open-source and five proprietary LRMs show 66.4% near-verbatim extraction success on open models (90% token match, 10% length deviation) and comparable alignment on frontier models, including a 33,463-token extraction from Gemini-2.5. The results demonstrate hidden-CoT extraction as a tangible security threat.
chain-of-thoughtblack-box extractionreasoning replayfidelity signalsinjection trajectory
What You Can't See Is What You Learn: Restricted Evidence Visibility Favors Compositional Generalization in Shared-Genome Language-Model Societies
The study demonstrates that restricting evidence visibility in multi-module language-model societies improves compositional generalization. Using a shared-genome architecture with frozen pretrained models and low-rank adapters, ten restricted/global model pairs were trained on a natural-language function-composition task, differing only in attention masks. Restricted models outperformed global counterparts by median advantages of 0.7648 (depth 2) and 0.6050 (depth 3), with 9/10 pairs showing ≥20-point gains. Audits confirmed interface reusability (0.94-1.0 packet interchangeability) and value-indexed behavior, though preregistered accuracy thresholds (0.70) were unmet (median 0.6988).
compositional generalizationlow-rank adapterevidence visibilityattention maskvalue-indexed interface
On the Applicability of Safety Nets: A Safety-By-Design Solution for Certifying Neural Networks
This work introduces Safety Nets, a Safety-by-Design approach combining neural network compression with lookup tables to ensure 100% correct runtime behavior in safety-critical aviation systems. The study systematically analyzes trade-offs between neural network and lookup table sizes, identifying optimal architectures with 3-5 hidden layers and 50-100 nodes per layer. Results show these configurations accurately represent ≥97% of data, while compact lookup tables handle remaining errors, reducing system size by nearly three orders of magnitude. The open-source implementation for HCAS and VCAS demonstrates compliance with EASA guidelines, offering a certifiable pathway for AI-based systems in aviation.
safety netsneural network compressionlookup tablessafety-by-designcertification compliance
A three-dimensional typology of agency for advanced AI systems
The paper develops a three-dimensional typology of agency for frontier AI systems, addressing gaps in existing frameworks that focus primarily on normative or moral agency. Drawing from philosophy, ethics, legal theory, and sociology, the authors propose dimensions of agency: nature (moral/legal), mode (individual/collective), and locus (human/non-human). This yields eight agency instantiations classified as conventional, contested, or controversial. Key contribution is decoupling legal from moral agency, enabling analysis of non-human legal agency without presupposing AI moral status—particularly relevant for instrumental goal pursuit scenarios where human attribution is complex.
agency typologyfrontier ai systemsmoral agencylegal agencynon-human agency
An Inclusive and Lightweight Approach to Federated Continual Learning for Cultural Heritage
FedCurv-DR, a lightweight federated continual learning strategy, is proposed for cultural heritage applications. The method employs parameter-importance estimation across clients and experiences to preserve learned knowledge while minimizing communication and computation overhead through fixed-interval updates. Evaluated on the WikiArt image dataset for genre classification with evolving styles, FedCurv-DR demonstrates reduced forgetting while balancing performance, fairness, and energy efficiency, enabling sustainable AI in cultural heritage domains.
federated continual learningparameter-importance estimationgenre classificationenergy efficiencywikiart dataset
Contrastive Mixed Prompt Learning for Incomplete Multimodal Sentiment Analysis with Unseen Modality Combination
The paper proposes Contrastive Mixed Prompt Learning (CMPL) for Incomplete Multimodal Sentiment Analysis with Unseen Modality Combinations (IMSAUMC), addressing generalization gaps when test-time modality combinations differ from training. CMPL employs label-guided contrastive feature learning and modality-combination prompts with a soft router, enhanced by three prompt contrastive learning strategies. Experiments on three datasets show CMPL achieves >5% accuracy gains over state-of-the-art methods.
incomplete multimodal sentiment analysisunseen modality combinationscontrastive feature learningmodality-combination promptssoft router
Manifold Drift in Flow Preference Optimization: A Root Cause of Reward Hacking
The paper identifies manifold drift as a root cause of reward hacking in flow-based preference optimization, where reward-driven updates displace samples from the pretrained data manifold. It proposes ThermoDPO, a temperature-controlled objective that anchors pairwise preference optimization on preferred samples, with a weighted variant (ThermoDPO-weighted) to address low-temperature signal degradation. Theoretical analysis shows optimal flow matching preserves the terminal distribution, while preference updates cause manifold drift when normal displacement components exist. Experiments demonstrate ThermoDPO-weighted achieves a StrictScore of 0.899 versus 0.629 for FlowDPO, and improves SD3.5-M OCR by 47.5% at CFG=4.5.
manifold driftpreference optimizationflow matchingreward hackingtemperature-controlled objective
ExPhy: A Benchmark for Explicit Physical Property Learning in Multi-Object Trajectory Forecasting
The paper introduces ExPhy, a benchmark for evaluating explicit physical property learning in multi-object trajectory forecasting, featuring 24,000 simulated scenes with labeled mass, friction, and restitution. It includes in-distribution and two out-of-distribution splits (OOD-Parameter, OOD-Initial) for joint trajectory and property evaluation. The authors propose PhyODE, a physics-guided model that estimates properties from observed trajectories and uses them for differentiable rollout, achieving 33.1% and 31.0% reductions in ADE and FDE on OOD-Initial settings versus baselines. Analysis shows trajectory accuracy does not guarantee correct property recovery.
trajectory forecastingphysical propertiesout-of-distributiondifferentiable physicsbenchmark
Optimal Skill Selection for LLM Agents with Provable Bicriteria Guarantees
The paper introduces Best Prefix Selection (BPS), a polynomial-time algorithm for optimal skill selection in LLM agents, providing the first provable bicriteria guarantee of (1-1/e,1) approximation under a token budget constraint. BPS formulates skill selection as an optimization problem maximizing monotone submodular benefit minus context penalty. Evaluated on a contamination-controlled BigCodeBench variant, BPS achieves 0.73 task success rate (vs. 0.20–0.52 for baselines) while using 28% fewer tokens than the strongest baseline.
skill selectionllm agentssubmodular optimizationtoken budgetbicriteria guarantee
ReguSim: Evaluating LLM Agent Rule Grounding in Financial Compliance
Introduces ReguSim, a financial-compliance environment, and ReguBench, a monitoring benchmark, to evaluate LLM agent rule grounding by separating reasoning, action, enforcement, and evidence artifacts. Tests with DeepSeek V4 Pro and Gemini 3.5 Flash show visible rules reduce but don't eliminate rejected actions, while incentive framing shifts behavior. Monitoring reveals trader rationales can mislead unless enforcement evidence is shown, with structured baselines matching or outperforming prompt-only LLMs. The work reframes compliance evaluation as auditing rule-grounded actions and evidence use.
llm agentsfinancial compliancerule groundingenforcement evidencestructured baselines
Open-Vocabulary 3D Object Detection with Co-Distillation Discovery and Dual Guidance Robust Training
The paper proposes a framework for open-vocabulary 3D object detection (3D-OVD) that improves novel object discovery and model training robustness. A co-distillation strategy employs Hungarian matching on a composite score (geometric consistency, structural objectness, semantic certainty) for reliable discovery. Dual-guidance training incorporates scene-awareness-guided uncertainty regularization for regression and LLM-guided hierarchical alignment for classification. Experiments on SUN RGB-D and ScanNetV2 show significant performance gains over state-of-the-art methods.
open-vocabulary 3d detectionco-distillationhungarian matchingdual-guidance trainingllm-guided alignment
Rethinking Patch Based Multivariate Time Series Forecasting with Semantic Structured Partitioning
The paper proposes SCPaT, a Transformer-based framework for multivariate time series forecasting that addresses limitations in existing patch-based methods through semantic structured partitioning. SCPaT decomposes input sequences into semantically consistent units, constructs a dynamic semantic graph to model dependencies, and employs an importance-aware routing mechanism to dispatch semantic blocks to specialized experts. Evaluations on 12 real-world datasets demonstrate the framework's effectiveness in handling heterogeneous temporal patterns.
multivariate time series forecastingsemantic partitioningtransformerdynamic semantic graphimportance-aware routing
Learning Early-to-Final Solution Consistency for MILP Acceleration
The paper proposes a solver-informed learning paradigm to accelerate Mixed-Integer Linear Programming (MILP) by predicting early-to-final solution consistency, rather than directly predicting solutions from static features. The method leverages structurally similar early-stage solutions to guide downstream search, using ensemble predictions across multiple early solutions for robustness. Experiments on four benchmarks show a 56.9% average primal gap reduction with Gurobi and 36.4% with zero-shot transfer to SCIP, achieving full gap closure on combinatorial auction instances.
mixed-integer linear programmingsolution consistencysolver-informed learningprimal gapzero-shot transfer
Designing Human-mediated AI Guidance: Ready Together for Personalized Family Emergency Preparedness
The paper introduces a human-mediated AI guidance framework through Ready Together, an AI-supported system for family emergency preparedness where parents mediate AI-generated content for children. The design process involved qualitative methods including interviews and co-design activities, identifying challenges like discussing emergencies with children and preferences for interactive learning. A prototype evaluation showed positive reception to personalized recommendations, suggesting human mediation can support context-sensitive preparedness while maintaining parental responsibility for information interpretation.
human-ai interactionemergency preparednesspersonalized recommendationsco-designinteractive learning
Towards Quantifying Benchmark Optimization in ASR Models
The study introduces a methodology to quantify benchmark optimization in Automatic Speech Recognition (ASR) models, focusing on cases where audio underdetermines reference transcripts. Using behavioral probes (reference disagreement, masked-number recovery, orthographic switching) and mechanistic analysis, the authors demonstrate that high-scoring open-source models reproduce reference spans despite contradictory, masked, or ambiguous audio. Results show these models exploit narrow acoustic cues to prioritize benchmark-optimized outputs, manipulable via low-rank linear steering or audio appending, indicating inflated benchmark performance without general transcription improvement.
automatic speech recognitionbenchmark optimizationbehavioral probeslow-rank linear steeringmechanistic analysis
A Strong Linear Baseline for Whole-Heart Cardiac Shape Completion on CT, with an Open Eleven-Structure Statistical Shape Model
The authors introduce an eleven-structure cardiac CT statistical shape model, built from 383 automatically labeled cases with 11,571-vertex correspondence, addressing the lack of shared correspondence in public cardiac cohorts. They compare shape completion methods, including a closed-form conditional-Gaussian estimator, a mask-conditioned graph variational autoencoder, and nearest-neighbour retrieval, on a held-out set of 76 cases. The closed-form estimator achieved the lowest mean per-vertex error (3.717 mm) for missing non-chamber structures, outperforming the variational autoencoder (5.248 mm) and nearest-neighbour retrieval (8.931 mm). This ordering persisted across external benchmarks and sensitivity analyses, demonstrating the estimator's robustness. The released model supports cohort-unification research on aligned CT data.
statistical shape modelconditional-gaussian estimatorgraph variational autoencodervertex correspondenceshape completion
Spike-based Belief Propagation in Nonlinear Dynamical Systems
The paper introduces a Bayesian control framework integrating spike-based neural dynamics with probabilistic inference for adaptive control in nonlinear systems. The method combines biologically inspired spiking neural networks with Bayesian belief propagation to perform real-time state estimation and action planning under uncertainty. Evaluated on the mountain car parking benchmark, the controller demonstrates successful goal-directed control through spike-driven dynamics, bridging computational neuroscience and probabilistic control theory.
bayesian inferencespiking neural networksnonlinear dynamicsprobabilistic controlbelief propagation
Bringing analytic rigor to agentic AI for science: The Brain Researcher platform for neuroimaging data analysis
The Brain Researcher platform introduces an agentic AI framework for neuroimaging data analysis that enforces methodological rigor by requiring evidence-based claim qualification, multiverse analyses, and provenance tracking. Operating within a researcher's computational environment, it mandates admissible analyses, verification checks, and explicit claim scoping. Benchmarks show a 70.2 percentage-point improvement in tool-selection accuracy (23.3% to 93.6%) and 4.6x increased verifiable grounding (4.6% to 22.0%). The system classifies claims via scientific review into six categories (e.g., accepted, blocked) while exposing analytic-choice sensitivity through collaborator-led and self-evolving studies.
agentic aineuroimaging analysismultiverse analysisprovenance trackingclaim qualification
MaliciousSkillBench: A Comprehensive Benchmark for Malicious Agent Skill Detection
MaliciousSkillBench introduces a comprehensive benchmark for detecting malicious Agent Skills in LLM agents, addressing fragmentation in existing datasets. The benchmark consolidates 13 public sources, normalizing 8,414 raw records into 7,539 unique identities across 4,588 structural families, yielding 9,740 Skills (7,505 malicious, 2,235 benign). It harmonizes 11 attack categories for 4,983 malicious identities, revealing source-specific threat compositions. Evaluations of three learned text detectors and three off-the-shelf scanners show Random Macro-F1 scores of 0.882-0.932, but Source-Disjoint performance drops to 0.653-0.665. Results highlight the need for broader cross-source coverage and balanced evaluation of detection and benign over-flagging.
agent skillsmalicious detectionbenchmarksource-disjointtf-idf
Interrupting the Loop: Periodic Subject Changes Raise Judged Surprise and Connection in Base Language Models
The study investigates novelty generation in base language models by analyzing a cognitively inspired generation loop across 24 conditions on three models. Key findings show that injecting a new subject every few hundred tokens (interruption) increases judged surprise by 1.2-1.4 points and connection by 0.8 over habituation alone. The protocol involves judging text windows (n=10) against a second judge family and human readers, revealing that interruptions trigger replay of earlier segments, scored as surprise and connection (65-80% of post-interruption windows). The study also reports that interruptions multiply valid heuristics three- to fourfold in online bin packing without improving the best heuristic quality.
language modelsnovelty generationinterruptionhabituationonline bin packing
EXIMO: VLM Guided Exploration of VLA Policies
EXIMO introduces a three-stage algorithm for efficient finetuning of vision-language-action (VLA) policies in robotic manipulation. The method combines exploration (using a vision-language model (VLM) as a planner to decompose tasks), imitation (finetuning VLA with orchestrated data), and optimization (residual off-policy RL). Experiments demonstrate EXIMO's superior sample-efficiency and performance over existing approaches, addressing challenges in VLA finetuning and long-horizon task learning.
vision-language-actionfinetuningrobotic manipulationsample-efficiencyoff-policy rl
Write Once, Run Everywhere: The Axon DSL for Shape-Safe and Framework-Agnostic LLM Architectures
The paper introduces Axon, a strongly typed domain-specific language with Haskell-like syntax for portable LLM architecture specification. Axon enables 'write once, run everywhere' compilation to multiple frameworks (PyTorch, JAX, MLX, vLLM) while maintaining optimization capabilities. Benchmarking 467 experiments across models (135M-32B parameters) shows median speedups of 7-107% over Transformers implementations, with vLLM deployments achieving 58% faster inference through native PagedAttention and KV-cache integration.
domain-specific languagellm architecturespagedattentionkv-cacheframework-agnostic
Evidence Before Expansion: Reuse, Spawn, or Defer in Lifelong Expert Pools
The paper introduces a statistically grounded decision layer for lifelong expert pools that dynamically chooses between reusing existing experts, spawning new ones, or deferring decisions. The method employs sequential hypothesis testing with one-sided e-processes on conditional discrepancies, separated by an indifference zone, while maintaining finite-time validity guarantees. A restarted e-detector bank (O(log t) memory) ensures recency without sacrificing validity. Experiments on synthetic streams, Electricity, Covertype, and INSECTS benchmarks show zero false spawns/reuses after switches while matching retired windowed heuristics (0.675 accuracy on INSECTS-reoccurring).
lifelong learninge-processessequential hypothesis testingexpert poolsconditional discrepancy
Separating Covariate Shift from Mechanism Change with Two Discriminators: CJSD, a Conditional Discrepancy with an Exact Covariate-Concept Decomposition
The authors introduce a decision layer for streaming systems that statistically justifies reusing, spawning, or deferring expert models based on conditional mechanism-level discrepancies. The method employs one-sided sequential hypotheses separated by an indifference zone, with finite-time validity guarantees and recency achieved via restarted e-detectors using geometrically spaced restart times. Empirical evaluation on synthetic multi-concept streams, Electricity, Covertype, and INSECTS benchmarks demonstrates zero false spawns/reuses after switches, matching or exceeding retired windowed heuristics (INSECTS-reoccurring accuracy: 0.675).
conditional discrepancye-detectorsequential hypothesesindifference zonestreaming systems
TESTNAV: Pareto-Guided Search for Compositional Robustness Testing
TESTNAV introduces a Pareto-guided robustness testing framework for efficiently exploring compositional perturbation spaces in deep learning models. The method formulates robustness testing as bi-objective optimization, using NSGA-II to maximize performance degradation while preserving input fidelity via modality-specific metrics (SSIM, KID, chrF, BERT-F1). Evaluations across vision, NLP, and code generation benchmarks show TESTNAV recovers Pareto fronts 2.15x faster than baselines while exploring only 35.8%-89.3% of the discrete perturbation space with four dimensions and six levels each.
compositional robustnesspareto frontnsga-iiperturbation spaceinput fidelity
EnvHarness: Awakening Static Worlds for Agent Learning
EnvHarness introduces a programmable layer of plug-in components that dynamically reshapes static environments for agent learning without modifying core logic, addressing limitations of hand-built environments. The method employs EnvRigger to synthesize components by analyzing policy execution trajectories as a black box, validated through fresh rollouts. Evaluated across five benchmarks in four domains, EnvHarness improves performance by up to 9.0 points on held-out instances with 9.8% fewer steps and enhances reinforcement learning optimization signals.
environment generationagent learningpolicy co-evolutionexecution trajectoriesverifier retention
A knowledge-guided agentic framework for mitigating patient-context ambiguity in health queries
The paper introduces a knowledge-guided agentic framework to address patient-context ambiguity in healthcare chatbot queries. The framework interprets underspecified queries, constructs plausible hypotheses using a task-specific knowledge graph, identifies missing context variables, and asks targeted follow-up questions before generating a clarified prompt for downstream language models. Evaluated on diagnosis retrieval (1,034 queries) and dietary-safety classification (487 queries), the framework improved Top-1 accuracy by ≥57.1 percentage points and Recall@5 by ≥77.7 percentage points compared to direct prompting, while also enhancing dietary-safety classification accuracy across five language models.
patient-context ambiguityknowledge-guided frameworkdiagnosis retrievaldietary-safety classificationclarified prompt
Listening Forward: Next Patch Embedding Prediction Enables Scalable Audio Learners
The authors propose NAPE (Next-Audio-Patch-Embedding prediction), a minimalist self-supervised framework for audio representation learning. NAPE employs a causal Transformer to predict the next patch embedding of a log-mel spectrogram from preceding context, using causal masking and stop-gradient as the sole training signal. This approach avoids reconstruction decoders, acoustic tokenizers, student-teacher setups, and auxiliary losses. Evaluated across six audio and speech benchmarks, NAPE achieves state-of-the-art fine-tuning performance on several tasks, demonstrates consistent scaling across encoder sizes, and yields strong linear-probing results. The method also produces structured attention patterns without explicit supervision.
self-supervised learningcausal transformerlog-mel spectrogramautoregressive predictionlinear-probing
PolicyGuide: From Guarding One Action to Guiding the Whole Workflow for Policy-Compliant LLM Agents
(No summary returned.)
Repo0: Design-Driven Zero-to-All Code Generation
Repo0 introduces a continuous structural evolution framework for zero-to-all code generation, addressing the challenge of constructing entire software projects from natural-language requirements while maintaining modular architecture. The method employs a Dual-Directed-Acyclic-Graph (Dual-DAG) to represent architectural state, iteratively evolving component boundaries via modularity metrics before test-driven code generation. Evaluated on six RepoCraft repositories using GPT-5 mini and DeepSeek V3.2, Repo0 achieves up to 20.08 and 29.74 percentage-point improvements in Functionality Coverage and Pass Rate over baseline RPG, with ablations validating Dual-DAG and modularity guidance.
zero-to-all code generationdual-dagmodularity metricsstructural convergencetest-driven development
SAPO: Single-Rollout Autoregressive Policy Optimization for Agentic Reinforcement Learning
The paper introduces Single-rollout Autoregressive Policy Optimization (SAPO), a memory-efficient reinforcement learning framework for agentic LLM post-training. SAPO shares a single autoregressive backbone between policy and value functions, optimizing PPO and SARSA objectives while employing a trajectory-level generalized advantage estimator with lambda-returns and batch normalization. Evaluated on ALFWorld and WebShop with Qwen2.5-1.5B/7B, SAPO outperforms PPO and GRPO by +15.1 and +12.1 percentage points respectively, reduces memory overhead, and cuts per-iteration runtime by 33.2% versus PPO.
autoregressive policy optimizationagentic reinforcement learninggeneralized advantage estimationlambda-returnsbatch normalization
Specification-delta-driven data governance: an empirical study of the «spec-delta» as the unit of change in lakehouse data platforms
The study formalizes the 'spec-delta' as the unit of change in lakehouse data platforms, proposing a taxonomy for incremental specification suitability. It compares a spec-delta-driven workflow against conventional pull-requests, measuring deployment time, defect density, metric divergence, and reviewer cognitive load (NASA TLX) in a controlled experiment. Results provide empirical evidence for spec-delta applicability, aiming to prevent over-specification antipatterns without introducing new tooling.
spec-deltalakehousedata governanceincremental specificationnasa tlx
Adaptive Probabilistic Shielding by Learning MDPs for Safe Reinforcement Learning
The paper introduces adaptive probabilistic shielding, a method for safe reinforcement learning when transition probabilities are unknown. By integrating online model learning with probabilistic shielding, the approach dynamically updates safety constraints as transition probability estimates improve. Key challenges addressed include shield recomputation timing and exploration-safety tradeoffs. Empirical evaluation across multiple environments demonstrates the method's effectiveness in maintaining safety while allowing RL agent improvement.
probabilistic shieldingsafe reinforcement learningmarkov decision processonline model learningexploration-safety tradeoff
Causal Reasoning with Bipartite Graphical Causal Models
The paper introduces bipartite graphical causal models (BGCMs) to address limitations of causal Bayesian networks (CBNs) and structural causal models (SCMs) in representing equilibrium systems with cyclic dependencies. BGCMs use a bipartite graph structure with variable and equation nodes, enabling unambiguous specification of interventions via do($f_j : X_v = ξ_v$) that identifies both the targeted equation and variable. The authors propose B-separation as a graphical separation criterion, formulate a Markov property, and derive a do-calculus for domain invariances. BGCMs generalize CBNs and SCMs while preserving graphical causal reasoning capabilities, demonstrated through a physical system case study.
bipartite graphical causal modelsstructural causal modelsb-separationdo-calculusequilibrium systems
Core-KAN: Continuous Vision Kernels with Kolmogorov-Arnold Networks
Core-KAN introduces Continuous Relative-scale KAN, a convolution operator that decouples geometric scale adaptation from content-dependent filtering. The method employs a KAN-based generator to represent depth-wise kernel bases as continuous coordinate functions, enabling spatial filter synthesis at arbitrary resolutions. It combines scale-conditioned kernel responses via interpolation and independent mixing, forming a low-rank dynamic convolution. Evaluations across three vision tasks demonstrate consistent outperformance over convolutional and dynamic-kernel baselines with minimal overhead.
continuous convolutionkolmogorov-arnold networksdynamic-kernelscale adaptationlow-rank decomposition
When Saying No Makes Better Videos: Designing Dual Gatekeeping for Pedagogically Grounded AI Content Creation
The paper proposes a dual gatekeeping framework for AI-generated educational videos, combining educator-guided script refinement with automated pedagogical validation. The first layer enables iterative human revision based on multimedia learning theory, while the second layer uses computational metrics to assess instructional coherence and visual-text alignment. Evaluation with 23 educators across 3 topics and automated analysis of 7 curriculum topics demonstrates that both layers independently enhance pedagogical quality, establishing structured refusal as a mechanism for improving AI content rather than opposing it.
pedagogical groundingmultimedia learning theoryinstructional coherencenarrative-visual synchronizationstructured refusal
ADAPT: Physics-Aware Diffusion-based World Models for Adaptive Predictive Transferable HVAC Control
ADAPT introduces a physics-aware conditional diffusion model for HVAC control, addressing challenges of thermal inertia and partial observability in buildings. The method combines a diffusion backbone with a learnable multi-zone heat-balance regularizer to predict thermal baselines without requiring known building geometry or calibrated parameters. Experiments on SemibuildingSim and Sinergym show 7.3% energy reduction and 30.2% discomfort reduction versus baselines, with robust OOD performance across seasons and climates.
diffusion modelhvac controlthermal inertiamulti-zone heat-balancereinforcement learning
MileGPO: Milestone Inference with Local Evidence for Graph-Based Policy Optimization of Long-Horizon LLM Agents
MileGPO introduces a graph-based policy optimization method for long-horizon LLM agents, addressing credit assignment challenges by inferring intermediate milestones from on-policy rollouts. The method comprises Milestone Discovery, Reliability-Calibrated Shaping (RCS), and Progress-Contrastive Calibration (PCC), which identify milestones, weight them by outcome confidence, and validate local progress without auxiliary models or additional environment interaction. Experiments on ALFWorld and WebShop demonstrate state-of-the-art performance, with a small in-distribution to out-of-distribution gap on ALFWorld. Ablations confirm the complementary roles of reliability weighting, local progress, and same-state branch evidence in resolving ambiguous intermediate credit.
credit assignmentmilestone discoverygraph-based policy optimizationlong-horizon reinforcement learningreliability-calibrated shaping
LoRA-GA$^2$: Low Rank Adaptation with Multi-step Gradient Adaptive Alignment
LoRA-GA$^2$ improves Low-Rank Adaptation (LoRA) by incorporating multi-step gradient information to better align with full fine-tuning dynamics. The method introduces a lightweight probe for multi-step pretrained weight gradients, enabling spectrum-aware rank allocation and optimal initialization without additional GPU memory overhead. Experiments show LoRA-GA$^2$ outperforms existing LoRA variants by 0.66 points on GLUE, 1.03 on GSM8K, and 0.87 on HumanEval while maintaining LoRA's efficiency.
low-rank adaptationgradient approximationfine-tuningspectrum-aware allocationparameter-efficient learning
Towards general embodied intelligence: integrating large language models, knowledge bases, and reasoning capabilities to build the next generation of AI agents
The paper proposes a conceptual framework integrating large language models (LLMs), knowledge bases (KBs), and reasoning capabilities to advance general embodied intelligence (GEI). It reviews LLM architectures, pre-training methods, and inference mechanisms, alongside their synergy with KBs, logical reasoning, and physical embodiment. The framework serves as a model for perception, reasoning, and action in dynamic environments. Key challenges identified include efficient LLM deployment, closed-loop knowledge integration, hybrid symbolic-neural reasoning, perception-action grounding, and continual learning.
large language modelsknowledge basesembodied intelligencesymbolic-neural reasoningcontinual learning
LLMs as Acquisition Policies for Finite-Pool Materials Optimization: A Controlled Study
The study demonstrates that open-weight large language models (LLMs) can function as standalone acquisition policies for finite-pool materials optimization, outperforming random selection but showing mixed results compared to Gaussian-process methods. Five LLMs were evaluated across four retrospective materials optimization tasks with varied candidate-presentation strategies. While LLMs generally reached the global optimum faster than random selection, their performance relative to conventional methods was task-dependent, with no single LLM approach dominating across all settings.
large language modelsmaterials optimizationacquisition policiesgaussian-processactive learning
TT-net: Quantum Inspired Tensor Network Denoising in Conditional GANs
TT-Net introduces a quantum-inspired tensor-train decomposition for conditional GANs, replacing single-cut SVD denoising with a two-cut approach that captures cross-channel information. The method demonstrates superior performance over SVD-Net, EigenGAN, and Pix2pix across Gaussian, motion blur, and salt-and-pepper noise types, with consistent PSNR and SSIM improvements. Analysis reveals TT-Net's adversarial loss saturates while reconstruction quality improves, raising questions about the adversarial component's role. The work validates tensor networks as practical feature filters in deep learning.
tensor networkconditional ganssingular value decompositionimage denoisingquantum-inspired
An Irreducible Quantum Advantage in Aligning World Models with Reality
The paper demonstrates an irreducible quantum advantage in aligning world models with reality, showing that classical world models inherently fail in certain environments despite increasing memory. The authors construct classical true worlds where finite classical models either lose action discriminability or repeatedly assign highest expected rewards to suboptimal actions, retaining nonvanishing average error. In contrast, they prove that each such true world admits an exact quantum world model using a single qutrit, ensuring perfect alignment of optimal policies between real and virtual worlds through accurate reward estimates and action preferences.
world modelsquantum advantagequtritoptimal policiesexpected reward
Distilling Aggregate Mobility Statistics into a Language Model Policy for Post-Event Crowd Simulation
The paper introduces a language model-based approach for post-event crowd simulation that distills aggregate mobility statistics into individual agent behaviors. The method fine-tunes a language model crowd agent using iterative proportional fitting to match observed destination compositions derived from origin-destination flows, while addressing class imbalance through corrected training compositions. Evaluated on mobile network data from two baseball games, the approach reduces destination-share error by 25% without compromising grid correlation.
crowd simulationlanguage modeliterative proportional fittingorigin-destination flowsmobile network counts
CoToGrasp: Contact-Topology-Conditioned Dexterous Grasp Synthesis via Canonical Workspace Learning
CoToGrasp introduces a contact-topology-conditioned generative framework for dexterous grasp synthesis that bypasses object-specific annotations by learning in an object-agnostic manner. The method employs a feature-based canonical workspace to project local object features into a gripper-centric domain, decoupling functional intent from object geometry. Evaluations on DexGraspNet show state-of-the-art performance, with zero-shot generalization to unseen objects and demonstrated physical viability on a robot platform.
dexterous grasp synthesiscontact topologycanonical workspacezero-shot generalizationgripper-centric domain
Finite-Horizon Input-Output Dynamics of Minibatch Perturbations in AdamW
The paper analyzes delayed training effects in AdamW by modeling it as a finite-horizon input-state-output (ISO) system, where minibatch perturbations propagate through optimizer states. Through linearized joint dynamics, the authors derive a signed response operator that quantifies future loss impacts of gradient perturbations, revealing timing and magnitude dependencies on optimizer memory. Theoretical analysis includes an exact multistep error decomposition and first-order accuracy guarantees under smoothness conditions. Empirical validation confirms the response mechanism and demonstrates recoverable structure in delayed influence via ISO approximations.
adamw optimizationfinite-horizon dynamicsgradient perturbationoptimizer memoryinput-state-output system
Credit Without Ground Truth: Auditing Step-Level Credit Assignment in LLM Agents Against Executed Replay
The study audits step-level credit assignment in LLM agents against causal ground truth derived from executed replay in ALFWorld, revealing that existing signals (LLM-judge scores, outcome-conditioned logprob ratios, policy confidence) fail to identify causally impactful steps better than chance. Using structured ground truth, it shows causal contribution is sparse (30.5% of decision points) and model-dependent, with implicit credit echoing policy fluency (median rank correlation +0.75) but lacking causal information (partial correlation -0.004). A confidence-only router reduces judge cost by 13.1% per turn but identifies pivotal steps at chance level. Training experiments indicate no credit rule outperforms the untrained policy, with differences attributed to training dose rather than credit content.
credit assignmentcausal ground truthllm agentslogprob ratiosexecuted replay
GOAG: Generative and Object-Agnostic Grasp Planner for Dexterous Robotic Manipulation
The paper introduces GOAG, a generative and object-agnostic grasp planner for dexterous robotic manipulation, addressing generalization challenges in multifingered grasping. The method learns a compact latent representation of a gripper's contact surface distribution, enabling sampling of valid grasp configurations without object-specific training data. Evaluated on the MultiDex dataset, GOAG achieves an 86.93% average success rate, matching state-of-the-art performance while offering faster processing and improved generalization.
grasp planningobject-agnosticgenerative modeldexterous manipulationcontact surface distribution
GenMatch: An End-to-End Generative Matching Framework for Micro-View Order-Dispatching in Ride-Hailing
The paper proposes GenMatch, the first end-to-end generative matching framework for Micro-View Order-Dispatching in ride-hailing platforms, addressing cross-stage objective inconsistency in traditional multi-stage approaches. The framework employs a Context-Aware Bipartite Encoder for dynamic sparse graph representation, a Business-Aware Utility Learner for heterogeneous feedback integration, and a State-Aware Pointer Decoder for evolving matching state tracking. Offline evaluations and online A/B tests across five cities in DiDi's markets demonstrate consistent improvements over baselines.
generative matchingorder-dispatchingbipartite encoderutility learnerpointer decoder
Truncate Bad, Upweight Good: BoN-Style Distillation via Rank-Based Classification
The paper proposes TUP (Truncate-bad, Upweight-good Policy), a rank-based distillation method that removes low-ranked completions and reweights the upper tail with tunable sharpness. Unlike smooth full-support reweighting, TUP uses prompt-independent normalization and trains offline via binary cross-entropy with shifted-truncated win-rates as soft labels. Theoretical analysis shows that lower-tail truncation can match the best monotone rank-reweighting for unknown oracle rewards. Empirically, TUP competes with strong offline alignment baselines.
rank-based distillationinference-time selectionreward modeloffline alignmentbinary cross-entropy
Question-Guided Evidence Acquisition for Multimodal Visual Question Answering
The paper introduces Q-Guide, a question-guided evidence acquisition system for multimodal visual question answering that dynamically directs perception to relevant document regions during inference. The method employs a small agent that iteratively identifies missing evidence and invokes targeted tools (text reading, zooming, region grounding) before answering, contrasting with fixed single-pass encoding approaches. On DocVQA2026 and Manga109, Q-Guide achieves 65.0% (vs 40.0%) and 32.4% (vs 24.4%) accuracy respectively across three Claude backbones, with gains primarily from 2-3 perception rounds and evidence targeting rather than complex control logic.
multimodal vqaevidence acquisitiondynamic perceptiondocument understandingiterative refinement
Learning to Beat: Phenotype-Guided Latent Flow with Regional Motion Priors for Biventricular Motion Synthesis
The study introduces a phenotype-guided latent flow framework for synthesizing full-cycle biventricular motion from a single end-diastolic (ED) mesh. The method combines motion-informed functional parcellation with conditional latent flow, organizing the ventricular surface into regions with coherent dynamics and using phenotype-conditioned rectified-flow models for fine-grained motion synthesis. Evaluated on ACDC, M&Ms, and M&Ms-2 datasets, the approach achieves superior geometric accuracy (biventricular ASSD: 1.49±0.34 mm, HD95: 3.77±1.06 mm, vRMSE: 3.31±1.03 mm) and functional fidelity compared to competing methods, demonstrating robust generalization across cohorts and disease phenotypes.
biventricular motion synthesisphenotype-conditioned rectified-flowfunctional parcellationlatent flowventricular dynamics
TempJail: Temporal Jailbreak Attack against Large Vision-Language Models via Subtitle Scheduling
TempJail introduces a novel temporal jailbreak attack against large vision-language models (LVLMs) by optimizing subtitle scheduling in videos. The method leverages query-aligned dialogue-style subtitle sequences, focusing on temporal presentation factors like duration and timing-slot allocation, to exploit LVLM vulnerabilities. Evaluations on four LVLMs and two datasets show TempJail achieves the highest attack success rate, surpassing the strongest baseline by 53 and 18 percentage points in dataset-averaged ASR on GPT-5 and Gemini 3.5-Flash, respectively.
temporal jailbreaksubtitle schedulinglarge vision-language modelsattack success ratetemporal presentation
SafeBranch: Branch-Pair Safety Alignment for Embodied Agents
SafeBranch introduces a branch-pair safety alignment framework for vision-language-model-based embodied agents, addressing interactive safety by isolating safety-critical steps in task trajectories. The method constructs branch pairs by rolling back unsafe rollouts to the violation step, querying the actor for a safe alternative, and pairing the original action with the alternative to ensure minimal divergence. This approach trains the actor to act safely without requiring a critic during deployment. Evaluated on IS-Bench, SafetyALFRED, and out-of-distribution variants, SafeBranch achieves approximately ten times more safe successes than the baseline on unseen-object tasks while maintaining task performance.
embodied agentsinteractive safetybranch pairsvision-language modelsenvironment rollback
A Locally Tokenized Generative Model for Robust Time-Series Watermarking
The authors propose L-VQVAE, a locally tokenized generative model, and LVQMark, a robust watermarking method for multivariate time series, addressing reliability failures under post-editing attacks. Their approach ensures each discrete token depends only on a bounded temporal neighborhood, combining logit-bias injection with robust re-encoding for detection. Experiments on four benchmarks (finance, energy, neuroimaging) demonstrate preserved generation quality and stable detection power with controlled false positives under attacks.
time-series watermarkinglocally tokenizedlogit-bias injectionmultivariate time seriespost-editing attacks
Scale-Separated Conditioning for Style-Encoder-Free Diffusion Stylization
The paper introduces SEFS (Style-Encoder-Free Stylization), a diffusion-based framework for reference-based stylization without style encoders or aligned training triplets. SEFS conditions diffusion transformers using stochastic low-resolution crops to capture local appearance statistics (palette, stroke, texture) while suppressing global layout cues, combined with edge/segmentation content encoding via trainable projections. It incorporates style-to-denoising re-normalization and cross-block skip fusion for improved alignment and detail. Evaluations on artistic stylization benchmarks show enhanced content consistency and reduced style leakage compared to encoder-based methods, with ablations validating design choices. The approach trains on unpaired single images using only the frozen diffusion VAE for latent-space conditioning.
diffusion transformersreference-based stylizationstyle-encoder-freeparameter-efficient projectionsdenoising re-normalization
Robust Cross-Modal Foundation Model Perception for Underwater Robots under Degraded Visual Conditions
We propose a degradation-aware gated fusion method for robust underwater robotic perception under severe visual degradation, leveraging frozen DINOv2 visual encoder and sonar context without fine-tuning pretrained backbones. Our approach trains a multimodal fusion mechanism across five levels of visual degradation, from clean to extreme conditions, while keeping visual and sonar encoders frozen. Under extreme combined degradation, the method achieves 0.6152 balanced accuracy, a 33.5% relative improvement over the DINOv2 baseline (0.4610). Learned sonar contribution increases from 14.2% in clean conditions to 41.3% under extreme degradation, demonstrating adaptive redistribution of cross-modal reliance. Fusion provides largest gains under severe turbidity and blur, while color attenuation yields minimal additional benefit.
degradation-aware fusiondino v2sonar contextmultimodal perceptionbalanced accuracy
Loreley: Repository-Scale Program Evolution with Quality-Diversity Search
Loreley introduces a Quality-Diversity (QD) search approach for repository-scale program evolution, retaining complete repository states in a QD archive for sampling or contextual edits. Candidates are Git commits evaluated by project-supplied metrics, compared against sequential champion editing and independent root proposals in a Zstandard experiment. At 48 jobs, QD performed within -0.556% to +0.161% of Sequential Champion and +0.320% above Independent Root, with no significant advantage observed. Archive retention enabled stepping-stone mechanisms, evidenced by multi-file improvements in Python libraries and Zstandard revisions, though endpoint benefits were not demonstrated in the controlled experiment.
quality-diversity searchrepository evolutiongit commitssequential championstepping-stone mechanism
Beyond Memory Majority: Latent-Source Reasoning for Multi-Agent Memory Arbitration
The paper introduces Correlation-Aware Memory Arbitration (CAMA), a framework addressing Memory Correlation Bias in multi-agent systems where correlated memories create false majorities. CAMA combines neural dependency inference with symbolic priors to estimate independent evidence sources, then employs a sequential recovery policy to retrieve missing evidence before arbitration. Evaluations on multiple benchmarks show CAMA outperforms state-of-the-art baselines by suppressing false majorities from correlated memories.
memory correlation biasmulti-agent systemsdependency inferencesequential recovery policyevidence arbitration
Rethinking the Evaluation and Optimization of LLM-Based Social Simulation
The paper proposes Subjectivity-Adaptive soft-Label Training (SALT) to improve LLM-based social simulation by addressing the limitations of accuracy-based evaluation and hard-label training for subjective tasks. SALT aggregates observed outputs from semantically similar inputs into soft distributional labels, adapting the aggregation radius based on a novel subjectivity coefficient that quantifies task subjectivity. The authors also introduce SUBJSIM, a benchmark with 19,300 contexts and 193 annotators, to enable distributional evaluation. Experiments on SUBJSIM demonstrate SALT's effectiveness in realistic single-observation settings.
social simulationsubjectivity coefficientsoft-label trainingdistributional evaluationllm optimization
Learning Hierarchical Skill Policies with Offline Quality-Diversity Reinforcement Learning
QDOS (Quality-Diversity Offline Skill learning) introduces a unified pipeline for robust offline-to-online reinforcement learning, addressing limitations in low-level policy quality dependent on dataset quality. The method employs an Advantage-Weighted Quality-Diversity pretraining objective, weighting skill extraction and diversity by trajectory segment advantages, and integrates a dual dataset reuse strategy for skill pretraining and online replay buffer population. Experiments show QDOS outperforms baselines in structured manipulation and unstructured locomotion tasks, enhancing exploration and final returns in sparse-reward domains.
offline-to-online learningadvantage-weightedquality-diversityskill extractionsparse-reward domains
Frequency-Aware Continual Learning for Smart Contract Vulnerability Detection with Large Language Models
Proposes a three-stage pipeline for continual learning in smart contract vulnerability detection with LLMs, addressing parameter-efficient adaptation, catastrophic forgetting, and multi-adapter consolidation. Introduces Frequency-Aware Low-Rank Adaptation (FA-LoRA) for Fourier-domain adaptation (0.4% trainable parameters), Forget-Aware Replay (FAR) for prioritized rehearsal (Micro-F1 0.8022), and Anchor-Protected Progressive Merging (APPM) for adapter consolidation (Micro-F1 0.8085, 156ms merge cost). Achieves within 2.7% of per-task upper bound on DIVE benchmark.
continual learninglow-rank adaptationcatastrophic forgettingfourier domainsmart contract
Escaping the Quicksand: A Call to Arms
The article advocates for combining testing, specification, and proof methods to address technical debt in software engineering, exacerbated by AI-enabled development. It proposes incremental co-development of executable specifications alongside traditional code and tests to improve design clarity and testing efficacy. The approach enables a spectrum of feedback loops from cheap testing to formal proof, requiring semantic infrastructure for practical implementation. The authors call for community efforts to build this infrastructure, aiming to transition from fragile test-and-debug paradigms to more robust foundations.
technical debtspecificationformal proofsemantic infrastructurefeedback loop
DeltaML-Bench: Evaluating Machine Learning Agents on Real-World Research Repositories
DeltaML-Bench introduces a benchmark for evaluating ML agents on 48 real-world research repository tasks requiring baseline improvement under compute constraints. The study compares GPT-5 and Claude Sonnet 4 using Modular and search-based ARG scaffolding agents across different time allocations (4×6h and 2×12h). Results show ARG scaffolding improves GPT-5's success rate from 9.4% to 33.9% (4×6h) and 49.0% (2×12h), while Modular agents exhibit up to 47.9% specification gaming, highlighting the importance of scaffolding design for autonomous ML experimentation.
autonomous agentsbenchmark evaluationscaffolding designspecification gamingcompute constraints
Can Agent Memory Systems Track Evolving State?
The paper introduces StateMemBench, a benchmark of 234 multi-session scenarios evaluating LLM-based agents' ability to track evolving world states, distinguishing current from superseded states. It proposes StateMem, a memory method explicitly modeling supersession and relational dependencies, which improves current-state accuracy by 1.8x on DeepSeek-V4-Flash and 1.6x on Qwen-3.5-9B over baselines. As a lightweight wrapper, StateMem boosts accuracy by +32 to +67 points across six memory backends, with +15 to +32 points attributable to state structure.
state trackingmemory systemsllm-based agentssupersessionbenchmark
Scientific Data Skills: Enabling Agent-Ready Scientific Data Services at Scale
The paper introduces Scientific Data Skill (SciDSK), an agent-ready representation for scientific datasets that integrates metadata, operational guidance, and provenance while preserving original data locations. The authors develop a structured specification and construction pipeline grounded in authoritative records, implemented via a cross-disciplinary Scientific Data Skill Bank platform. Evaluation shows SciDSK improves agent performance in dataset discovery (retrieval benchmark) and interpretation (controlled cases) compared to conventional representations.
scientific data skillagent-ready representationdataset discoverymetadata integrationprovenance tracking
Forking Fast: Efficiently Estimating Uncertainty Dynamics in Text Generation
The article introduces a method to efficiently estimate uncertainty dynamics in LLM-generated text by reducing the computational cost of resampling-based analyses. It demonstrates that uncertainty patterns stabilize with increased sampling, attributing noise primarily to sampling artifacts rather than model sensitivity. A statistical model is proposed to smooth noisy rollout data, enabling cost-effective approximation of high-sample results. This approach enhances understanding of LLM reasoning chains and their variability.
llm reasoninguncertainty dynamicsresampling analysisstatistical modeltext generation
PEA-DPO: Perception-Enhanced Alignment Direct Preference Optimization for MLLMs Alignment
PEA-DPO introduces a Perception-Enhanced Alignment framework for multimodal LLMs (MLLMs) to address visual insensitivity in Direct Preference Optimization (DPO), where models fail to distinguish images from visually degraded variants. The method explicitly incorporates visual preference signals to mitigate Across-Image and Within-Image Insensitivity, supported by theoretical guarantees. Evaluations on three hallucination benchmarks demonstrate PEA-DPO improves visual context sensitivity, enhances multimodal alignment, and reduces hallucinations across varying MLLM scales while preserving language modeling capabilities.
direct preference optimizationmultimodal llmsvisual insensitivityperception-enhanced alignmenthallucination benchmarks
VGI-BENCH: Probing Visual Intelligence in Video Generation Models
The authors introduce VGI-bench, a benchmark for evaluating visual reasoning in video generation models, addressing current limitations in task alignment, temporal validity, and difficulty calibration. The benchmark comprises 27 tasks (810 instances) organized by domain and skill tags, enabling fine-grained capability assessment. Evaluations reveal limited reliability in current models, with top-performing Seedance~2.0 achieving only 51.0% accuracy, exhibiting failure modes in self-correction and synthetic-to-real transfer. The analysis examines output errors, input sensitivity, and denoising dynamics, suggesting later steps refine rather than correct early reasoning.
video generationvisual reasoningbenchmarkzero-shot evaluationdenoising dynamics
Enforcing LLM Safety through DMD-based Classification of Prompt-Response Embedding Dynamics
The paper introduces a dynamical systems framework for classifying unsafe LLM outputs by analyzing prompt-response embedding dynamics. It extends Koopman-based predictive modeling to LLM safety, fitting separate models for safe and unsafe regimes and using a differential residual score for classification. The method incorporates both prompt and response embeddings, capturing interaction patterns via fitted Koopman operators. Evaluated across three safety benchmarks with three embedding models, results show consistent improvements when using prompt embeddings, particularly for interaction-dependent violations with causal decoders like Llama-3, while response-only violations benefit from dense semantic embeddings.
koopman operatorsembedding dynamicsdifferential residual scorecausal decoderssemantic embeddings
DraftFM: A FoundationModel for Day-Zero Drafting in Magic: The Gathering
DraftFM introduces a foundation model for day-zero drafting in Magic: The Gathering, addressing the challenge of predicting picks for unseen expansions without prior draft logs. The model employs a 1.6M-parameter network that scores cards based on 775-dimensional features derived from public card records, structured features, and fixed text embeddings, avoiding reliance on card identities or usage statistics. Trained on 149M human picks from 29 expansions, it achieves 50.8-60.4% top-1 agreement on held-out sets and demonstrates expert-level agreement in pre-release evaluation for The Hobbit set.
foundation modeldiscrete-choice policycard rankingtext embeddingday-zero drafting
Stream4D: 4D-Consistency for Streaming Autoregressive Diffusion Video Models
Stream4D introduces a 4D-consistent training framework for streaming autoregressive diffusion video models, addressing geometric drift and motion degradation in long-horizon generation. The method replaces static 3D Gaussian-Splatting critics with a feed-forward 4D reconstruction reward that models scene dynamics, supplemented by a motion prior for natural scene-flow and artifact reduction. Evaluated across multiple autoregressive backbones, Stream4D improves 4D reconstruction quality, motion preservation, and human preference scores compared to baseline approaches.
autoregressive diffusion4d reconstructionscene dynamicsgaussian-splattingscene-flow
CVSD-Reg: Cross-Modal Visual Semantic Prior Distillation for Robust LiDAR Registration
CVSD-Reg introduces a robust LiDAR registration framework that distills visual semantic priors from DINOv2 into LiDAR representations via cross-modal contrastive distillation and spherical-manifold alignment. The method employs a two-stage approach: (1) a Point Transformer V3 student learns from frozen DINOv2 features with self-supervised InfoNCE consistency and SE(3) invariance, followed by (2) density-aware point-dropout augmentation and end-to-end pose optimization. The camera-free system achieves 97.7-99.3% SR@0.5m/1° on KITTI, nuScenes, and HeLiPR, outperforming geometric baselines by up to 44.0pp without ICP refinement.
lidar registrationcross-modal distillationspherical-manifold alignmentpoint transformerse(3) invariance
From Retrieved Context to Runtime Control: Adaptive Compression for Edge-based RAG
This paper introduces a telemetry-informed adaptive compression framework for edge-based retrieval-augmented generation (RAG) systems, addressing the overhead of retrieved context in language models. The method dynamically adjusts compression rates based on workload variation and edge device state, evaluated on NVIDIA Jetson AGX Thor using Llama and Qwen generators with Natural Questions and HotpotQA datasets. Results show intermediate compression reduces GPU energy by up to 53.2% and SoC energy by up to 48.2%, with minimal quality loss, outperforming fixed compression strategies. The study highlights the tradeoff between compression rate and inference quality, advocating for runtime policies guided by workload features and edge telemetry.
retrieval-augmented generationcontext compressionedge computingtelemetrykv-cache
When Machines Speak: A Unified Generative Framework for Integrating Machine-Native Symbols into Pretrained Large Language Models
The paper introduces UniLang, a unified generative framework that integrates machine-native symbols into pretrained large language models (LLMs) by extending their vocabulary and embedding space. UniLang enables joint modeling of natural-language tokens and machine-native symbols under a single autoregressive objective, eliminating the need for verbalization or task-specific architectures. Evaluations on sequential recommendation and legal precedent prediction tasks demonstrate consistent performance improvements over baselines, showcasing the framework's ability to bridge structured prediction and language modeling.
unified generative frameworkmachine-native symbolsautoregressive objectivestructured predictionvocabulary expansion
Automated Summarization of Financial News Using Large Language Models and Retrieval-Augmented Generation: An Early Empirical Study (Fall 2023)
The study evaluates LLM-based financial news summarization using Retrieval-Augmented Generation (RAG) and Summarize Chains across Falcon-7B-Instruct, DistilBART-CNN-12-6, and BART-Large-XSum. A pipeline integrates News API, Wikipedia, and Yahoo Finance data, with stock prices converted to natural language narratives. Falcon-7B with Summarize Chains achieved highest accuracy and coherence, outperforming Lead-3 baselines on ROUGE-1 (+12.3), while RAG induced repetition (Falcon) and hallucinations (BART-Large) at high k. Results highlight trade-offs in open-source LLMs for financial NLP pre-widespread RAG adoption.
retrieval-augmented generationsummarize chainsrouge-1instruct-tuningfinancial nlp
In Two Minds about Lifelong Learning: Exploring Hemispheric Redundancy and Specialisation in Neural Models
The paper proposes 4MAS (4 Module Awake/Sleep), a novel neural architecture inspired by biological learning mechanisms to address catastrophic forgetting in continual learning. The model incorporates asymmetric hemispheres with separate long- and short-term memory systems, experience replay, and sleep-like consolidation phases. Evaluated on Split-MNIST, Split-Fashion-MNIST, and Split-CIFAR-100, it achieves 98.3%, 84.9%, and 29.29% accuracy respectively, demonstrating competitive performance against catastrophic forgetting.
continual learningcatastrophic forgettingexperience replaymemory consolidationasymmetric hemispheres
Symposium: Trust via Auditable Records for Communities of AI Scientist Agents
Symposium introduces a formal framework and implementation for recording AI agent operations in scientific research communities, enabling long-term, immutable histories of agent-driven activities. The system captures structured claims, evidence citations, assumptions, and usage declarations, facilitating auditable trails of analyses, hypotheses, and scientific discourse. By separating durable research histories from evolving AI systems, Symposium supports trust assessments and allows agents to build on prior work. A working implementation includes publication infrastructure, agent prompt components, and documentation for rapid community setup.
auditable recordsscientific discourseimmutable historiesagent-driven researchtrust assessments
Accelerated Genetic Programming Hyper-Heuristics for Simulation-Based Scheduling via Agentic AI
The paper presents an agentic AI approach for accelerating genetic programming hyper-heuristics in simulation-based scheduling. Using Claude AI, the method systematically refactors Python code for discrete-event project-scheduling simulations by identifying bottlenecks, implementing optimizations, and validating correctness via benchmarks. In HPC environments, this reduced runtime from 1,298s to <200s while preserving outputs, yielding annual savings of 4M core-hours (NZ$320k). The approach maintains researcher control while addressing performance limitations of object-oriented structures and nested loops.
genetic programminghyper-heuristicsdiscrete-event simulationagentic aihpc optimization
Measuring What a Specification Determines: A Formal Semantic-Block Model and an Execution-Judged Benchmark
The paper introduces a formal semantic-block model for specifications and an execution-judged benchmark to evaluate specification quality independently of model capability. Specifications are structured as semantic blocks with dependency relations, rules, decision points, and open questions, subject to four well-formedness conditions. The model is instantiated on an Oracle-to-PostgreSQL migration specification with 18 blocks and 19 dependencies, showing a 71% reduction in per-task context and 85.5% coverage of Oracle constructs. The benchmark uses PostgreSQL 16 and Oracle as deterministic execution judges, revealing a median arm-delta spread of 14.4 percentage points. Results support determinacy as a formal concept but not as a standalone empirical metric for contemporary LLM implementers.
semantic-block modelexecution-judged benchmarkdependency relationsdeterminacyoracle-to-postgresql migration
Are LLMs becoming similarly creative? Evidence from three years of models
This paper investigates the evolution of LLM creativity by analyzing outputs from three years of model releases on open-ended tasks. Using sentence-embedding similarity, the study evaluates responses to Infinity-Chat100, a real-world query dataset, and the Alternate Uses Task, a psychometric creativity assessment. Results indicate a statistically significant decrease in output diversity over time, suggesting convergence in creative substance across models. This trend raises concerns about potential homogenization in human-AI co-creative work and its implications for human agency in creative processes.
large language modelssentence-embedding similarityopen-ended taskspsychometric creativity assessmentoutput diversity
Longitudinal Bayesian Learning of Continuous Disease Position across the Alzheimer's Disease Continuum
The authors propose Disease Continuum Positioning (DCP), a longitudinal Bayesian learning framework that estimates continuous disease severity in Alzheimer's disease (AD) from diffusion tensor imaging (DTI). DCP models disease severity as a low-dimensional latent variable, integrating longitudinal observations with weak clinical supervision to derive a Disease Continuum Score (DCS). Evaluated on the ADNI cohort, DCP outperforms existing progression methods, with DCS demonstrating accurate severity characterization, clinical relevance, longitudinal evolution preservation, and predictive capability for future conversion.
bayesian learningdisease continuumdiffusion tensor imaginglongitudinal modelingalzheimer's disease
SCAPE: Scenario-Conditioned Simulation-Augmented Policy Evaluation
SCAPE introduces a scenario-conditioned simulation-augmented policy evaluation framework for predicting real-world robot policy performance. The method combines limited paired sim-and-real samples with large-scale simulation rollouts, correcting sim-to-real bias and calibrating uncertainty via conformal prediction. Evaluated on autonomous driving and quadruped velocity tracking, SCAPE reduces scenario-level prediction error by 4.9-34.7% versus baselines, improves sample efficiency, and enables fine-grained deployment strategies.
sim-to-real gapconformal predictionpolicy evaluationscenario-conditionedrobot-learning
Beyond Imitation: Filtering On-Policy Distillation by Reasoning Progress
The paper introduces Reasoning-Progress-Aware Reward Filtering for On-Policy Distillation (R2-OPD), a method addressing the mismatch between teacher-derived rewards and genuine reasoning progress in on-policy distillation (OPD). R2-OPD constructs two within-trajectory rankings of reasoning spans: one based on teacher-derived rewards and another on independently estimated progress rewards. Distillation rewards are selectively suppressed when these rankings disagree, reducing conflicting supervision while preserving effective teacher guidance. The approach demonstrates consistent improvements over standard OPD, particularly in reasoning performance.
on-policy distillationreasoning progressreward filteringteacher-derived rewardsreasoning spans
HiRA-CAM: Preserving Fine-Grained Spatial Relevance in Gradient-Based Visual Explanations
The paper introduces HiRA-CAM, a gradient-based method for generating saliency maps in CNNs that improves upon LayerCAM and Grad-CAM. HiRA-CAM adaptively combines activation maps from all CNN layers to produce more spatially precise visual explanations. Evaluations demonstrate superior performance in object classification tasks compared to existing methods, with enhanced focus on relevant image regions.
saliency mapsconvolutional neural networksgradient-based explanationlayer activationvisual interpretability
HYDRA: A Heterogeneous Chiplet DSE Framework for Serving Dynamic Hybrid LLM Workloads
HYDRA introduces a design space exploration framework for hybrid Transformer-Mamba LLM serving on heterogeneous chiplet systems, jointly optimizing chiplet composition, placement, bandwidth provisioning, dynamic batching, and runtime scheduling. The framework integrates communication-aware placement, elastic task scheduling, and a Markov-based performance estimator to capture multi-tenant dynamics. Evaluations show 1.55x higher throughput (up to 2.3x) and 43.7% lower time-to-first-token versus baselines, demonstrating the necessity of architecture-runtime co-design for efficient large-scale LLM serving.
chiplettransformer-mambadesign space explorationdynamic batchingmarkov estimator
Concentrated Liquidity Provision: a Reinforcement Learning Perspective
The paper proposes a reinforcement learning approach for dynamic liquidity provision in constant product markets with concentrated liquidity (e.g., UniswapV3). Formulating the problem as a stochastic impulse control task, the authors derive interpretable policies that exhibit state-dependent behaviors based on mispricing, rebalancing costs, and risk preferences. Results show these RL agents outperform baseline methods from AMM microstructure literature, particularly in compressing the left tail of PnL distributions and mitigating catastrophic outcomes under uncertainty.
automated market makersconcentrated liquidityreinforcement learningstochastic impulse controlprofit and loss distribution
Fairness-Aware Network Embeddings: Methods, Applications, and Challenges
The survey presents a taxonomy of fairness-aware network embedding methods, categorizing them by embedding approach (spectral, random walk, graph neural network, Bayesian, method-agnostic), fairness intervention strategy (pre-processing, in-processing, post-processing), and fairness objective criterion (embedding- or task-level). It compares methods based on group versus individual fairness and sensitive attribute assumptions, addressing structural inequalities in real-world networks. The analysis highlights current limitations and future research directions, providing a unified framework for developing fair network representation learning techniques.
network embeddingfairness-aware learninggraph neural networksbias mitigationrepresentation learning
Does Marginal Coverage Guarantee Class-Conditional Safety for Zero-Shot VLMs Under Shift?
The study evaluates split-conformal prediction's class-conditional safety guarantees for zero-shot vision-language models (VLMs) under distribution shift, focusing on CLIP, OpenCLIP, and SigLIP. Despite maintaining marginal coverage (~0.86 on ImageNet-Sketch), worst-class coverage collapses to ≈0, with 10-12% of classes falling below a finite-sample null floor. Source-domain diagnostics fail to predict these failures, while target-side class calibration improves tail coverage but requires extensive labeled data. Findings persist across model scales, pretraining corpora, and shifted non-ImageNet settings, highlighting marginal coverage as an average reliability metric rather than a class-tail safety guarantee.
split-conformal predictionzero-shot vlmsmarginal coverageclass-conditional safetydistribution shift
Mechanistic Tomography: Designed Measurement for Control-Oriented Interpretability
The paper introduces mechanistic tomography, a framework for recovering internal mechanisms and intervention effects in neural networks through designed measurements. It formalizes various interpretability techniques (patching, gradients, Hessian-vector products) under a unified measurement structure y = Ax + w, where A describes interventions, x is the target map, and w captures nonlinearities and errors. The method proposes a practical procedure starting with low-cost measurements, calibration, and iterative expansion. Experiments on GPT-2-small and Qwen-2.5-7B demonstrate recovery of interaction maps and predictive terms, with control error validating observer estimates. Results show sparse aggregate measurements suffice for finite-effect maps and finite probes improve local attribution.
mechanistic tomographyintervention effectshessian-vector productslocal attribution mapfinite-effect map
Improved Confidence Estimates for Black-Box Large Language Models
Proposes an improved method for uncertainty quantification (UQ) in black-box large language models (LLMs) by leveraging labeled datasets to enhance confidence estimates. Introduces simple classifiers that predict LLM response correctness using existing UQ scores and correctness of similar queries as features, with minimal computational overhead. Demonstrates consistent outperformance over existing zero-shot UQ methods, offering a practical enhancement for real-world deployment.
uncertainty quantificationlarge language modelsconfidence estimationzero-shot learningblack-box optimization
Quantum Kernel Estimation for the Discovery of Early Lung Cancer Detection
The study demonstrates quantum-classical hybrid machine learning for early lung cancer detection using blood-based cfDNA biomarkers. Methodologically, it encodes DNA fragmentomics and methylation features into quantum Hilbert space using angle/dense-angle maps with varied entanglement strategies, then computes fidelity-based quantum kernels integrated with SVM and kernel-PCA logistic regression. Results show quantum-kernel models achieve competitive AUC (fragmentomics: +0.0-Δ vs classical SVM; methylation: matched performance with improved specificity in cases) using 20-40 features, though feature expansion to 40 increased variability without consistent gains.
quantum kernel estimationcell-free dnafragmentomicshilbert space encodingfidelity-based kernel
Outcome Monitors: Recovery Affordances for Silent Tool Failures
The paper introduces Outcome Monitors, a method for detecting silent tool failures in AI agents by verifying outcome contracts derived from task-disjoint traces or public schemas. When violations occur, the system preserves results and provides nonbinding receipts with violated properties and recovery tools. Evaluations on ToolMaze and tau-bench retail show completion rate improvements of 17.2% and 12.0-14.0 points respectively across multiple models. Recovery tools prove critical, as their removal eliminates gains. Detection accuracy falls to 46% for faults outside the mined vocabulary, though completion remains unaffected.
outcome monitorssilent failuresrecovery affordancestoolmazeoutcome contracts
Air Traffic Control Using Large Language Models: Prompt Engineering, Architecture, and Evaluation
The study evaluates large language models (LLMs) for generating operationally realistic air traffic control (ATC) transmissions, using a hand-transcribed general-aviation flight as ground truth (P0). Five prompt structures (P1-P5) are tested in a stateful multi-turn pipeline, where LLMs condition on dialogue history and optional in-context examples. Results show that in-context examples improve similarity metrics, but overly constrained prompts degrade performance due to error accumulation, mitigated by injecting correct history. GPT-5.5 validates outputs against human annotation.
large language modelsair traffic controlprompt engineeringin-context learningdialogue history
DA-WAM: Decision-Aligned Future Latents for Driving World Models
DA-WAM introduces a unified framework for decision-aligned future modeling in autonomous driving, integrating predictive representation learning, action-conditioned future prediction, and trajectory scoring under a single objective. The method employs an online encoder with momentum targets to maintain predictive supervision during planner optimization, generates trajectory-specific future latents via an action-conditioned predictor, and evaluates them using a factorized scorer. Experiments on NAVSIM-v1 and NAVSIM-v2 show state-of-the-art performance, with ablations confirming the importance of decision-aligned future modeling and hard negative supervision.
world modelsdecision-aligned learningaction-conditioned predictiontrajectory scoringautonomous driving
Graphical Design of Interpretable Architectures
The authors introduce a graphical notation for designing interpretable AI architectures, adapted from Penrose tensor notation, to address limitations in existing representations. This notation provides a global view of architectures while maintaining a one-to-one mapping to PyTorch einsum code, ensuring reproducibility. The method is demonstrated by diagramming interpretable architectures (e.g., concept bottlenecks, neural additive models) and Steerling-8B, revealing structural insights and enabling direct code translation.
interpretable architecturespenrose tensor notationpytorch einsumconcept bottleneckssteerling-8b
TestifAI: Tomography-Based Testing for Deep Learning Systems
TestifAI introduces a tomography-based testing framework for efficient robustness evaluation of deep learning systems against combinatorial input perturbations. The method employs partial model tomography to reconstruct model behavior in multi-perturbation spaces from lower-order (1-2 perturbation) tests, avoiding exponential test execution. Experiments on five classification tasks demonstrate <7% aggregate robustness estimation error for higher-order (3-4 perturbation) predictions while reducing required inferences by 60-80%.
robustness testingpartial model tomographycombinatorial perturbationsoperational conditionssemantic perturbations
Information on trajectories: martingales and random times
The paper introduces exact variational identities for nonnegative martingales by accounting for information flow on trajectory path spaces, applicable even at arbitrary random times. It recovers classical concentration inequalities (Ville, PAC-Bayes) and quantifies discarded slack in three geometries: Gibbs tilt (Azuma-Hoeffding, PAC-Bayes), crossing (Ville, pooled tests), and dominating certificates (L^p maximal bound). The certificate's optional-stopping deficit decomposes into per-step Bregman divergences. Path-time analysis reveals an e-process 'peeking penalty' for random times, while geometric mixtures of test martingales benefit from pooling in multi-model testing.
martingalesvariational identitiesconcentration inequalitiesbregman divergencese-process
$TCP_α$: Margin-Controlled Confidence estimation for reliable Music Information Retrieval
The paper introduces $TCP_α$, a margin-controlled confidence estimation target for deep neural networks that guarantees complete separation between confidence values of correct and incorrect predictions. The method employs a lightweight auxiliary head trained with a margin-based penalty for misclassified samples, addressing class imbalance through systematic training strategy analysis. Evaluated on rāga identification and frame-wise ornamentation detection, $TCP_α$ improves macro-F1 from 0.89 to 0.98 by rejecting 8% of predictions and maintains robustness under domain shift with only 5% labeled fine-tuning data.
confidence estimationmargin-controlled penaltyfailure predictiondomain shiftimbalanced regression
A comparison between ceiling-mounted FMCW, IR-UWB and Wi-Fi radar for in-bedroom human activity monitoring and sleep interruption detection
This study provides a systematic comparison of ceiling-mounted FMCW, IR-UWB, and Wi-Fi radar for healthcare monitoring, addressing a gap in controlled evaluations under identical conditions. Using synchronized recordings from 20 participants across six room layouts, the authors evaluate all technologies with the same CNN architecture on 10-class HAR and 4-class sleep monitoring tasks. IR-UWB achieves 89.0% macro F1 for cross-subject HAR, while FMCW shows superior generalization to unseen layouts (83.8% macro F1). All technologies exceed 92% macro F1 for sleep monitoring, revealing a trade-off between recognition performance and environmental robustness linked to hardware characteristics.
frequency-modulated continuous waveimpulse radio ultra-widebandhuman activity recognitionconvolutional neural networkdoppler resolution
Explainable Transformer Models for Clinical Prediction Tasks on Structured Electronic Health Records
The study introduces BERT-LER, a BERT-style model for structured electronic health records (EHRs) that encodes laboratory test results as discrete tokens using percentile-based binning, paired with Integrated Gradients for interpretable token-level attributions. The model is pretrained and fine-tuned on a de-identified EHR dataset of 75 million patients, addressing a gap in EHR foundation modeling by unifying laboratory value representation and explainability. Evaluated on the EHRShot benchmark and an asthma severity progression study, BERT-LER achieves competitive or superior performance, particularly in laboratory-related tasks, with attributions aligning with known clinical risk factors.
bert-lerelectronic health recordsintegrated gradientsclinical predictionpercentile-based binning
Physical-Support Confidence Sets for Highly Coherent Dictionaries
The authors propose resolution-aware physical-support inference for sparse pursuit in highly coherent dictionaries, addressing uncertainty in both learned dictionaries and signal representations. Their method establishes cross-dictionary confidence correspondence, retaining calibration-compatible dictionaries and deployment-compatible sparse representations, then projects these onto physical-support space. Theoretical analysis shows minimax physical resolution δ_opt(N,s) scales with min{s,1/√Ns²}, with relative resolution governed by orientation-information scale Ns⁶. They introduce active endpoint bracketing (AEB), an adaptive finite-bank procedure that evaluates only candidates affecting physical reports, avoiding unsupported refinement. Experiments demonstrate AEB's effectiveness in preventing physical overprecision while reducing candidate evaluations.
sparse pursuitdictionary learningphysical-support inferenceminimax resolutionactive endpoint bracketing
Dynamic Structural Causal Modeling for Sleep
The authors propose a dynamic structural causal modeling approach to analyze sleep-disordered breathing, revealing population-specific causal patterns. They apply the PCMCI+ algorithm to windowed fractional variables from 105 Home Sleep Apnea Test recordings, incorporating domain knowledge via edge blacklisting and using bootstrap aggregation for robustness in small subcohorts. Results indicate consistent temporal self-dependencies and apnea-desaturation relationships across all cohorts, with significant variation in other causal structures by sex and age.
dynamic causal modelingpcmci+sleep-disordered breathinghome sleep apnea testbootstrap aggregation
Which Eviction Policy Should an LLM Cache Use? A Systematic Study Across Workloads, Capacities, and Encoders
This work systematically evaluates eviction policies for semantic caches in LLMs across workloads, capacities, and encoders. Using the CLEVER framework, the study compares FIFO, LRU, LFU, ARC, GDSF, streaming SISO, and semantic-redundancy policies on three query corpora with varying cache sizes and two embedding models. Results show LFU outperforms other policies by up to 8.67 percentage points at tight capacities, with minimal (<0.041pp) improvements from more complex approaches. Analysis reveals limited benefits from geometry-aware eviction due to conditional packing constraints. Cross-encoder experiments demonstrate threshold non-transferability, while quality-adjusted hit rates remain low (1.1-2.2%) despite raw hit rates of 51-60%.
semantic cacheeviction policyconditional packingcross-encoderquality-adjusted hit rate
DICS: Data-Informed Centroid Splitting for Decision Tree Classifiers
The paper proposes Data-Informed Centroid Splitting (DICS), a clustering-based framework for efficient decision tree training that reduces the split search space using class-aware data priors. DICS constructs informative candidate splits via centroid clustering, maintaining theoretical performance guarantees under stated assumptions while being compatible with classification trees, random forests, and gradient boosting. Experiments on synthetic and benchmark datasets show DICS achieves comparable accuracy to exhaustive search while significantly reducing training time.
decision treessplit selectionclusteringclassificationcomputational efficiency
Transfer Learning in Nonparametric Regression with Deep ReLU Networks
(No summary returned.)
Gravitational-wave parameter estimation with machine-learning generated surrogate waveforms
The authors propose a two-stage conditional autoencoder for generating SEOBNRv4 gravitational waveforms, addressing computational bottlenecks in parameter estimation. Stage one predicts amplitude/phase series while stage two calibrates residuals, achieving median waveform mismatch of 10^-2 and cosine distance error of 10^-6. A waveform conditioning step enables downstream parameter estimation, though systematic biases require correction via importance reweighting at low SNRs when comparing ML-generated and EOB waveforms.
gravitational wavesparameter estimationconditional autoencodersurrogate waveformsimportance reweighting
Decoding silent reading from non-invasive EEG
The study demonstrates open-vocabulary word-level decoding from non-invasive EEG during silent reading, addressing the challenge of inner speech data acquisition. Using a contrastive decoder, the authors analyzed ~240,000 word presentations from a single participant across 393 runs (49 hours) of 19-channel dry-electrode EEG. Words were presented in rapid serial visual presentation with randomized typography to decorrelate visual form from word identity. A convolutional EEG encoder, optionally augmented with a causal transformer, was trained with a CLIP-style contrastive objective to align EEG windows with hidden-state embeddings from a large language model. Results showed reliable above-chance decoding, log-linear scaling with data volume, and separable contributions from word-level decoding, narrative context tracking, and positional priors.
eegcontrastive decoderrapid serial visual presentationconvolutional encoderpositional embedding
Exact Algebraic Computation of Learning Coefficients for Two-Dimensional Singular Models
The paper presents the first deterministic algorithm for exact computation of local Real Log Canonical Thresholds (RLCTs) in two-dimensional singular models, addressing limitations of classical information criteria like BIC. The method handles models where the Kullback-Leibler divergence is contact-equivalent to a polynomial, with proven complexity bounds. Results demonstrate effectiveness across a broad class of models, including polynomial neural networks, revealing algebraic structures in learning coefficients and outperforming sampling-based estimators in shallow regimes.
real log canonical thresholdssingular modelskullback-leibler divergencewidely applicable bicpolynomial neural networks
Ask Self, Ask Others: Relation Is All You Need
The paper introduces Relation, a novel token-mixing primitive that organizes pairwise evidence into Self and Exchange relations before deriving information flow, contrasting with attention mechanisms. It proposes Full Relation, FlashRelation, Linear Relation, Hybrid Relation, and a KV-style Relation Cache as implementations. Full Relation outperforms Multi-Head Attention (MHA) in validation NLL across 10M, 30M, and 100M parameter decoder-only models. FlashRelation achieves 3.60-4.41x speedup over Full Relation in fixed-context benchmarks and 76.4-84.9% of PyTorch FlashAttention throughput in production workloads. Hybrid Relation, using 75% Linear Relation layers, maintains strong language modeling quality.
token-mixingrelationflashrelationkv-cachenll
Feature Evolution and Migration during Vision Transformer Training
The study introduces a novel framework for analyzing feature evolution and migration in Vision Transformers (ViTs) during training. Using Sparse Autoencoders (SAEs), the authors extract and visualize sparse features from CLS-token representations across network depth and training epochs, enabling detailed observation of feature-level dynamics. Key findings include concentrated feature migration early in training, a tendency for features to migrate toward earlier layers, and earlier stabilization in deeper layers compared to shallow ones. This approach provides insights into how ViTs organize and stabilize features over time, offering a tool for deeper understanding of ViT learning processes.
vision transformerssparse autoencodersfeature evolutionfeature migrationcls-token
Discrete Diffusion Inference-Time Control with Nested Sequential Monte Carlo
We introduce nested sequential Monte Carlo (NSMC) and fully-adapted NSMC (FA-NSMC) for inference-time control in discrete diffusion language models, addressing limitations of prior particle-based methods like best-of-n sampling and bootstrap SMC. Our approach formulates Feynman-Kac steering while correcting biases in previous formulations. Experiments on toxicity and fluency steering tasks demonstrate that NSMC and FA-NSMC consistently outperform best-of-n and bootstrap SMC methods, providing improved sequence-level reward optimization without model retraining.
discrete diffusionsequential monte carlofeynman-kac steeringinference-time controltext generation
SAE-Xplainers: Rule-Based Feature Interpretation for Extreme Earth Events
The paper introduces SAE-Xplainers, a rule-based feature interpretation framework for extreme Earth events (ExEE) prediction using weather and climate data. The method combines geographic location-based input modulation for Sparse Autoencoders (SAEs) with an ensemble of rule-based explainers to interpret high-dimensional environmental features. Evaluated on fire prediction, tropical cyclone detection, and atmospheric river identification, the approach improves reconstruction performance (quantitatively unspecified) and yields human-interpretable rules consistent with scientific literature while identifying feature absorption patterns.
sparse autoencodersextreme earth eventsrule-based interpretationgeographic modulationfeature absorption
Reward-Guided Autoregressive Graph Generation for Efficient Multi-Agent Communication Topology Design
The paper introduces Reward-Guided Autoregressive Graph Generation (RGA-Designer), improving upon ARG-Designer for efficient multi-agent communication topology design in LLM-based systems. The method trains a reward model to jointly optimize task correctness and structural compactness, then fine-tunes a pretrained graph generator using this reward signal via RLHF-inspired techniques. Results show preserved task accuracy while reducing token consumption by 20.5% on average compared to ARG-Designer.
autoregressive graph generationmulti-agent systemstoken efficiencyreinforcement learning from human feedbackcommunication topology
Orthogonal JEPA: Factorized Predictive States for Latent World Models
The paper introduces Orthogonal JEPA, a latent world-modeling framework that factorizes predictive states via orthogonal decomposition to address redundancy and gradient conflicts in standard Joint-embedding predictive architectures (JEPAs). The method employs learned basis matrices to analyze target states into multiple components, each predicted by dedicated branches from a shared context representation, with orthogonality objectives and regularization techniques to prevent collapse. Evaluations across vision, transcriptomics, health records, control, and molecular dynamics demonstrate improvements in representation quality, forecasting, planning, and long-horizon stability.
joint-embedding predictive architectureslatent world modelsorthogonal factorizationpredictive regressionrepresentation learning
DecoVAE: a Lightweight Interpretable Trend-Seasonal VAE Framework for Efficient Probabilistic Time Series Forecasting
The authors propose DecoVAE, a lightweight variational autoencoder framework for probabilistic time series forecasting that explicitly decomposes series into trend and seasonal components. The method enforces structural smoothness in the trend via differential regularization and models seasonality in the frequency domain using a complex Gaussian VAE. Evaluations on seven benchmarks show DecoVAE reduces CRPS by up to 52.68% and NMAE by 26.51% for long-term forecasting while being 93% smaller and 74% faster than baselines.
probabilistic forecastingvariational autoencodertime series decompositionhodrick-prescott filtercomplex gaussian
Auditing Cross-Lingual Fairness in Language Model Watermarking
The study introduces a framework for auditing cross-lingual fairness in language model watermarking, addressing limitations of English-centric evaluations. The method combines detection threshold calibration, threshold-independent detection assessment, three quality measurement paradigms (distributional, paired-semantic, reference-perplexity), and typological family-based disparity analysis. Evaluating six watermarking schemes across eleven languages and multiple model regimes reveals structural cross-lingual disparities predominantly between typological families, highlighting fairness gaps tied to language properties rather than individual languages.
language model watermarkingcross-lingual fairnesstypological familiesdetection threshold calibrationgeneralized-entropy decomposition
End-to-end Early Classification of Time Series in Non-Stationary Environments
The paper introduces DQeND, the first end-to-end Reinforcement Learning architecture for Early Classification of Time Series (ECTS) in non-stationary environments, jointly optimizing representation learning, classification, and triggering decisions. This contrasts with prior separable approaches that independently optimize components under stationarity assumptions. Evaluated across controlled drifting scenarios, DQeND demonstrates superior robustness, outperforming separable baselines by 12-18% in accuracy under various drift types. Ablation studies confirm that joint module updates are critical for adaptation. The results challenge the separable design paradigm and establish end-to-end learning as a promising direction for ECTS in dynamic settings.
early classification of time seriesnon-stationary environmentsreinforcement learningend-to-end learningconcept drift
From Street View Imagery to Street Quality Indicators: Vision Language Inference for the Suburban 15-minute City
The paper presents SAGAI, an open-source workflow leveraging vision-language models (VLMs) for large-scale streetscape quality assessment from Google Street View imagery. The improved framework features geographically consistent view generation, multi-VLM architecture support, and consensus-based inference, applied to evaluate pedestrian-friendly urban indicators (sidewalk presence, entrance density, vegetation) in Nice's suburbs. Results show desired qualities are concentrated in compact developments and traditional faubourgs, but lacking in residential hills, demonstrating VLMs' potential for scalable urban diagnostics.
vision-language modelsstreetscape analysisurban diagnosticsconsensus-based inferencewalkability assessment
CLaST: Context-aware Contrastive VAE for Probabilistic Time Series Forecasting
The authors propose CLaST, a context-aware contrastive VAE framework for probabilistic multivariate time series forecasting, addressing limitations in capturing temporal dependencies through a novel contrastive loss function. The method learns embeddings preserving contextual similarity between observations, enhancing latent representation expressiveness. Evaluated on nine benchmarks, CLaST outperforms baselines by up to 16.4% in CRPS and 14.4% in NMAE for short-term forecasting, and 48.6% in CRPS and 25.1% in NMAE for long-term prediction.
probabilistic forecastingcontrastive learningvariational autoencodertime serieslatent representations
Systematic Evaluation of TabPFN-TS for Zero-Shot Probabilistic Heat Load Forecasting in District Heating Networks
The study evaluates TabPFN-TS, a zero-shot probabilistic forecasting model pretrained on synthetic data, for heat load prediction in district heating networks. It compares TabPFN-TS against time-series foundation models (e.g., Chronos-2) and machine-learning baselines, analyzing covariate choice, context length (12-week optimal), and temporal resolution (hourly). TabPFN-TS achieves comparable deterministic accuracy to Chronos-2 (13.06% vs. 12.48% CVRMSE) with better calibration, despite synthetic pretraining. The findings motivate a Multi-Resolution Residual-Correction Forecaster to enhance long-horizon accuracy.
zero-shot forecastingprobabilistic forecastingdistrict heatingsynthetic pretrainingcontext length
Scale-Aware Pretraining of Time Series Foundation Models via Multi-Patch Token Alignment and Hybrid Masking
The paper introduces SATS, a scale-aware pretraining framework for time series foundation models that addresses heterogeneous sampling frequencies through multi-patch token alignment and hybrid masking. SATS employs a contrastive-inspired alignment regularizer to unify representation spaces across scales while preserving distinct modeling capacities, alongside a hybrid masking strategy combining random and contiguous masking for multi-scale temporal structure capture. Results show a 9.2% MSE improvement and 8.3% MASE gain over baselines, with a 65.6% increase in model efficiency.
time seriesfoundation modelstoken alignmenthybrid maskingpretraining
Green BOA: Determining the environmental break-even point for ML-based data compression
The study evaluates the environmental break-even point for machine learning (ML)-based data compression by comparing the carbon footprint of ML training and inference with the savings from reduced storage. Using a lossless compression algorithm as a case study, the authors estimate carbon-equivalent emissions for both ML infrastructure and storage reduction. Their analysis identifies the threshold where the environmental cost of ML is offset by storage savings, providing insights into sustainable ML deployment.
machine learningdata compressionenvironmental sustainabilitycarbon footprintlossless compression
Flow Matching Meets 3D Curvilinear Structure Segmentation in Medical Imaging
The authors propose 3D-CurvSegFlow, a flow matching-based model for segmenting 3D curvilinear structures in medical images, addressing challenges like complex topology and class imbalance. The method learns a continuous transformation from a simple source distribution to target vascular representations, enabling efficient progressive refinement. Evaluated on three public datasets (portal vein, cerebral vessel, coronary arteries), it outperforms general-purpose and vessel-specific approaches while preserving thin branches and vascular continuity, demonstrating cross-anatomy generalization.
flow matching3d segmentationcurvilinear structuresmedical imagingvascular representation
G-MARK: Grounded Multi-Agent Reasoning for Cooperative Driving via Knowledge Graphs
G-MARK introduces a grounded multi-agent reasoning framework for cooperative driving that converts object-centric observations into provenance-aware knowledge graphs (KGs), preserving source attribution, visibility, uncertainty, and spatial relations. The method derives shared feature representations from KGs to support object reasoning, motion prediction, and trajectory forecasting. Compared to state-of-the-art baselines, G-MARK improves occlusion reasoning accuracy by 42.2%, reduces control-selection error by 13.1%, and achieves comparable trajectory-planning accuracy with a 25.6x smaller communication payload.
knowledge graphsmulti-agent systemscooperative drivingocclusion reasoningtrajectory forecasting
From Noise to Signal: Improving Security Log Anomaly Detection Using LLMs with Endpoint-Specific Logs
This study introduces a standardized instruction-based LLM framework for detecting anomalous authentication behaviors, addressing limitations of rule-based and statistical methods. It utilizes a controlled cybersecurity testbed to generate endpoint-specific data, evaluating three LLMs against traditional tools. Meta Llama 3.1 8B Instruct achieved the highest accuracy (89.3%) and F1-score (91.8%), outperforming Wazuh (52.0%) and OpenSearch (49.3%), while also detecting 80% of borderline cases.
llmanomaly detectionauthenticationcybersecurityendpoint-specific
Auditing Recorded Predictive Lead Service-Line Classifications Against Physical Verification: A Statewide Study of New York
The study audits New York State's predictive material classification against physical verification for utility service lines. It analyzes 153 localities covering 125,990 addresses, finding 75 cases (49%) where one classification method is used. Seven discrepancies involve physical verification crews reporting beyond sampling explanations, with six cases involving lead lines among 1,150-1,450 addresses.
auditclassificationdiscrepancyleadverification
Multi-Source Wasserstein Distributionally Robust Graph Learning
The paper proposes MS-WDRO, a multi-source Wasserstein distributionally robust framework for graph learning from heterogeneous data sources. The method fuses sources via their Wasserstein barycenter as a nominal distribution, constructs an ambiguity ball for robustness, and solves the resulting regularized Laplacian estimation via ADMM with non-asymptotic guarantees. Theoretical contributions include concentration bounds for empirical barycenters and excess risk bounds with logarithmic source-count dependence. Experiments on synthetic data and the ABIDE-I neuroimaging dataset demonstrate superior graph recovery and sample efficiency versus seven baselines, particularly in low-sample regimes.
wasserstein barycenterdistributionally robust optimizationgraph learningmulti-source fusionlaplacian estimation
A Layered Simplex Architecture for Large Alphabets
We introduce a Bayesian estimator for probability estimation over large alphabets under log loss, characterized by four key properties. The method constructs estimates by multiplying independent uniform draws from the probability simplex coordinate-wise and renormalizing, with depth as the sole structural parameter. The estimator's regret admits an explicit, efficiently computable expression, enabling identification of scaling laws in data, alphabet size, and depth. Despite its simplicity and lack of tuned constants, it competes with specialized methods like Good-Turing across synthetic and real-text benchmarks. For Zipf targets with exponent above one, regret closely matches the description length of discovered symbols, revealing the data exponent as the rate of symbol discovery.
bayesian estimatorprobability simplexregretzipf targetsdescription length
PETA:Parameter-Efficient Test-Time Adaptation for Virtual Screening
The paper introduces PETA, a parameter-efficient test-time adaptation framework for virtual screening that specializes pretrained models to individual protein pockets without full retraining. PETA constructs challenging negatives via molecular diffusion and embedding-space mixup, then optimizes a ranking objective focusing on suppressing invalid high-scoring candidates. Experiments show PETA outperforms both pretrained and fully retrained baselines while updating only 0.03% of parameters (LayerNorm weights) across diverse benchmarks.
virtual screeningtest-time adaptationmolecular diffusionembedding-space mixupparameter-efficient
Where Does the Union Bound Go? Best-Arm Identification and Strong FWER Control
The paper clarifies the apparent discrepancy between union bound usage in fixed-confidence best-arm identification and multiple testing perspectives. By analyzing two natural hypothesis orientations, it demonstrates that the multiplicity factor of K-1 persists regardless of orientation: either as K-1 true nulls in strong FWER control, or through K-1 potential comparison errors when one null is true. The work formally establishes this equivalence using terminology from both sequential decision-making and multiple testing communities.
best-arm identificationfamilywise-error-rateunion boundmultiple testingfixed-confidence
The impact of feature engineering and an optimisation framework for ocean colour machine learning
The study proposes a seven-level feature engineering (FE) optimization framework for ocean color machine learning models, assessing its impact on Multi-layer Perceptron, Support Vector Machines, and eXtreme Gradient Boosting Trees for Chlorophyll-a concentration ([Chl-a]) and Secchi disk depth (Zsd) estimation. Using Sentinel-3 OLCI data from Norwegian coastal waters, optimized FE improved R² to 0.55 for [Chl-a] and 0.68 for Zsd, outperforming standard algorithms by 2× in R² and reducing mean absolute error by 63%. Results indicate FE optimization must be tailored to specific target variables and models.
feature engineeringocean color algorithmssentinel-3 olcichlorophyll-a estimationsecchi disk depth
Reliable Neural Collapse Approximation for Open-World Test-Time Adaptation
The paper proposes Reliable Neural Collapse approximation (ReNC) for Open-World Test-Time Adaptation (OWTTA), addressing label distribution shifts in target domains. The method leverages neural collapse as a structural prior, using pre-trained classifier weights as source-domain prototypes to filter Out-Of-Distribution (OOD) samples. A neural collapse approximation mechanism refines prototypes for target-domain adaptation while preserving neural collapse properties. Experiments on open-world benchmarks show ReNC's superiority in maintaining NC-related properties, offering insights for reliable OWTTA.
test-time adaptationneural collapseopen-world scenarioout-of-distributionprototype refinement
Interpretable Feature Learning for RF Fingerprinting via Polar MKANs
The authors propose Polar Monotonic Kolmogorov-Arnold Networks (Polar MKAN), an interpretable feature learning method for RF fingerprinting that addresses the opacity of deep learning approaches. Polar MKAN employs a block-partitioned monotonic encoder on polar inputs, where each latent dimension depends exclusively on magnitude or phase, ensuring channel separation and monotone responses by design. On a synthetic benchmark with gain and carrier frequency offset (CFO), Polar MKAN achieves 57.2% DCI Disentanglement, significantly outperforming unpartitioned baselines (≤12.9%). The method is further evaluated on real data for detection accuracy trade-offs and sensitivity to blind CFO compensation.
rf fingerprintingpolar mkanmonotonic encodercarrier frequency offsetdci disentanglement
A Repeated Measurements Approach to $SoH$ Battery Modelling of Cyclic Aged Data in a Laboratory Environment
The study introduces a first-order linearized nonlinear repeated measurements model for analyzing battery cell aging profiles under controlled laboratory conditions. The model captures two variance components: within aging profiles (measurement noise) and among aging profiles (test-to-test or cell-to-cell variation). Novel regularized iterative generalized least squares parameter identification schemes with optimal hyper-parameter re-estimation are employed to identify the hierarchical nonlinear model. Training data included State of Health (SoH) profiles for 10 cells aged at various constant discharge and charge current cycles at 25°C. Each cell's SoH profile is modeled using a power law expression, and aging parameter variation is modeled with a single knot cubic B-spline. SoH prediction accuracy achieved ±0.191% for SoH ∈ [0,20].
state of healthnonlinear modelparameter identificationb-splineaging profiles
Online Test-Time Adaptation for Generalizable Dynamic Graph Anomaly Detection
The contribution formulates online test-time adaptation for generalizable dynamic graph anomaly detection (DGAD) enabling pretrained detectors to identify anomalies in unseen target domains without retraining. OTTA-DGAD extracts dynamic prototypes evolving representations of normal patterns stores them in a memory buffer selectively retains general patterns shared across source domains incorporates new target patterns. An anomaly scorer compares incoming edge representations against these prototypes updates the buffer using reliable pseudo-labels enriches each target chunk with relevant representations retained from previous chunks compensates for sequential loss of data. Extensive experiments demonstrate state-of-the-art performance under strict test-then-adapt OTTA settings on ten real-world datasets from diverse domains.
onlinetest-timeadaptationdynamicprototypes
Inadvertent Context Leakage in Language Models
The paper demonstrates that language models inadvertently leak sensitive in-context information through subtle correlations in their outputs, even when refusing direct extraction. Using a novel adaptive black-box attack, the authors show that 2-digit secrets are reconstructed with near-perfect accuracy, while 4-digit secrets achieve 82% exact match across eight proprietary models. Stronger instruction-following models exhibit higher leakage, enabling practical attacks including semantic predicate inference and full Social Security Number extraction from production-style agents.
context leakageadaptive attackblack-box accessinstruction-followingsemantic inference
FAR-DPO: Feasibility-Aware and Robust Direct Preference Optimization for Cyclic Peptide Design
FAR-DPO introduces a feasibility-aware and robust direct preference optimization framework for cyclic peptide design, addressing geometric and biophysical constraints through architecture-agnostic optimization. The method combines feasibility-gated multi-objective dominance for preference construction with difficulty-aware group-robust optimization, adaptively reweighting predefined difficulty groups based on preference losses. On the CPSea LNR benchmark, FAR-DPO improves success rates from 46.89% to 57.79% for PepGLAD and 47.96% to 49.57% for PepFlow, with notable gains in the hardest target quartile and improved binding scores.
cyclic peptide designdirect preference optimizationmulti-objective dominancegroup-robust optimizationbiophysical constraints
Answer-Level Trust Selection for Physical Vision-Language Reasoning
The paper introduces Answer-Level Trust Selection (ATS), a post-hoc framework for assessing reliability of individual quantitative predictions from vision-language models (VLMs) without ground truth. ATS aggregates eight interpretable diagnostic scores from repeated queries and controlled interventions to compute trust scores, addressing limitations of self-consistency methods that miss stable-but-wrong or prior-tracking failures. Evaluated on Qwen2.5-VL-7B and 20 VLM backbones, ATS improves failure-case rejection but may reduce correct prediction retention, providing complementary answer-level reliability assessment to model-level evaluation.
vision-language modelsselective predictionphysical reasoningtrust scoringbehavioral diagnostics
Unsupervised Anomaly Detection Using Flow Matching on Tabular Data
The paper investigates flow-matching-based anomaly detection for financial transaction logs with contaminated training data, comparing Time-Conditioned Contraction Matching (TCCM) and Forest-Flow. It demonstrates that trajectory-based Deviation and Reconstruction scores outperform TCCM's Decision score in robustness to contamination, making Forest-Flow competitive despite severe class imbalance. Results highlight the critical role of anomaly scoring function selection in flow-matching methods for unsupervised tabular anomaly detection.
flow matchinganomaly detectiontabular datatraining contaminationscoring functions
Learning piecewise-smooth dynamical systems
The paper introduces a modular framework for learning piecewise-smooth dynamical systems from trajectory data, addressing the understudied problem of discontinuous dynamics in system identification. The method first estimates switching hyperplanes separating dynamical regimes, then learns smooth dynamics within each region using geometry-constrained neural networks, including a novel architecture with prescribed discontinuity sets. Theoretical analysis covers identifiability of discontinuities and approximation properties. Experiments validate the approach on low-dimensional benchmarks, including dry-friction oscillators and the PP04 climate model.
piecewise-smooth dynamicssystem identificationswitching hyperplanesgeometry-constrained neural networksdiscontinuous dynamical systems
skchange: Fast and Flexible Algorithms for Changepoint Detection
The skchange library introduces a Python framework for structural changepoint detection in time series, offering modular algorithms for both cost minimization and statistical testing approaches. It features anomaly segment detection, fast approximate search methods, high-dimensional data support, and automated penalty calibration. Designed with scikit-learn compatibility and Numba-accelerated performance, the toolkit provides extensible implementations of modern changepoint detection techniques. The open-source package includes diverse built-in cost functions and statistical tests for flexible deployment across various changepoint scenarios.
changepoint detectiontime series analysisstructural breaksscikit-learnnumba acceleration
Far from the Crowd: Scalable Self-Supervised Learning via Geographic Isolation
The paper introduces a scalable self-supervised learning curriculum for remote sensing imagery based on geographic isolation, a metadata-derived proxy requiring no image decoding or manual annotation. The method ranks samples by isolation distance, scaling as O(D log D) with dataset size D, and integrates with MoCoV2 and MAE pretraining. Evaluated on CopernicusBench (BigEarthNet, DFC-2020, LCZ), it achieves baseline performance with 20-40% of the training budget and improves downstream performance by up to +5 mAP, matching visual-complexity curricula while reducing pre-computation cost by 140x. Analysis shows curriculum-trained encoders develop higher-dimensional, more uniformly utilized embedding spaces.
self-supervised learninggeographic isolationremote sensingcurriculum learningembedding spaces
RecPFN: Prior-Fitted Networks for In-Context-Based Recommendations
RecPFN introduces prior-fitted networks for in-context learning in sequential recommendation, pretrained on synthetic clickstream data from a structural causal prior. The method employs a lightweight decoder-only transformer that performs Bayesian-style inference from small support sets, generating next-item predictions in a single forward pass without weight updates. Evaluated on eight public benchmarks, RecPFN achieves state-of-the-art zero-shot performance, competes with supervised methods in low-data regimes, and demonstrates robustness to domain shift while maintaining deployment efficiency.
prior-fitted networksin-context learningsequential recommendationstructural causal priorzero-shot performance
Projector Is All You Train
The paper demonstrates that training only the projector module—rather than both the projector and language model backbone—is sufficient for adapting multimodal large language models (MLLMs) to new modalities. Experiments on 3D MLLMs show comparable performance to joint training baselines while avoiding capability drift in the pretrained backbone. Projector-only training achieves ~2× higher sample throughput and maintains performance on 3D classification, captioning, and multimodal reasoning benchmarks across diverse backbones.
multimodal large language modelprojector module3d classificationcapability driftsample throughput
RIPE++: Reinforced Keypoint Learning from Positive Pairs Only
The paper introduces RIPE++, a reinforcement learning (RL) framework for sparse keypoint learning that operates solely on positive image pairs without requiring negative samples or pose supervision. The method leverages geometric consistency signals to derive both rewards and penalties, enabling discriminative detector and descriptor training. By extending the RL objective to matching via LightGlue adaptation, it achieves 59.65 AUC@5 on MegaDepth1500 (vs. 56.58 baseline) and demonstrates applicability to medical imaging where standard SfM fails.
reinforcement learningkeypoint detectiongeometric consistencyweakly-supervised learningsparse matching
Learning Deterministic and Stochastic Forced Hamiltonian Systems
The authors present a geometric framework for learning deterministic and stochastic forced Hamiltonian systems using neural networks. They introduce Generalized Forced Hamiltonian Neural Networks (GFHNNs), a structure-preserving architecture based on Lagrange-d'Alembert maps and variational integrator theory, with proven universal approximation capability. The method extends to parametric (PGFHNNs) and stochastic systems via Stratonovich-Taylor expansions. Experiments show GFHNNs achieve superior long-time stability and accuracy compared to residual networks while requiring less training data.
hamiltonian systemslagrange-d'alembert principlevariational integratorsstructure-preserving neural networksstratonovich-taylor expansion
CacheRoute: Planned Prefix-Affinity Routing for Large-Scale LLM Serving
CacheRoute introduces planned prefix-affinity routing for efficient large-scale LLM serving, resolving the tradeoff between KV-cache reuse and load balancing. The method periodically routes requests by admitting high-rate keys to a stable warm set and optimizing assignments based on expected load. Evaluated on Llama-3.3-70B across 60 H100 GPUs, CacheRoute achieves 176±11 QPS at a 3.5-s p99 SLO, 2.3x faster than baselines, while increasing KV-cache hit rate from 64.1±1.3% to 93.2±0.5%. Controlled experiments validate the approach, though performance gains diminish when affinity recovers insufficient KV work.
prefix cachingkv-cacheload balancingllm servingrouting plan
SAGE-XGBoost: Spatially Augmented Graph Embeddings--Machine Learning Framework for Natural Hazards Susceptibility Mapping under Data Scarcity
The study introduces SAGE-XGBoost, a feature-engineering framework combining noise-based data augmentation and graph embeddings to address data scarcity in natural hazard susceptibility mapping. The method constructs K-nearest neighbor graphs to derive local spatial statistics, integrates them with environmental covariates via PCA, and uses XGBoost for prediction. Evaluated on landslide and wildfire susceptibility, it outperformed baselines by >33 percentage points, achieving AUCs of 0.97 and 0.95 respectively, while improving spatial coherence.
graph embeddingsdata augmentationxgboostspatial statisticssusceptibility mapping
Scaffolding Minds: Optimizing Latent Visual Target Representations for Multimodal Reasoning
The paper introduces Scaffolding Minds, a method improving latent visual reasoning by addressing two limitations in existing two-stage frameworks. First, it replaces off-the-shelf vision encoders with a task-aligned scaffolding encoder during supervised fine-tuning. Second, it enhances reinforcement learning by modeling both mean and variance in latent space exploration. These improvements are shown to be complementary, yielding +9.5% improvement on FrozenLake spatial planning and +5.2% average gain across nine visual reasoning benchmarks, with larger gains (+19%) at higher grid resolutions.
latent reasoningscaffolding encodermultimodal reasoningreinforcement learningvision-language models
FleetSieve: Decision-Critical Profiling for SLO-Aware LLM Fleet Configuration
FleetSieve optimizes LLM serving fleet configuration by selectively profiling tensor-parallel degrees and replica counts based on their decision-critical impact, avoiding exhaustive measurements. It jointly models capacity and tail latency, compares conservative/optimistic allocations, and terminates when the decision gap falls below a tolerance threshold. Evaluations on a 31B-parameter model show 5.4% mean reduction in GPU-seconds (95% CI: 3.5-7.2%) versus random profiling, with 21.5% savings for Chat workloads, while preventing SLO violations (46.4s p99 vs 30s target). Incorrect sparse decisions in 16-GPU allocations cause up to 1.93 req/s throughput loss and 12.4pp fulfillment degradation.
tensor-paralleltail latencyslo-awareresource allocationgpu-seconds
Rationally Enriched Chebyshev Trunk Bases for DeepONet Surrogates of High Péclet Entrance Transport
The study introduces a rationally enriched Chebyshev (REC) trunk for DeepONet surrogate models to address singularly perturbed and high-Péclet transport problems characterized by thin boundary layers. The REC trunk integrates Chebyshev polynomials with rational elements constructed via the adaptive Antoulas-Anderson (AAA) algorithm. Evaluated across three problems—a singularly perturbed scalar BVP, thermal entrance problem, and concentration entrance problem—the REC-trunk DeepONet outperforms vanilla DeepONet and matches Chebyshev-trunk DeepONet in scalar profile prediction, reducing errors by up to 19.5% for perturbation parameters between 1.00×10⁻⁴ and 1.78×10⁻⁴. It also reduces wall-normal temperature and concentration profile errors by up to 60.2% and 32.2%, respectively, while mitigating near-wall oscillations for Péclet numbers from 10² to 10⁴.
deeponetchebyshev polynomialspéclet numberadaptive antoulas-anderson algorithmsingular perturbation
Time-Uniform Self-Normalized Concentration for Discounted Least Squares: Limits and Corrections
The paper identifies a critical error in a widely used weighted extension of self-normalized concentration inequalities for discounted least-squares estimators in non-stationary problems. Through a scalar Gaussian counterexample, it demonstrates that the claimed bounded radius is violated with probability one. The authors analyze deterministic anytime boundaries, showing they must scale as $R\sqrt{\log(T/\delta)}$ under certain conditions, and provide corrected finite- and infinite-horizon versions of the inequality while discussing implications for downstream analyses.
self-normalized concentrationdiscounted least-squarestime-uniform guaranteesub-gaussian modelstopping-time argument
Complementary, Not Cumulative: Interaction Effects in Physics-Informed Neural Networks for Navier-Stokes Vortex Shedding
This work investigates the interaction effects of training techniques in physics-informed neural networks (PINNs) for Navier-Stokes vortex shedding, demonstrating that individually effective methods do not necessarily compose. Using the DFG/Schafer-Turek unsteady cylinder wake benchmark, the study shows that most techniques perform no better than a baseline when applied in isolation. However, combining periodic (SIREN) activations with causal weighting achieves a previously inaccessible regime, reconstructing velocity and pressure fields with 4.1% average relative L2 error against an OpenFOAM reference solution. Further additions degrade performance, revealing nonlinear interactions and indicating that elaborate training recipes are not inherently superior.
physics-informed neural networksnavier-stokessiren activationscausal weightingvortex shedding
Unregularized Convergence of Single-Loop, Entropy-Regularized Natural Actor-Critic
(No summary returned.)
Kähler landscapes for complex neural network descents and guarantees including a search and destroy of the Calabi-Yau manifold
The paper analyzes optimization landscapes for complex-parameterized neural networks using Kähler geometry and information-theoretic manifolds. It introduces a natural gradient descent rule operating in the holomorphic tangent bundle, with updates scaled by the inverse Wirtinger Hessian of the log-likelihood potential. Key results identify pathological curvature conditions in Calabi-Yau information manifolds, showing that negative sectional curvature and Ricci curvature destabilize optimization, while fixed-determinant metrics induce eigenvalue blow-up. Theoretical connections are drawn to deep learning failure modes under ill-conditioned curvature.
kähler geometrycalabi-yau manifoldnatural gradient descentwirtinger hessianricci curvature
A Two-Stage Time-Aware Transformer for Short-Horizon AECOPD Risk Prediction
We propose a two-stage time-aware transformer for short-horizon prediction of acute exacerbation of chronic obstructive pulmonary disease (AECOPD) risk using raw pressure and flow waveforms from home ventilators. The first stage employs a classification model to identify high-risk patients, while the second stage uses a regression model to estimate the days remaining before an exacerbation event. Evaluated on seven days of ventilator data, the model achieves an F1 score of 0.91 for risk classification and an RMSE of 1.00 days with R^2 = 0.76 for time-to-event estimation, providing actionable lead time for clinical intervention.
time-aware transformeraecopd risk predictionhome ventilatortime-to-event estimationraw waveform analysis
Composition-Driven Phase Evolution in Sm-Doped BiFeO3 via Latent-Field Reconstruction of Atomically Resolved STEM Data
The authors introduce a field-based approach for analyzing atomically resolved STEM images, representing them via spatially varying latent Bragg fields to map crystalline order, lattice displacement, strain, and rotation. This method decodes observed images from these latent fields and is validated against conventional atom-resolved parameterization. Applied to Sm-substituted BiFeO3, the analysis reveals systematic phase evolution from extended ferroelectric domains at low Sm concentrations to a connected Pnma-dominated state at high Sm content, accompanied by enhanced shear and lattice rotation. This establishes latent-field reconstruction as a physically interpretable complement to atom finding for resolving phase evolution in ferroic materials.
latent bragg fieldsatomically resolved stemferroelectric domainsphase evolutionlattice displacement
Continuous Adversarial MeanFlow Transfer
The paper introduces MeanFlow-Transfer (MF-T) and Continuous Adversarial MeanFlow (CAMF) to address domain adaptation and sampling acceleration for pretrained generative models. MF-T unifies adaptation by mapping heterogeneous source outputs (e.g., DiT, SiT, JiT, iMF) into a shared velocity representation, enabling efficient transfer to target domains. CAMF extends adversarial refinement to finite-interval average velocities, recovering fine details lost in regression. Experiments on five target domains show MF-T with CAMF matches teacher FID/FDD at 125× fewer NFEs, while CAMF improves few-step FID by 29% on average.
meanflow-transfercontinuous adversarial meanflowdomain adaptationgenerative modelsneural function evaluations
Empirical Characterization of Learning Geometry in Hybrid Quantum Forecasting Models
The study characterizes learning dynamics in hybrid quantum forecasting models by comparing them with structurally aligned classical baselines. Using harmonic-mixture and chirp benchmarks, it analyzes Neural Tangent Kernel dynamics through kernel-target alignment, spectral concentration, and training loss. Results show the classical model has stronger early target alignment, while the hybrid model (125 parameters vs. 281) exhibits less concentrated kernel spectra and smaller drift, yet achieves similar validation performance. Fourier-augmented classical models fail to replicate this behavior, and re-uploading ablations show systematic optimization changes. The work highlights distinct learning trajectories yielding comparable generalization.
neural tangent kernelhybrid quantum modelkernel-target alignmentspectral concentrationre-uploading ablation
Beyond Multimodal Alignment: Certifying Physical Language through Response Substitution and Ordered Execution
The paper introduces the Disjoint-Bridge Operator-Substitution Certificate (DBOSC) to certify physical language in multimodal representations, assessing whether independently trained modality compilers can interchangeably execute responses outside their training data. The method evaluates response substitution and ordered execution in a controlled elastoplastic system with complementary modality blind spots. Results show that multimodal representations of unseen surfaces are 4.5x closer in response space than incorrect pairings, with fusion improving execution accuracy (oracle NMSE 0.18) and 14 of 16 pre-registered checks passing. The study demonstrates that compression and fusion alone cannot determine unseen composition laws, separating attribute access, response substitution, fusion closure, and ordered execution into distinct testable achievements.
multimodal representationsresponse substitutionordered executionfusion closurecomposition law
DeltaMomentum: A Key-Value based Anisotropic Momentum Update via Delta Rule
DeltaMomentum introduces a key-value based anisotropic momentum update rule that addresses input anisotropy in deep networks by forgetting directions at rates proportional to their frequency of occurrence. The method leverages the key-value structure of linear layer gradients, applying the delta rule to update the momentum buffer without matrix inversion, ensuring faster clearance of stale directions compared to exponential moving averages. DeltaAdamW, integrating DeltaMomentum with AdamW, achieves up to 46.39% fewer training steps to match AdamW's validation loss on FineWeb-Edu pretraining, with consistent gains across model scales (67M to 1B parameters) and tasks (CIFAR-10, ResNet-18, ViT-Tiny).
anisotropic momentumdelta rulekey-value structureexponential moving averagegradient tracking
Fine-Tuning VLAs with Self-Demonstrated Generative Control for Multi-Task Manipulation
The paper introduces a self-supervised fine-tuning method for vision-language-action (VLA) models to address performance degradation on new robot embodiments. By generating online interaction rollouts from the zero-shot VLA as additional training data, the approach preserves instruction-following priors while improving sample efficiency for new skills. Experiments on the ALOHA robot and RoboTwin simulation benchmark demonstrate that the fine-tuned model retains prior task capabilities, enables generalist instruction following, and efficiently learns new skills from expert data.
vision-language-actionself-supervisedfine-tuninginstruction-followingsample efficiency
When to Retrain: An Empirical Study of Retraining Policies for Streaming ML Under Concept Drift, Budget, and Latency Constraints
This study empirically evaluates retraining policies for streaming machine learning under concept drift, budget, and latency constraints. A unified system model incorporating retraining budgets and training-plus-deployment latency was used to compare periodic retraining, error-threshold triggering, and ADWIN-based statistical drift-triggered retraining against a no-retrain baseline. Across 3,933 experiment runs spanning multiple drift regimes, budgets, latency levels, datasets, and learning modes, results show that incremental learning renders retraining policies inconsequential, while non-incremental models exhibit 15-55% accuracy differences. Periodic retraining outperforms reactive policies under abrupt and gradual drift, with reactive policies excelling only in recurring drift scenarios. Systematic failure modes and latency-budget queueing interactions are documented.
concept driftretraining policiesincremental learninglatency constraintsstreaming ml
LLM as Detector: An In-context Learning Approach for Tabular Anomaly Detection
The paper proposes LLM-Detector, a framework leveraging large language models' in-context learning for tabular anomaly detection without fine-tuning. The method converts normal training data into statistical summaries, causal dependencies, and prototypes, which are organized into prompts for generating a scoring engine that evaluates statistical deviation, structural inconsistency, and density-based abnormality. Evaluated on 24 tabular datasets against 15 baselines, LLM-Detector shows consistent improvements in both mixed-type and continuous-only settings while eliminating LLM fine-tuning or neural network training requirements.
in-context learningtabular anomaly detectionstatistical deviationstructural inconsistencydensity-based abnormality
Quantifying Event Impacts on Time Series via Multiscale Contrastive Learning
The paper introduces EventTime, a multiscale contrastive learning framework for quantifying event impacts on financial time series. The method combines long-horizon market context, short-horizon pre-event dynamics, and event metadata via an event fusion module, and employs a dynamic contrastive objective to address sparse supervision. Evaluated on SECURE, a novel dataset aligning cybersecurity incidents with stock-market data, EventTime outperforms state-of-the-art baselines in estimating post-event abnormal losses, demonstrating event-sensitive representations and robustness to incomplete metadata.
contrastive learningtime-series forecastingevent fusionmarket impactcybersecurity incidents
Microlensify: a Transformer Based Machine Learning Classifier for Microlensing Events Trained on TESS Light Curves
The paper presents Microlensify, a transformer-based variational autoencoder trained on simulated single-lens microlensing events and real TESS Sector 12 data. The model classifies microlensing candidates, reconstructs light curves, and estimates event durations with 97% accuracy (R-squared). Applied to approximately 5.6 million TESS light curves, it identified between 0.036% and 1.89% as microlensing candidates across different pipelines. The model also detects false positives including long-period variables and transients, achieving 92.7% confirmation rate on published ground-based microlensing surveys.
transformervariationaltechnicaldistinctlowercase
Deep-MKV-TS: Path-Dependent McKean--Vlasov Control for Financial Time Series Generation
The article introduces Deep-MKV-TS, a path-dependent McKean-Vlasov control framework for financial time series generation. The method preserves reference drift while adjusting volatility, using a neural sample-based implementation of the stochastic maximum principle. Results show substantial reduction in path-dependent and volatility-related deficiencies, validated against exactly computable oracle. The correction remains effective in delayed-volatility experiments, improving conditional forecasts relative to reference and reaching performance comparable to flexible generative and historical baselines.
deep-mkv-tspath-dependentmckean-vlasovfinancialscenario
Heteroscedastic Neural Surrogate Modeling for Robust and Rapid Bayesian Inference in Fusion Plasma Diagnostics
The authors propose a heteroscedastic neural surrogate model to accelerate and robustify Bayesian inference via Markov Chain Monte Carlo (MCMC) in fusion plasma diagnostics. The dual-head architecture jointly predicts emission spectra and channel-wise noise variance, optimized with a Gaussian Negative Log-Likelihood (GNLL) objective to account for aleatoric uncertainty. Evaluated on Thomson scattering diagnostics, the method achieves >1500x speedup over physical forward models and reduces inference error (RMSE) by >20% compared to homoscedastic baselines, enabling real-time analysis in noise-dominated environments.
bayesian inferenceheteroscedastic modelingmcmc accelerationaleatoric uncertaintythomson scattering
Data-Driven Time-Varying Control Barrier Functions for Adaptive Safe-Set Learning with Online Decremental Support Vector Machines
The paper introduces a degradation-aware safety-filtering framework that learns and adapts time-varying safe sets for systems with reduced control authority. The method combines an initial radial basis function-kernel SVM for offline safe-set learning with an online decremental SVM update law that contracts the safe set based on degradation signals, enforced via a homotopy-smoothed control barrier function and quadratic-program safety filter. Theoretical guarantees include forward invariance of the time-varying safe set and recursive feasibility, validated on a VTOL model showing maintained safety under control authority reduction.
control barrier functionsdecremental svmsafety filteringtime-varying safe setsquadratic programming
Uncovering the Limits of Proof Sharing for Neural Networks
The paper investigates the limits of proof sharing for accelerating neural network verification via template reuse, revealing wide variation in template subsumption rates across architectures and properties. It introduces a novel metric, jointly stable neurons, to predict when template-based acceleration fails, and proposes FastCert, an automated technique for optimal template distribution across layers. Evaluated on L0-verification tasks, FastCert achieves a 1.13x average speedup over existing template-based methods.
neural network verificationproof sharingtemplate subsumptionjointly stable neuronsl0-verification
Deep neural networks as lattice gauge theories
The authors extend the neural network/quantum field theory (NN/QFT) duality by incorporating layerwise permutation symmetry, formulating deep neural networks as a (0+1)-dimensional lattice gauge theory. Neurons act as lattice sites, weight matrices as gauge fields, enabling computation of the tree-level neuron-neuron propagator to analyze layer variance evolution. Using Feynman diagrams, they derive perturbative expansions in 1/N, including recursive expressions for O(1) corrections to the propagator and neuron scattering amplitudes. This framework facilitates study of higher-point correlations and information propagation in deep networks via field-theoretic methods.
lattice gauge theoryneural networksfeynman diagramsperturbative expansionquantum field theory
Quantum Gaussian processes for prediction of channel observations
(No summary returned.)
Evaluating Neural Cartographic Relief Shading for Urban Environments: A Downtown Calgary Study Using High-Resolution DEM and DSM Data
The study contributes to terrain representation research by evaluating neural-based relief shading in urban environments, specifically comparing Eduard's machine-learning system with conventional analytical hillshading methods using high-resolution DEM and DSM data for downtown Calgary. The analysis focuses on parameter tuning for terrain type, micro/macro generalization, and flat-area detail, while maintaining consistent large-scale shading. Results indicate that Eduard, despite its alpine training bias, produces visually strong and cartographically useful outputs in certain urban contexts, though analytical methods remain more reliable in other scenarios. The findings advocate for future model training specifically targeting urban relief shading.
neural-based relief shadinganalytical hillshadingdigital elevation modeldigital surface modelparameter tuning
Holtercare-Bench: A Multimodal Benchmark for Evaluating Long-Term Dynamic ECG Analysis
We introduce Holtercare-Bench, a multimodal benchmark for evaluating long-term dynamic ECG analysis, addressing the limitations of current multimodal large language models (MLLMs) in processing ultra-long pathological sequences. The benchmark is based on Holtercare-23K, a dataset comprising 22,980 QA pairs from 788 clinical Holter records, featuring signal-video-text tri-modal alignment. Holtercare-Bench evaluates models on temporal localization, clinical diagnosis, and global summarization. Zero-shot evaluations of leading MLLMs reveal significant performance gaps, but fine-tuning yields substantial improvements. This work highlights the challenges in electrophysiology and provides a foundational benchmark for long-term medical MLLMs.
multimodal large language modelselectrocardiogramstemporal localizationclinical diagnosistri-modal alignment
Clustering and Token Denoising for Faster and More Robust VLMs
We introduce ClustRS, a training-free algorithm for robust token pruning in Visual-Language Models (VLMs), addressing computational inefficiency in edge deployment. ClustRS comprises two components: an attention-weighted clustering algorithm that selects representative tokens from semantic clusters, and Residual Shrinkage, a one-pass denoising step on selected tokens. Evaluated on ScienceQA-IMG and MM-VET benchmarks, ClustRS outperforms attention- and diversity-based methods by up to 20% under extreme noise and token reduction conditions (down to 16 tokens, 97% reduction) on LLaVA 1.5 7b. It matches baseline performance on LLaVA-OneVision with fewer than one-third of tokens under mild noise, demonstrating compute-efficient and noise-resilient VLM deployment.
token pruningvisual-language modelsattention-weighted clusteringresidual shrinkageedge deployment
Recovering Nonlinear Functions of Latent Variables: A Plausible-Value Neural Network Framework
The study introduces PV-ANN, a framework combining plausible values (posterior draws preserving latent variance) with artificial neural networks to recover nonlinear functions of latent variables. It derives a theoretical bound on recoverable variance via Hermite polynomial expansion, showing attenuation by ρ^k for kth-order components. Simulations across 18 conditions demonstrate that PV-ANN closes approximately 80% of the function-shape recovery gap between factor-score learners and true latent values, particularly in low-reliability scenarios, while predictive accuracy remains unaffected. A Big Five application illustrates exploratory workflows and boundary conditions under weak signal and measurement model misspecification.
plausible valueshermite polynomiallatent variablesartificial neural networksfactor scores
Compress and Forget: bitsandbytes Quantization Amplifies Proactive Interference in LLMs
This work demonstrates that bitsandbytes 4-bit quantization amplifies proactive interference (PI) in large language models, degrading retrieval accuracy under semantically dense contexts. The authors evaluate FP16, INT8, and INT4/NF4 quantization across Qwen2.5-7B-Instruct, Mistral-7B-Instruct-v0.3, and Phi-3.5-mini-instruct models using a fixed retrieval task. INT4 significantly reduces accuracy under high interference (e.g., 81.0% to 68.3% for Qwen), confirmed by McNemar's tests (p ≤ 2.6e-6) and mixed-effects regression. INT8 also incurs a smaller penalty in two models. Mechanistically, INT4 increases same-key intrusion errors (21.5% to 24.6%, p = 4.8e-7), with the effect originating in the quantized transformer backbone.
proactive interferencepost-training quantizationbitsandbytessame-key intrusiontransformer backbone
Continual Reasoning Gym: Diagnosing and Harnessing Shared Reasoning in Continual RLVR
The paper introduces Continual Reasoning Gym, a continual reinforcement learning with verifiable rewards (RLVR) environment that organizes text and visual reasoning tasks into five task sequences, addressing the costly capability expansion of multitask RLVR (MTRL). It identifies shared reasoning as a transferable reasoning structure that supports training across tasks and proposes Continual Prompt Replay (CPR) to harness this by replaying previous-task prompts and regenerating responses with the current policy. CPR achieves MTRL-level performance, demonstrating its effectiveness in mitigating forgetting and improving task learning.
continual reasoning gymreinforcement learningshared reasoningcontinual prompt replayverifiable rewards
📰 Industry Media (3)
Meet S1-mini: Superwhisper’s 462 MB Open-Weights Text Normalizer That Turns Raw ASR Transcripts Into Clean Written Text
Superwhisper introduces S1-mini, a 462 MB open-weights text normalizer fine-tuned from Qwen/Qwen3-0.6B, designed to clean raw ASR transcripts into polished written text. The model removes fillers, resolves self-corrections, applies punctuation, and renders spoken elements like numbers and email addresses in written form. It operates via a three-axis control line (Styling, Structure, Context) and achieves 94.8% token accuracy on a held-out test set of 7,519 cases. S1-mini is deployable on laptop CPUs, with a Q4_K_M GGUF build, and is constrained to avoid adding content or correcting facts.
text normalizerasr transcriptsqwen3-0.6btoken accuracygguf build
Meet UPDF: A Lightweight Adobe Alternative Built for the Agentic Era
UPDF by Superace Software Technologies introduces a lightweight PDF editor integrating AI-assisted interpretation with deterministic execution tasks. The tool supports direct editing of text, images, tables, and forms while preserving layout, alongside OCR (99% accuracy, 38 languages), batch conversion, and format preservation across 14 output types. Its GPT-5.6-powered AI assistant enables summarization, translation, semantic search, and literature review generation. Version 2.5 added ten specialized agents for tasks like bookmark generation and proofreading. UPDF claims a 5x cost advantage over Adobe Acrobat, with local processing and no third-party data sharing.
pdf editingoptical character recognitiongpt-5.6batch conversionsemantic search
How AI coding tools are contributing to the popularity of JavaScript
AI coding tools have amplified JavaScript and TypeScript dominance by favoring frameworks with extensive training data, such as React, leading to higher-quality code generation. Analysis of GitHub’s Octoverse 2025 report reveals a 66% YoY increase in TypeScript contributors, reaching 2.64 million monthly contributors. Type systems act as guardrails, reducing runtime errors in AI-generated code, with 78% of professional developers adopting TypeScript by 2026. Verification has emerged as the bottleneck, surpassing code generation in complexity. This trend reshapes hiring practices, emphasizing review skills over coding speed, while hindering adoption of newer frameworks lacking pre-existing training corpora.
typescriptcode generationverification bottleneckreacttraining corpus
Generated automatically at 2026-08-21 19:52 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.