Daily Digest — 2026-07-16
221 items · 10 research labs, 210 arxiv papers, 1 industry media
MarkTechPost: all feed URLs failed (last tried: https://www.marktechpost.com/feed/)AI News: all feed URLs failed (last tried: https://artificialintelligence-news.com/feed/)
🏛️ Research Labs (10)
The US is advancing AI safety through state and federal action
The article outlines a 'reverse federalism' approach to AI governance in the US, where state-level legislation (California, New York, Illinois) establishes common safety frameworks for frontier AI models, converging toward a federal standard. Key elements include documented risk assessments, incident reporting, and independent audits. The proposed federal framework, led by agencies like CAISI, would standardize model testing, cybersecurity evaluations, and international alignment. The approach aims to balance innovation with safety while avoiding regulatory fragmentation, emphasizing democratic oversight and global leadership in AI governance.
reverse federalismfrontier airisk assessmentsindependent auditscybersecurity evaluations
GPT-Red: Unlocking Self-Improvement for Robustness
OpenAI introduces GPT-Red, an automated red-teaming model trained via self-play reinforcement learning to identify vulnerabilities in LLMs. The system adversarially trains production models (e.g., GPT-5.6 Sol) by generating diverse prompt injections, reducing failure rates by 6x on hard benchmarks. GPT-Red achieves 84% attack success on novel scenarios (vs. 13% for humans) and demonstrates real-world exploit capabilities against agentic systems. Results show monotonic robustness improvements, with GPT-5.6 Sol resisting 99.95% of direct injections while maintaining baseline capabilities.
red-teamingprompt injectionself-play reinforcement learningadversarial trainingagentic systems
MUFG aims to become AI-native with OpenAI
Mitsubishi UFJ Financial Group (MUFG) deployed ChatGPT Enterprise to 35,000 employees, aiming to transform financial operations and customer experiences through AI-native integration. The initiative involved enterprise-wide training, custom GPT workshops, and departmental AI champions to foster adoption. Results include 100% training participation, 1,800 custom GPTs created in four months, and 20-30% workload reduction in research tasks. MUFG also explores AI-driven customer interfaces, such as natural language queries in wealth management apps and AI concierges for personalized financial advice.
chatgpt enterprisegenerative aiai-nativecustom gptsrobo-advisory
What building Shippy taught us about building agents
The Skylight team developed Shippy, a reliable AI agent for maritime domain awareness, focusing on deterministic tooling and rigorous evaluation. Shippy combines a system prompt ('soul'), markdown-based skills, and configurable runtime components (e.g., Claude Opus 4.6, OpenClaw framework). It interfaces with live Skylight data via a purpose-built CLI to avoid API misuse, and operates in isolated Kubernetes sandboxes per user. Evaluation uses subject-matter rubrics and LLM judges, achieving consistent performance on data retrieval (e.g., EEZ boundaries) while identifying edge cases like overstepping tactical recommendations.
agent-skills specdeterministic cliephemeral kubernetes deploymentllm judgemaritime domain awareness
Model Routing Is Simple. Until It Isn’t.
The authors demonstrate that model routing in agentic systems is fundamentally a systems optimization problem rather than a classification task, requiring simultaneous consideration of cost, latency, quality, and compliance. They analyze three key dimensions: cost dynamics influenced by caching behavior, task complexity beyond surface-level difficulty, and latency factors extending beyond model speed. Their optimization-based router achieves a 21% cost reduction and 9% latency improvement compared to single-model baselines while maintaining 84% accuracy on the AppWorld Test Challenge, with lightweight overhead of 6ms and 2kB per task. The approach highlights the importance of treating routing as a holistic system optimization rather than isolated model selection.
model routingsystems optimizationcaching behaviorlatency factorstask complexity
Welcome Inkling by Thinking Machines
Thinking Machines introduces Inkling, a 975B parameter multimodal Mixture-of-Experts (MoE) model with native support for text, image, and audio inputs, featuring a 1M token context window. The architecture employs hybrid attention (5:1 sliding window to global layers), relative positional encoding, and short convolutions, with 256 experts (41B active parameters). Trained on 45T multimodal tokens, it demonstrates capabilities in cross-modal reasoning and domain adaptation. The model is available in BF16 (2TB VRAM) and NVFP4 (600GB VRAM) variants, with optimized inference support in transformers, SGLang, vLLM, and llama.cpp.
multimodalmixture-of-expertsrelative attentionnvfp4kv-cache
Introducing Real World VoiceEQ: Measuring the human quality of voice AI
Real World VoiceEQ introduces a benchmark for evaluating the human-like quality of voice AI systems, addressing limitations in existing metrics like word error rate (WER) and latency. The benchmark assesses over 40 proprietary and open-source voice models across 15+ dimensions and 60+ metrics, including Automatic Speech Recognition (ASR), Text-to-Speech (TTS), and Speech-to-Speech (S2S). Developed from 1M+ human ratings, it highlights specialized capabilities in voice models, such as emotional understanding and conversational intelligence, rather than a single 'best' model. Results show significant variation in performance, particularly in handling noise, accents, and emotional cues, underscoring the need for context-aware evaluation beyond traditional benchmarks.
word error ratetext-to-speechspeech-to-speechemotional understandingconversational intelligence
Take our I/O 2026 quiz, vibe coded in Google AI Studio.
Google AI Studio demonstrates low-code application development using the Antigravity coding agent and Gemini models, enabling non-developers to create interactive content. The method involves prompt engineering with Gemini for task specification, followed by iterative refinement in Google AI Studio's preview interface. A quiz about I/O 2026 announcements was developed by an editor without coding expertise, showcasing the platform's accessibility. The workflow included uploading reference materials and prompt optimization based on generated outputs.
google ai studioantigravity coding agentgemini modelsprompt engineeringlow-code development
9 demos of Gemini Omni and Gemini 3.5 in action
Google introduced Gemini Omni and Gemini 3.5 Flash, showcasing multimodal and agentic capabilities. Gemini Omni enables video editing through natural language prompts, maintaining scene consistency and physics, while allowing iterative refinements. Gemini 3.5 Flash excels in long-horizon agentic tasks, leveraging Antigravity for multi-step workflows and coding, and powers interactive web UIs and personal AI agents. Demonstrations include video transformation, UX generation, and custom fitness trackers. Both models are integrated into Google’s ecosystem, with Gemini Omni available via Gemini app and YouTube, and Gemini 3.5 Flash accessible through Antigravity, AI Studio, and Search.
multimodalagenticantigravityworkflowsux
Check out real-life AI prototypes from the Futures Lab.
The Google-funded Futures Lab at University of Waterloo developed AI-powered educational prototypes through an eight-week interdisciplinary workshop. Computer science, business, and natural science students co-created tools like Kanji Garden (Japanese learning via AI-generated stories), SignFluent (real-time ASL feedback), and MuscleMemory (AI-based calisthenics form correction). Led by Dr. Edith Law, the program emphasizes practical applications of AI in education and future work scenarios, moving beyond theoretical research. Prototypes demonstrate novel integrations of generative AI, computer vision, and real-time feedback systems for immersive learning experiences.
generative aicomputer visionreal-time feedbackimmersive learninginterdisciplinary prototyping
📜 arXiv Papers (210)
Do AI Agents Know When a Task Is Simple? Toward Complexity-Aware Reasoning and Execution
The paper introduces E3 (Estimate, Execute, Expand), a method for task-aware execution-scope estimation in LLM agents, formalizing minimum-sufficient execution and the Agent Cognitive Redundancy Ratio (ACRR). E3 reduces computational cost by first estimating task difficulty, then executing a minimal viable path before expanding scope if verification fails. On MSE-Bench (121 deterministic edits), E3 achieves 100% success while reducing cost by 85%, tokens by 91%, and inspected files by 92% compared to baselines. Real-world validation with gpt-4o on LLM-Case confirms E3's efficiency gains without compromising task success.
task-aware executionminimum-sufficient executionagent cognitive redundancy ratiollm agentsexecution-scope estimation
TerraZero: Procedural Driving Simulation for Zero-Demonstration Self-Play at Scale
TerraZero introduces a procedural driving simulator and self-play training stack for zero-demonstration reinforcement learning at scale. The system combines a fast C-based engine (1.3M agent-steps/sec on one GPU) with procedural generation of diverse scenarios using real-world map geometry, randomized dynamics, and rule-based traffic agents. Policies train via compute-efficient self-play without human demonstrations, achieving zero-shot generalization across cities and left-hand traffic. TerraZero tops the InterPlan long-tail benchmark, excels in safety metrics (best collision/TTC scores on val14), and matches realism of reference-anchored methods on Waymo Open Sim Agents, while supporting multi-agent control of vehicles, pedestrians, and cyclists.
procedural generationself-playzero-shot generalizationreinforcement learningtraffic simulation
PalmClaw: A Native On-Device Agent Framework for Mobile Phones
PalmClaw introduces a native on-device agent framework for mobile phones, addressing limitations of existing GUI-based mobile agents by directly exposing device capabilities through structured tools with explicit execution boundaries. The framework manages sessions, memory, skills, tools, and agent loops locally, enabling direct access to device functions while maintaining control. Experimental results demonstrate an 11.5% relative improvement in task success, a 94.9% reduction in completion time, and lower setup overhead compared to baselines.
mobile agentsdevice capabilitiesexecution boundarieson-device frameworkstructured tools
Audio-Native Speech Recognition with a Frozen Discrete-Diffusion Language Model
The paper introduces a non-autoregressive speech recognition system using a frozen discrete-diffusion language model (DiffusionGemma, 26B MoE) with a Whisper encoder for acoustic features. A 42M-parameter adapter (0.16% of backbone) enables audio grounding via CTC loss, overcoming gradient flow limitations in native training objectives. The system achieves 6.6% WER on LibriSpeech test-clean, processes utterances in ~8 parallel steps independent of length, and supports multilingual transcription (English, Hindi, Mandarin) with a single adapter.
discrete diffusionspeech recognitionnon-autoregressivelow-rank adaptersconnectionist temporal classification
Dynamic Resource Allocation for Ensemble Determinization MCTS
The paper proposes two dynamic resource allocation methods for Ensemble Determinization Monte Carlo Tree Search (MCTS): Dynamic Number of Determinizations, which adapts the count of determinization trees based on search behavior, and Dynamic Simulation Allocation, which nonuniformly distributes simulations across trees for optimal knowledge gain. These enhancements were evaluated on three tabletop games (Jaipur, Lost Cities, Splendor) in iteration- and time-based settings, showing statistically significant performance improvements in specific configurations.
ensemble determinizationmonte carlo tree searchdynamic resource allocationdeterminization treessimulation budget
Win by Silence: Deletion Non-Monotonicity, Autonomous Exploitation, and Typed-State Gating in LLM Plan Evaluation
The paper identifies and analyzes a non-monotonicity failure in LLM plan evaluators where strategic omission of transitions improves scores despite semantic incompleteness. It formalizes the deletion effect through Proposition 1 (Δₖ = (∏ᵢ<ₖ pᵢ)[cₖ + (1 - pₖ)Rₖ₊₁]) and validates it empirically on 26 routes (57 deletions, all matching theory). A score-seeking optimizer exploited this to outperform baselines in 21/26 cases. The GATE mechanism blocked all 26 silenced routes (0 false positives) and improved coverage in 47/54 revisions. PCSC detection prevented omission splices while exposing registry-provenance boundaries in cooperative settings.
non-monotonicityplan evaluationtyped-state gatingomission incentiveregistry-provenance
Resist and Update: Counterfactual Report Coordinates for Incentive-Compatible LLMs
The paper introduces counterfactual report coordinates (CRCs) as a method for ensuring incentive-compatibility in aligned language models by enforcing causal contracts on model reports. Using interchange interventions on a Bayesian-witness benchmark, the authors identify low-rank report coordinates for answer, confidence, and caveat that are near-orthogonal and independently controllable. A training-free CRC clamp achieves resist and update scores of 1.00 (Wilson 95% CI [0.99,1.00]) on the benchmark, with deployable single-pass compilation showing reduced performance (0.73/0.97). The method generalizes across three model families and transfers to SycophancyEval.
counterfactual report coordinatesincentive-compatibilityinterchange interventionsbayesian-witness benchmarksycophancyeval
FormalAnalyticGeo: A Neural-Symbolic Based Framework for Multimodal Analytic Geometry Problem Generation
The authors introduce FormalAnalyticGeo, a neural-symbolic framework for multimodal analytic geometry problem generation, addressing the scarcity of annotated samples in this domain. The framework employs four specialized LLM components: Generator, Formalizer, Measurer, and Quality Verifier, operating sequentially with structured feedback enabling automatic retry loops. A formal intermediate representation, Condition Description Language (CDL), bridges problem text and precise diagram rendering via a Signed Distance Field (SDF) engine. The framework produces AnalyticGeo7K, a dataset of 7K verified multimodal problems with aligned text, diagrams, formal annotations, and ground truths, achieving a median relative error of 0.70% and 82.3% accuracy within 5% of exact solutions.
analytic geometryneural-symboliccondition description languagesigned distance fieldmultimodal
Form, Not Content? A Preregistered, Placebo-Controlled Evaluation of Learned Error-Conditioned Self-Repair Through Prompts and Weights in Frozen Small Code Models
The paper introduces PoPE (Popperian Placebo-controlled Evaluation), a methodology for assessing whether frozen small code LLMs (0.5-1.5B parameters) can operationally use execution counterexamples to self-repair. PoPE employs placebo controls to ablate task-relevant content or derange task-error assignments, evaluating models through prompt and weight channels (adapter training) with preregistered rules. Results show no significant superiority of error-content over placebos: 12 vs. 10 unlocks in prompt channel (content-ablated), and an 8-8 tie in weight channel (error-content vs. baseline). The study finds no evidence of equivalence or non-inferiority, suggesting error information may not externally guide conjecture testing.
popperian evaluationfrozen llmsself-repairplacebo-controlledadapter training
ViHoRec: A Quality-Controlled Vietnamese Hotel Recommendation Dataset and Cold-Start Benchmark
The authors introduce ViHoRec, a quality-controlled Vietnamese hotel recommendation dataset addressing three challenges: cross-platform entity resolution, reproducible quality metrics, and privacy-preserving benchmarking. The dataset comprises 18,267 interactions from 6,832 users and 560 hotels, collected from Booking.com, Traveloka, and Ivivu. Key contributions include a reproducible construction pipeline with entity resolution, HMAC pseudonymization for privacy, and a public cold-start benchmark with temporal splits. Results show sharp performance degradation for users with short histories (BPR-MF Recall@10 drops from 0.120 to 0.065), with UserKNN outperforming learned models, establishing ViHoRec as a sparse, cold-start-dominated benchmark.
recommender-systementity-resolutioncold-starthmacbpr-mf
Knowledge- and Gradient-Guided Reinforcement Learning for Parametrized Action Markov Decision Processes
The paper introduces Neuro-Symbolic Knowledge- and Gradient-Guided Reinforcement Learning (KGRL), a novel algorithm for Parametrized Action Markov Decision Processes (PAMDPs) that combines domain knowledge with gradient-based optimization. KGRL uses a Datalog knowledge base to prune inapplicable actions and constrain parameter spaces, then refines parameters via gradient descent. This approach improves sample efficiency and episodic return compared to state-of-the-art RL baselines, while providing procedural explanations through rule activation tracking. Experiments demonstrate superior performance in both metrics across PAMDP environments.
pamdpneuro-symbolicdataloggradient-based optimizationsample efficiency
Real-time fall detection based on vision for low-power edge platforms
The paper introduces a physics-informed fall detection framework that models falling as a stability-loss event in a coupled dynamical system. The proposed dual-LTC architecture comprises Center-of-Mass (CoM) and Base-of-Support (BoS) subsystems, instantiated as Liquid Time-Constant (LTC) neural networks, to continuously model inertial trajectory evolution and ground-contact adjustment. A learnable coupling module emulates physical interaction, while a Stability Manifold classifier detects boundary crossing using Lyapunov-inspired metrics. The architecture supports a three-state prediction paradigm (Normal, Falling, Fallen) and achieves competitive accuracy with a sub-50K-parameter network, enabling real-time inference on low-power edge devices.
liquid time-constantstability manifoldlyapunov-inspired metricscenter-of-massbase-of-support
MemOps: Benchmarking Lifecycle Memory Operations in Long-Horizon Conversations
The paper introduces MemOps, a benchmark for evaluating lifecycle memory operations in long-horizon conversational agents, moving beyond traditional black-box question-answering assessments. It structures memory as explicit operations (remembering, forgetting, updating, reflecting) with detailed traces of triggers, targets, and state transitions, embedded in task-oriented dialogues. Evaluations across long-context, retrieval-based, and parametric systems reveal distinct failure modes, showing session-level retrieval outperforms turn-level approaches and long-context models struggle with ordered state reconstruction. The benchmark enables interpretable, operation-level diagnosis of memory reliability.
long-term memorylifecycle operationsstructured tracesmemory-state trajectoriesinterpretable evaluation
UR-VC: Unsupervised Robotic Value Correction for Time-Derived Progress Proxies
UR-VC introduces an unsupervised method for correcting time-derived progress labels in robot learning, addressing the noise in using normalized time as a proxy for physical task progress. The method aggregates time-derived labels from similar states across multiple episodes, requiring no manual annotations or additional models. Evaluated on bimanual cloth manipulation, UR-VC captures non-uniform progress and local regressions, improving upon monotonic time-based signals. When used for advantage-conditioned policy learning, it shows promising results in real-robot task success.
robot learningprogress labelsunsupervised correctiondeformable-object manipulationadvantage-conditioned policy
A Multi-Agent System for Autonomous, Fine-Tuning-Free Clinical Symptom Detection: Development and Validation Study
The authors present Pythia, a multi-agent system for clinical symptom detection that autonomously generates and optimizes extraction prompts without fine-tuning or manual engineering. The system operates on locally hosted open-weights models, using development-set sensitivity and specificity for prompt selection. Evaluated on 72 signs/symptoms across 400 clinical notes, Pythia achieved mean sensitivity 0.76 and specificity 0.95, outperforming a lexicon-based approach (0.82 sensitivity, 0.76 specificity) in 20/62 comparable concepts and maintaining high specificity (0.97) for overflagged cases.
multi-agent systemclinical symptom detectionprompt optimizationopen-weights modelspecificity transfer
Unveiling Complex Collective Behaviors from Simple Rewards
The paper introduces a two-stage EEC explanatory framework to analyze emergent collective behaviors in Multi-agent Reinforcement Learning (MARL) for robot swarms, focusing on simple reward structures. A key contribution is the Agent Response Map (ARM), a novel analytical tool that spatially maps agents' decision-making patterns, identifying regions of aggregation and avoidance. ARM reveals that robots implicitly learn geometric fields of the environment and use these structures for coordinated movement. This finding is validated in two tasks: cooperative multi-robot shape assembly and competitive predator-prey pursuit-evasion, demonstrating ARM's ability to uncover hidden geometric structures in MARL policies.
multi-agent reinforcement learningagent response maprobot swarmsgeometric fieldscollective behaviors
ChartGenEval: Corruption-Tested Multi-Dimensional Feedback for Rhythm-Game Chart Generation
ChartGenEval introduces a six-question evaluation framework for rhythm-game chart generation, focusing on timing fidelity rather than note reconstruction. The method employs corruption-tested core outputs with dose-controlled failures, validated across 80 song groups and 40-song development panels. Results show sensitivity to phase shifts (15-60ms), pattern rewriting (37% perplexity reduction), and loop collapse (62% self-similarity increase), providing multi-dimensional feedback for generator comparison and iteration.
rhythm-game chart generationcorruption-tested evaluationtiming fidelitydose-controlled failuresmulti-dimensional feedback
Reproducible Reservoir Computing with Thermally Driven Superparamagnets: Controlling Temperature Sensitivity
This paper introduces a method to stabilize the performance of superparamagnetic nanodot ensembles for reservoir computing under ambient temperature fluctuations. By simulating magnetization dynamics and analyzing thermal activation effects, the authors propose heterogeneous nanodot patterns with varying sizes and characteristic timescales to mitigate temperature sensitivity. Benchmarking on the NARMA-10 task demonstrates that optimized heterogeneity maintains stable performance across a 5-35°C range with minimal performance loss. The study also characterizes the trade-off between performance and temperature stability, tunable via reservoir hyperparameters, advancing the practical deployment of these ultra-low energy computing substrates.
reservoir computingsuperparamagnetic nanodotsthermal activationheterogeneous patternsnarma-10
Accelerating Masked Diffusion Large Language Models: A Survey of Efficient Inference Techniques
The survey presents a unified latency decomposition framework for diffusion large language models (dLLMs) to analyze inference efficiency, addressing the gap between theoretical parallel generation advantages and practical deployment. It categorizes acceleration techniques across algorithmic innovations (e.g., diffusion-aware caching), architectural optimizations, and inference-time scaling, while emphasizing the need for rigorous benchmarking. The framework disentangles trade-offs between algorithmic, architectural, and system-level factors affecting end-to-end latency, offering guidelines for reproducible evaluation and highlighting open challenges in parallel generation.
diffusion large language modelsparallel generationlatency decompositioninference efficiencyalgorithmic innovations
Solution of the Hempel's statistical ambiguity problem and Causal AI
The paper resolves Carl Hempel's statistical ambiguity problem in inductive-statistical inference by proving that predictions derived from Maximally Specific Causal Relationships (MSCRs) are consistent. Using Nancy Cartwright's probabilistic causality framework, the authors introduce Causal Rules and a semantic probabilistic inference procedure that incrementally refines these rules with statistically relevant information. Theorem 1 establishes the consistency of MSCR-based predictions, addressing the ambiguity problem. The method enables applications in Causal AI and Causal Machine Learning, with connections to invariant feature learning and spurious association.
statistical ambiguitymaximal specificitycausal rulesprobabilistic inferencecausal ai
Human-AI Agent Interaction as a Neuroplastic Training Environment
The article introduces a framework leveraging human-AI agent interaction as a neuroplastic training environment to counteract conditioned reactive patterns. It posits that iterative AI interactions, such as generative image prompting, evoke high-frequency contact events that reinforce negative neural pathways through long-term potentiation. The framework proposes replacing reactive re-prompting with behind-the-scenes observation at the regulatory gap, thereby weakening these pathways through long-term depression. The method is characterized by three observation layers and two application modes: user-guided and agent-assisted. Results indicate that while observed and unobserved sessions behaviorally resemble each other, they neurologically oppose, demonstrating the framework's potential to mitigate negative neural reinforcement.
neuroplastic traininglong-term potentiationregulatory gapcontact eventsreactive patterns
Visual Access Boundaries in Vision-Language Model Reasoning
The study investigates the role of visual token access in Chain-of-Thought (CoT) prompting for Vision-Language Models (VLMs) by introducing Visual Access Sweep, a causal intervention that masks attention from generated tokens to image tokens. It defines the Visual Access Boundary (VAB) as the minimal access region preserving task accuracy. Experiments on Qwen2.5-VL and InternVL3 models reveal that CoT prompting does not prolong direct image-token access but extends language-side computation over image-derived hidden states. Results show CoT gains are constrained by perceptual readout, with improvements observed only when visual attributes are reliably extracted, placing the bottleneck at readout rather than counting.
chain-of-thoughtvision-language modelsvisual access boundarycausal interventionperceptual readout
PixelLoop: Shortcut Topological Navigation with Pixel-Level Loops
PixelLoop introduces dense pixel-level loop closures for topological navigation, contrasting with sparse image-level edges or metric SLAM corrections. The method grounds topological maps in pixel-space relative 3D geometry, enabling shortcut connections that directly alter planning connectivity and cost propagation rather than merely aligning coordinates. Experiments show 35% absolute improvements in Success Rate and SPL over image-relative baselines, with particularly strong gains in shortcut exploitation scenarios, validated through real-world robot deployments.
topological navigationloop closurepixel-level geometrycost propagationshortcut exploitation
Autonomous Tracking and Terminal Guidance of Moving Targets for Fixed-Wing UAVs
The paper presents a unified control framework for fixed-wing UAVs with pan-tilt cameras, enabling end-to-end target tracking and terminal engagement. The method combines vision-based target acquisition, UKF-based state estimation fusing YOLO detections with inertial data, and constraint-aware NMPC with CBFs to prevent self-occlusion, followed by quaternion-based BPNG for terminal guidance. Simulations demonstrate robust tracking and precise interception while respecting dynamic limits and field-of-view constraints.
nonlinear model predictive controlunscented kalman filtercontrol barrier functionsbiased proportional navigation guidancefixed-wing uav
The One-Word Census: Answer-Choice Conformity Across 44 Language Models
The One-Word Census introduces a minimalistic instrument to measure answer-choice conformity across 44 language models, revealing structured convergence patterns. Using 31 single-turn prompts with broad categories (e.g., 'Name a tree'), each model was queried four times without system prompts, and responses were analyzed via exact-match token normalization. Results show extreme convergence, with one answer dominating 80% of responses in 7 of 31 categories, and significant variation in conformity across models. Persona- and community-tuned models diverge most, while newer flagship models exhibit high conformity. Conformity increases across generations within Claude, GPT, Qwen, and Grok lineages, with reversals in the latest Claude and GPT models. The field is more concentrated than human norms in 18 of 20 shared categories.
conformitylanguage modelsexact-matchsurprisallineages
Silent Alarm: A J-Space Protocol for Comparing Danger Recognition Across Models and Quantization Levels
The paper introduces JADR (Jacobian Assessment of Danger Recognition), a protocol for evaluating LLM safety mechanisms by analyzing internal representations in Jacobian space (J-space) before token generation. It measures top-k J-space tokens across layers, grouped into six behavioral axes, comparing danger (StrongREJECT) and safe (XSTest, OKTest) prompts without external judges. The SafetyAUC metric, with bootstrap CIs, assesses model robustness across quantization (BF16, INT8, INT4) and fine-tuning. Evaluated on six models (Qwen3 variants, Gemma 2 9B), JADR significantly distinguishes safety mechanisms and quantization effects, validated against StrongREJECT grader.
j-spacejacobian assessmentsafetyaucquantizationdanger recognition
Who Grades the Grader? Co-Evolving Evaluation Metrics and Skills for Self-Improving LLM Agents
The paper introduces Double Ratchet, a co-evolutionary framework for simultaneously improving LLM agent skills and evaluation metrics without pre-existing reliable metrics. The method evolves metric compositions through a lifecycle involving drawback detectors, anchored reference sets, consensus regularization, and held-out audits. Results show 88-110% retention of performance lift compared to ground-truth-driven skill loops across code generation (MBPP+), text-to-SQL (Spider 2.0-Snow), and report generation tasks. Safety is maintained via anchor discipline and outer audits, with evolved outputs preferred over baselines in 77% of cases after rubric gaming was detected and corrected.
self-evolving agentsevaluation metric co-evolutionanchored reference setdouble ratchetlifecycle-managed skills
Do We Really Need Multimodal Emotion Language Models Larger Than 1B Parameters?
The paper challenges the necessity of large (>1B parameter) multimodal emotion recognition (MER) models by proposing Light-MER, a lightweight framework using knowledge distillation from large teacher models to sub-billion-parameter students. Key innovations include an optimal transport loss combining Sliced Wasserstein Distance with hidden-state alignment and a multi-reward GRPO optimization strategy balancing MER performance and efficiency. Experiments on nine benchmarks show Light-MER achieves state-of-the-art performance with improved inference efficiency, demonstrating the viability of small multimodal emotion language models.
multimodal emotion recognitionknowledge distillationoptimal transporthidden-state alignmentgrpo optimization
When Close Enough Is Not Enough: Autoregressive Drift in Quantum Circuit Synthesis
The paper investigates quantum circuit synthesis using a 44.8M-parameter encoder-decoder transformer, focusing on exact equivalence in fault-tolerant computing. For parameterized circuits (3-6 qubits), a hybrid approach combining transformer structure with classical optimization achieves median fidelity 1.000. On Clifford+T circuits (3-6 qubits), autoregressive drift causes exact equivalence to degrade from 88% (≤9 gates) to near zero (>26 gates), mitigated by inference-time candidate generation (22.5% exact-match) and 2.5x more data (39.5%). Post-processing rescues approximate outputs, but exact discrete correctness remains limited by drift.
quantum circuit synthesisautoregressive driftclifford+t circuitstransformert-count
HSEmotion Team at the 11th ABAW Challenge: Multi-Task Learning and Ambivalence/Hesitancy Video Recognition
The HSEmotion team introduces a multi-task learning framework for affective behavior analysis, achieving competitive performance in the 11th ABAW Challenge. The method employs frozen lightweight facial extractors (MT-EmotiDDAMFN, MT-EmotiEffNet-B0) with separate heads, systematic post-processing, and weighted backbone fusion for simultaneous prediction of valence, arousal, facial expressions, and action units on the s-Aff-Wild2 dataset. For ambivalence/hesitancy video recognition on the expanded BAH dataset, an audiovisual pipeline extends to video-level Macro F1 via late fusion of face, HuBERT audio, and RoBERTa text classifiers, temporal aggregation, and a global-text gate. Results show validation set Frame-level Weighted F1 improving from 0.74 to 0.79, with video-level Macro F1 reaching 0.73 on public test, outperforming ConvNeXt baselines without fine-tuning heavy backbones.
multi-task learningaffective behavior analysislate fusiontemporal aggregationweighted backbone fusion
Accuracy and Normalized Accuracy under Length Bias: Analysis, Guidelines, and a Bayesian Alternative
The paper introduces Bayesian accuracy, a scoring rule that mitigates length bias in multiple-choice benchmarks by computing posterior probabilities under an explicit length prior. Analyzing standard and length-normalized accuracy, the authors characterize their biases and demonstrate that normalization often over-corrects. Bayesian accuracy, requiring no additional forward passes, reduces empirical length bias across benchmarks and few-shot settings compared to existing methods. Results show it consistently outperforms both likelihood-based and normalized scoring in fairness across completion lengths.
length biasbayesian accuracymultiple-choice benchmarksscoring rulesnormalized accuracy
Constraint-Aware Aggregation for Federated Reinforcement Learning in Microgrid Energy Coordination
The paper introduces constraint-aware aggregation methods for Federated Reinforcement Learning (FedRL) in microgrid energy coordination, addressing the limitations of standard FedAvg in handling system-level constraints. The proposed penalty-based aggregation rule, $w_i \propto R_i - \alpha V_i$, integrates local performance ($R_i$) and estimated constraint violation ($V_i$) without requiring dual optimization or local training modifications. Evaluated on DairyGridEnv with synthetic and real-world datasets (Finland and German FIELD), the method reduces constraint violations by 15-20% while improving rewards compared to FedAvg. A combined reward-violation scheme offers tunable trade-offs via $\lambda$ but exhibits lower stability. The results highlight lightweight aggregation strategies' effectiveness in enhancing safety and performance in FedRL.
federated reinforcement learningconstraint-aware aggregationmicrogrid energy coordinationpenalty-based ruledairygridenv
Practical Judgment, Virtue, and Intuition in the Use of Opaque AI-Enabled Systems
The article proposes leveraging practical judgment, virtue, and intuition to mitigate ethical and operational challenges posed by opaque AI-enabled systems, particularly in military applications. It argues that humanistic, non-quantifiable values should guide training and deployment protocols to bridge the gap between system opacity and domain-specific norms. The framework emphasizes cultivating distinctly human capabilities to ensure ethical compliance and effective decision-making across domains deploying autonomous black-box systems.
opaque aipractical judgmentvirtue ethicsautonomous systemshumanistic values
Hallo4D: Multi-Modal Hallucination Mitigation for Consistent Spatio-Temporal Generation
Hallo4D introduces a unified framework for mitigating spatiotemporal hallucinations in 3D and 4D content generation. The method employs a generation-detection-correction paradigm, leveraging large multimodal language models (LMMs) to identify inconsistencies from multi-view and multi-frame renderings. It guides image-space consistency optimization through multi-model voting, incorporating motion-aware keyframe sampling, LMM-guided initialization, and appearance alignment. Additional techniques include exposure-aware optimization and visibility pruning for robustness. Experiments show Hallo4D outperforms baselines across diverse 3D and 4D generation settings, offering a scalable solution for consistency-aware content generation.
spatiotemporal hallucinationsmultimodal language modelsconsistency optimizationkeyframe samplingvisibility pruning
Weakly Supervised Spatio-Temporal Candidate Discovery of Dairy Farm Sites from Seasonal Satellite Imagery
A weakly supervised pipeline is proposed for ranking dairy farm candidate clusters from seasonal Sentinel-2 imagery and OpenStreetMap priors. The method employs a Barlow Twins encoder to learn multi-season tile embeddings without farm labels, combined with rule-based scoring incorporating farm proximity, seasonal pasture evidence, and summer greenness. Spatial smoothing via geographic proximity and embedding similarity generates ranked candidate clusters. Evaluated on 26,722 tiles from County Cork, Ireland, the pipeline identifies 535 high-confidence tiles forming 71 clusters, achieving 0.60 precision within 500m and 0.80 within 1000m for the top 5 clusters. Results demonstrate the efficacy of seasonal representation learning and weak geographic priors in reducing large satellite image collections to compact candidate sets.
weakly supervisedbarlow twinssentinel-2openstreetmapspatiotemporal
Tracing Agentic Failure from the Flow of Success
The paper introduces OAT, a lightweight unsupervised method for failure attribution in LLM-based agentic systems, trained exclusively on successful trajectories without step-level supervision. OAT frames the problem as one-class learning using neural controlled differential equations to model the dynamical patterns of successful trajectories in latent space. At inference, it assigns anomaly scores to steps in failure trajectories based on deviations from learned dynamics, identifying error steps. Experiments demonstrate OAT is 200--5000× faster than prompting-based baselines and achieves +20% and +7% F1 score improvements on in-domain and out-of-distribution datasets, respectively, using only 100 successful trajectories.
failure attributionneural controlled differential equationsone-class learninganomaly scoreagentic systems
LLMs Can See the Smoke but not the Fire: Evaluating Abductive Reasoning with Elenchos
The paper introduces Elenchos, a generative evaluation framework for assessing abductive reasoning in LLMs through structural inverse problems. Using formal systems like lambda-calculus, it tests whether models can detect mutations and infer the underlying rule modifications causing behavioral differences. Results show a detection-attribution dissociation: models often recognize system alterations but struggle to identify specific mutations, especially with interacting mutations, and exhibit limited improvement with increased reasoning budgets.
abductive reasoninglarge language modelsstructural inverse problemlambda-calculusdetection-attribution dissociation
Learning-based Probabilistic Load Forecasting with Post-hoc and In-model Uncertainty
The authors propose a unified probabilistic forecasting framework for one-day-ahead smart-building load prediction, addressing uncertainty propagation from reconstructed inputs. They compare post-hoc residual-quantile and integrated quantile-learning schemes across three deep learning backbones: recurrent, hybrid recurrent, and Temporal Fusion Transformer (TFT) models. Results show that integrated quantile learning with TFT achieves superior reliability, yielding 2.2-3.6% MAPE and 28-83W RMSE, with intervals 5x narrower than modular approaches. Reconstruction-sensitivity tests reveal a 106% increase in Quantile Score without widening intervals, indicating reconstruction uncertainty is not automatically absorbed. Findings highlight limitations of post-hoc residual quantiles in inference scenarios dependent on reconstructed inputs.
probabilistic forecastingtemporal fusion transformerquantile learningreconstruction uncertaintydiebold-mariano test
Bulkhead: Automated Semantic Detection and Remediation of Container Escape Vulnerabilities
Bulkhead introduces an automated framework combining large language models (LLMs) and formal methods to detect and remediate container escape vulnerabilities caused by filesystem isolation weaknesses. The system employs a multi-agent approach, using high-risk functional patterns to locate cross-boundary interaction entry points and call-chain patterns to reconstruct execution paths. It identifies vulnerabilities like missing security checks and TOCTOU flaws, generates proof-of-concept exploits for validation, and produces verified patches via assertion-driven model-checking templates. The method addresses limitations of static rule matching and manual auditing in existing defenses.
container escapepath traversalformal methodsmulti-agent systemmodel-checking
Line-Anchored Feedback Cuts Token Costs and Improves Correctness in AI Code Editing
The paper demonstrates that line-anchored feedback in AI code editing significantly reduces token generation costs while improving correctness. The authors compare holistic prompts with structured, line-anchored feedback using FileMark, a VSCodium extension. Results show token reductions of 22% (Claude Opus) to 58% (Claude Sonnet), with correctness improvements of +2.0 points pooled and +5 to +7 points for three local models. Further experiments reveal that correctness nearly triples for 100+ line files when edit-application is automated.
line-anchored feedbacktoken reductiongenerative aicode editingcorrectness improvement
MaxSAT-Based Feedback for Guiding Vision-Language Models in Sudoku
A neuro-symbolic approach integrating Maximum Satisfiability (MaxSAT) reasoning enhances Vision-Language Models (VLMs) in solving Sudoku puzzles by enforcing logical consistency. The method encodes VLM-generated candidate placements as soft clauses and Sudoku constraints as hard clauses in a partial MaxSAT formulation. When inconsistencies occur, the MaxSAT solver identifies a largest mutually consistent subset, providing structured textual and visual feedback for refinement. Evaluation on a Sudoku dataset across multiple VLMs shows improved logical consistency and increased solved instances, particularly in full-board refinement mode, demonstrating symbolic optimization's efficacy in vision-language reasoning.
maxsatvision-language modelsneuro-symboliclogical consistencysudoku
Less Experts, Faster Decoding: Cost-Aware Speculative Decoding for Mixture-of-Experts
We propose EcoSpec, a cost-aware speculative decoding framework for Mixture-of-Experts (MoE) models that optimizes draft token selection by considering expert activation costs. EcoSpec introduces a lightweight expert predictor and dynamic expert buffer to favor draft paths preserving high acceptance likelihood while reusing experts already activated, addressing expert scattering in confidence-driven speculative decoding. Evaluated on three large-scale MoE models (DeepSeek-V3.1, Qwen3-235B-A22B, GPT-OSS-120B) across reasoning, coding, question-answering, and dialogue benchmarks, EcoSpec reduces active expert footprints and achieves up to 1.62× end-to-end decoding speedup, demonstrating the importance of expert activation cost awareness in MoE inference.
speculative decodingmixture-of-expertsexpert scatteringexpert activationdraft token
From Critic to Confidence: PPO for Language-Based Quantitative Prediction with Confidence Estimation
CARE-PPO introduces a reinforcement learning framework for language-based quantitative prediction with confidence estimation, addressing LLM hallucinations and overconfidence. The method combines actor-critic Proximal Policy Optimization (PPO) fine-tuning with a Confidence-Aligned Reward for Estimation, enabling joint learning of accurate numerical predictions and reliable confidence signals. Evaluated on healthcare and finance tasks using Qwen-3 models (4B and 8B parameters), CARE-PPO outperforms logit-based and verbalized baselines in prediction accuracy and confidence alignment, demonstrating robustness under out-of-distribution scenarios. It also reduces task-specific overfitting, highlighting RL fine-tuning's generalization advantages over supervised approaches.
ppo fine-tuningconfidence estimationlanguage-based predictionactor-criticout-of-distribution robustness
Text-Aided Multi-Modal Panoptic Symbol Spotting for CAD Floor Plan Drawings
The paper proposes TextCAD, a multimodal framework for panoptic symbol spotting in CAD floor plans that jointly models graphical primitives and textual annotations. It introduces a Type-Attribute Correlation Encoder (TACE) to capture compositional annotation semantics and a Semantic Hierarchy Alignment framework with Multi-level Semantic Filtering (MSF) for cross-modal fusion. Evaluations on building-design datasets demonstrate state-of-the-art performance in symbol spotting accuracy.
panoptic symbol spottingmultimodal fusiontype-attribute correlationsemantic hierarchy alignmentcad floor plans
Internet of Agentic Things: Networked AI Agents for Closed-Loop IoT Orchestration
The paper proposes the Internet of Agentic Things (IoAT), an architectural framework integrating agentic AI, IoT, cyber-physical systems, Physical AI, edge computing, and digital twins for closed-loop orchestration. The architecture comprises cloud, edge/fog, and physical IoT layers interconnected via autonomous AI agents that perceive, reason, coordinate, and actuate across distributed cyber-physical environments. IoAT is formalized as a coupled workflow-control problem using a hylomorphic dynamic programming framework, enabling nested strategic and tactical decision-making. A smart-building orchestration use case is presented, alongside discussions on safety, security, governance, resilience, and trustworthy deployment challenges.
agentic aicyber-physical systemsedge computingdigital twinsdynamic programming
Jetson-PI: Towards Onboard Real-Time Robot Control via Foresight-Aligned Asynchronous Inference
Jetson-PI introduces Foresight-Aligned Asynchronous Correction for efficient Vision-Language-Action (VLA) model deployment on low-power devices like NVIDIA Jetson Orin. The method employs a lightweight future correction module to predict environment representations conditioned on committed actions, mitigating perception-execution misalignment. It combines confidence-based scheduling optimization with system-level accelerations (CUDA graph reuse, GPU-resident buffering, flow unrolling) to reduce reaction time. Experiments show 8.66× and 5.41× control frequency improvements over naive PyTorch and vla.cpp, respectively, and a 14.8% higher success rate than VLASH on the LIBERO benchmark.
vision-language-action modelsasynchronous inferencefuture correction moduleconfidence-based schedulingcuda graph reuse
Evidence-Grounded Verified Agentic Reasoning: A Path Toward Eliminating LLM Hallucination in Empirical Inference via Tool-Attested Kernel Proofs
EG-VAR (Evidence-Grounded Verified Agentic Reasoning) introduces a Lean 4-based tool-calling architecture that ensures LLM outputs are verified through tool-attestation axioms and kernel-checked inference chains. The system guarantees that all verified claims structurally descend from attested tool calls (Thm. 3.1) and valid inferences (Thm. 3.2), with residual outputs marked as Abstain. On TableBench (n=120), EG-VAR achieved 100% accuracy versus 95% for same-tool baselines; in counterfactual tests, it maintained 100% source-faithfulness versus 80-90% for baselines. Semantic-formalization errors were 3.3% (Sonnet) and 1.7% (Opus). EG-VAR serves as a governance interface for high-stakes empirical claims, enabling auditable proof obligations and abstention conditions.
evidence-grounded reasoninglean 4tool-attestationkernel-checked inferenceformal sidecar
A Learning-Rate-Gated Failure of GRPO in a Small Language and Vision-Language Model Web Agent: A Controlled Null and Its Mechanism
This work investigates whether Group Relative Policy Optimization (GRPO) enhances the performance of small-scale (4B-8B) language and vision-language model web agents beyond their supervised baselines. Through a controlled grid of 18 runs varying learning rate, KL weight, seed, initialization, and clipping, the study finds no credible improvement in success rates for tasks the agent has mastered. Notably, moderate to high learning rates degrade performance on text tasks, while GRPO only improves performance when the sampled policy outperforms the greedy one. Mechanistic analysis reveals that learning rate regimes differentially affect model components, with effective rank in late layers tracking capability at 4B but not at 8B, indicating scale-dependent behavior.
group relative policy optimizationlearning ratevision-language modeleffective rankscale-dependent
Atomic Units of X: The Compression Layer of Intelligence
The paper proposes the Compression Calculus, a formal framework for intelligence as atomic compression and compositional reuse across cognitive, computational, and biological systems. It introduces the Compounding Cascade thesis, arguing that abstraction layers multiplicatively enhance representational efficiency rather than yielding incremental gains. The analysis suggests contemporary AI systems (e.g., large language models) operate suboptimally by relying on token/document-level processing instead of concept-level atomic structures. The framework unifies knowledge representation, explainable AI, and adaptive system design through the lens of compression and abstraction.
compression calculuscompounding cascadeatomic unitsknowledge representationabstraction layers
Can Induced Emotion Bias LLM Behaviors in Sequential Decision Making?
The study examines whether induced emotion biases sequential decision-making in Large Language Models (LLMs), contrasting with known human behavioral effects. Using the Iowa Gambling Task (IGT) and an imagination-based emotion induction procedure, researchers validated LLMs' ability to perceive emotions and learn sequentially at human-like pace. Results show induced emotion does not significantly alter LLM decision dynamics on average, but anger specifically reduces penalty sensitivity and early-stage exploration, revealing context-dependent affective modulation distinct from human responses.
large language modelssequential decision-makingiowa gambling taskaffective modulationemotion induction
Agentic Service-Oriented Computing: A Manifesto for the Next Frontier of Service-Oriented Computing
The paper introduces Agentic Service-Oriented Computing (ASOC) as a novel research area addressing the engineering challenges of deploying LLM-powered agents in complex distributed workflows. ASOC focuses on six foundational principles: harness-ability, composability, lifecycle engineering, trustworthiness by design, goal-driven orchestration, and observability/accountability. It proposes a five-dimensional research agenda covering agentic services foundations, composition and orchestration, governance and accountability, security and trust, and evaluation and certification. The authors argue that the Services Computing community is uniquely positioned to provide the conceptual and engineering framework needed to transition agentic AI from fragmented prototypes to dependable, service-based systems.
agentic service-oriented computingllm-powered agentsservice orchestrationtrustworthiness by designagentic qos
Multi-Perspective Agentic Program Repair via Code Property Graphs and Temporal Execution Graphs
CT-Repair introduces a multi-perspective agentic program repair framework leveraging Code Property Graphs (CPGs) and Temporal Execution Graphs (TEGs) to address limitations in LLM-based APR. The method employs a three-stage filtering pipeline to compact TEGs, with three finite-state-machine-guided agents analyzing bugs from static, dynamic, and hybrid perspectives to produce evidence-grounded repair strategies. Evaluation on 854 Java bugs from Defects4J v3.0 shows CT-Repair correctly repairs 489 bugs in mixed-model configuration and 388 bugs in GPT-5.4-mini configuration, outperforming ReinFix and RepairAgent by 19 and 30 bugs respectively. The union of three evidence perspectives repairs 99 more bugs than the strongest individual perspective, while execution and behavior filtering reduce candidate method scope by 94.85% and runtime records by 55.97% respectively.
code property graphstemporal execution graphsagentic program repairfinite-state-machineexecution filtering
Vertical Standardisation for High-Risk AI Systems under the EU AI Act: A Domain-Specific Framework for Algorithmic Hiring
The paper proposes a vertical, domain-specific framework for algorithmic hiring systems to comply with the EU AI Act's requirements, focusing on ranking-based recruitment. Unlike horizontal AI governance approaches, it maps Act requirements to concrete standardisation recommendations across key areas: lifecycle discrimination risks, fairness-aware data governance, explainability, human oversight, and post-deployment monitoring. Recommendations were informed by the FINDHR project but remain implementation-agnostic, enabling alternative methods or tools. The framework addresses high-risk AI system challenges in recruitment, emphasizing compliance with Act-mandated standards for risk management, data quality, logging, and technical documentation.
algorithmic hiringeu ai actfairness-aware data governancepost-deployment monitoringranking-based recruitment
Explainable-by-Design Audio Deepfake Detection via Wiener-Hopf Linear Prediction
The authors propose an explainable-by-design audio deepfake detection framework combining Wiener-Hopf linear prediction with a lightweight 2D CNN, offering transparent classification based on acoustic signal properties. The method achieves competitive performance on benchmark datasets while maintaining lower computational complexity than state-of-the-art approaches. Grad-CAM analysis reveals the model focuses on low-order predictor coefficients and silence/transition regions, capturing reverberation artifacts and statistical inconsistencies in synthetic speech, with demonstrated robustness against common post-processing degradations.
wiener-hopf linear predictionaudio deepfake detectioninterpretabilityconvolutional neural networkgrad-cam
Deep Learning-based Surrogate Modelling of the LOD Method for Multiscale Problems
The paper introduces LOD-MSNO, a hybrid neural operator combining Localized Orthogonal Decomposition (LOD) with operator learning to address multiscale PDEs with rough coefficients. The method leverages LOD's problem-adapted basis functions as a multiscale prior while mitigating computational costs through data-driven operator learning. Theoretical error bounds are provided. Experiments show LOD-MSNO outperforms existing neural operators in accuracy for high-contrast elliptic PDEs while maintaining computational efficiency, achieving mean relative errors below 2% on benchmark problems.
multiscale pdesneural operatorlocalized orthogonal decompositionsurrogate modelingoperator learning
Traceback Translators Against Forgetting in Continual Fake Speech Detection
The paper proposes a forgetting-resilient approach for continual fake speech detection using domain translators within a frozen detector. The method employs traceback translator networks to remap new feature spaces into original ones, mitigating catastrophic forgetting. Experiments demonstrate high detection rates compared to traditional retraining, with reduced computational effort and maintained accuracy on prior data.
continual learningfake speech detectioncatastrophic forgettingdomain translatorstraceback translator
Mind the Gap: Promises and Pitfalls of Hierarchical Planning in LeWorldModel
The study introduces Hi-LeWM, a hierarchical extension of LeWorldModel (LeWM) for long-horizon goal-conditioned control, freezing the pretrained low-level LeWM and adding high-level planning over latent subgoals. Evaluations on PushT and Cube tasks reveal that hierarchy does not universally improve performance: short horizons favor a one-step high-level horizon, while longer horizons expose a mismatch between learned high-level action spaces and inference-time search distributions. Experiments demonstrate that constrained search around macro-actions encoded from training trajectories, with proper subgoal timing, recovers hierarchical regimes, improving over flat LeWM by +11.3% at medium-range horizons and +14.7% at the longest PushT horizon.
hierarchical planninglatent subgoalsgoal-conditioned controlmacro-actionstemporal abstraction
Evidence-Grounded AI for Musculoskeletal Care
The paper introduces OrthoPilot, an evidence-grounded AI system for longitudinal musculoskeletal care, integrating hospital data streams with external knowledge via a large language model. It autonomously processes real-time imaging, lab, and pathology data to generate evidence-based decisions from diagnosis to rehabilitation. Evaluated on a specialist-validated benchmark of 1,000 disease codes, OrthoPilot outperformed 81 orthopaedic physicians in diagnostic reasoning and management planning, and surpassed other intelligent systems across 60 clinical centres. In a prospective study of 1,870 complex cases, it improved full-chain management success by 10.6%, and in an 8-month deployment with 8,240 inpatients, increased cumulative cases per bed by 9.7%.
orthopilotmusculoskeletal carelarge language modelclinical aievidence-based decisions
OOD-RL-Bench: A Benchmark Framework for Out-of-Distribution Detection in Reinforcement Learning
The authors introduce OOD-RL-Bench, a benchmark framework for evaluating out-of-distribution (OOD) detection methods in reinforcement learning (RL). The framework supports configurable anomaly injection and detector evaluation across temporal RL trajectories, addressing limitations of static OOD benchmarks. Using a Deep Q-Network in LunarLander-v3, they assess detectors via metrics like matched-time AUROC and detection delay, finding performance varies significantly by anomaly type (e.g., high accuracy for observation perturbations but poor performance on action-conditioned dynamics). The framework is released as a reproducible artifact.
reinforcement learningout-of-distribution detectionbenchmark frameworkanomaly injectiontemporal dynamics
The Model Knows Your Project, Not You: Measuring Recognition in LLMs with NameRank
The authors introduce NameRank, a [0,1] recognition score measuring parametric knowledge in LLMs by probing 4,685 entities across 54 cohorts with open-ended questions. Independent judges evaluate responses against curated gold standards, penalizing hallucinations, context echoes, and guesses. Results reveal recognition favors named, indexable artifacts over credentials or titles, with Nobel, Turing, and Fields laureates achieving high scores. Tools outrank creators, and named methods or awarded papers propagate recognition better than contributions to celebrated artifacts. Bibliometrics poorly predict recognition, and top-density institutions outperform peers at matched citations. Recognition correlates with peak salience in news events, not persistence.
namerankparametric knowledgehallucinationbibliometricssalience
TRACE: An Operational Reasoning Schema for Auditable Agentic Commitments
The paper introduces TRACE (Typed Reasoning And Commitment Evidence), a schema for recording reasoning traces in agentic systems, with a focus on auditability. It proposes a typed, versioned record structure, an eight-stage reference writer, and a consumer contract to operationalize commitments. The method critiques language models' reasoning limitations (autoregressive association, lack of formal constructs) and demonstrates through examples (music lessons, search-and-rescue) how TRACE enables auditable decision-making. Results include schema specifications, benchmark protocols, and convergence metrics, though empirical evaluation is deferred.
reasoning tracesauditabilityautoregressive associationcommitment evidenceoperational interface
From Observation to Insight: Mechanistic World Models and the Quest for Autonomous Discovery
The paper proposes Mechanistic World Models, a paradigm shift from predictive to explanatory AI by centering reusable mechanisms in representation and learning. Drawing on philosophy of science, it formalizes computational requirements for discovery, identifies design principles favoring explanatory knowledge, and analyzes mechanism-centric model anatomy. The work synthesizes disparate approaches (mechanistic interpretability, causal representation learning, equation discovery) under a unified framework, arguing current ML lacks the structural organization needed for autonomous scientific discovery despite strong predictive performance.
mechanistic world modelsexplanatory mechanismsautonomous discoverycausal representation learningequation discovery
Agent-Safety Evaluations as Load-Bearing Evidence: A Vendor-Neutral, Cross-Harness Reconstructability Metric
The paper introduces a vendor-neutral reconstructability metric for agent-safety evaluations, addressing the lack of load-bearing evidence in current assessments. It proposes an eight-class property-level metric and a cross-harness adapter that generates Evidence Sufficiency Cards per decision, enabling monitor-coverage checks. A counterfactual-replay protocol and replayability-precondition probe are implemented to quantify claim-evidence gaps. Evaluations on public traces show sufficiency scores ranging 0.458-0.833, with replay preconditions unmet in all cases. The method successfully distinguishes between raw (0.542) and instrumented (0.667) variants in synthetic release tests, advocating for reconstructability vectors in safety claims.
reconstructability metricevidence sufficiencycounterfactual-replaymonitor-coverageagent-safety
An Omnilingual-ASR-Based Speech-LLM System for the 2nd MLC-SLM Challenge
The paper presents a cascaded diarization-then-recognition system for Task 1 of the 2nd MLC-SLM Challenge, combining DiariZen-Large-s80 segmentation, CAM++ embedding-based clustering, and a LoRA-adapted omniASR LLM 7B v2 recognizer. The system achieves a macro tcpMER of 29.27% on the Development set (150 conversations, 21 language/accent categories) versus the baseline's 79.15%, and 50.23% on the Evaluation set. Key findings show embedding-based clustering outperforms end-to-end speaker assignment, and overlap-aware segmentation increases tcpMER by duplicating transcriptions.
diarizationtcpmerlora-adaptedomniasrembedding-based clustering
Function-Aware Fill-in-the-Middle as Mid-Training for Coding Agent Foundation Models
The paper introduces function-aware fill-in-the-middle (FIM) mid-training to enhance coding agent foundation models by leveraging the structural isomorphism between agent action-observation loops and function call sites. The method masks functions selected via program dependency graph analysis and a complexity-inferability criterion, applied to a 2.6B-token decontaminated corpus from 968 GitHub repositories. Results show improvements of +2.8/+3.0 on SWE-Bench-Verified for Qwen2.5-Coder-Instruct (7B/14B) and +3.2 for Qwen3-8B, with additional gains on SWE-Bench-Lite (+3.7/+4.0/+5.4) and mitigation of capability erosion on non-agent benchmarks.
fill-in-the-middlemid-trainingprogram dependency graphcoding agentself-supervised objective
Do We Really Need Transformers for Global Spatial Information Extraction in Traffic Forecasting?
The study challenges the necessity of transformer-based attention for global spatial information extraction in traffic forecasting, proposing instead a simpler global aggregation operator. Through controlled ablation experiments replacing only spatial mixing modules, the authors compare attention mechanisms with uniform full-range mixing across six traffic benchmarks. Results show comparable performance (0.14% mean MAE difference), with each method outperforming on three datasets, while full-range mixing reduces complexity from O(N²) to O(N). Analysis reveals spatial attention decomposes into a row-uniform global background and dataset-dependent residual components, suggesting attention's benefits may not justify its complexity universally.
traffic forecastingspatial attentionglobal aggregationcomplexity analysisablation study
EVOQUANT: Self-Evolving Verifier-Guided Strategy Optimization for Robust Quantitative Trading
EVOQUANT introduces a self-evolving verifier-guided framework for automated quantitative trading strategy optimization, addressing limitations of manual tuning and LLM-induced errors. The method employs LLMs for performance bottleneck diagnosis, generates semantically controlled candidate edits, and selects optimal strategies via a multi-stage verification pipeline, while distilling optimization experience for continual self-improvement. Evaluated on seven strategies from A-share and Crypto markets, EVOQUANT significantly improves Sharpe ratios, with average test Sharpe increasing from -0.298 to 0.538 and best-performing strategy achieving 199% relative improvement. Ablation studies and stress tests confirm the framework's robustness, transforming strategy optimization into an automated, verifiable process.
quantitative tradingsharpe ratiomulti-stage verificationsemantic controlself-evolving
The Computational Basis of Confidence in Large Language Models
The study establishes statistical decision confidence (SDC) as a normative framework for analyzing confidence signals in large language models, demonstrating that answer-logit differences (LD) behave as monotonic readouts of latent decision variables in perceptual and memory tasks. Using SDC's qualitative signatures, including the correct/error folded-X pattern, the authors evaluated three multimodal non-reasoning models and one reasoning model across four tasks. Results showed LD satisfied SDC predictions in simpler tasks but only partially in complex visual reasoning, revealing the framework's current limitations without explicit normative process models. This work bridges confidence analysis in biological and artificial intelligence systems.
statistical decision confidenceanswer-logit differencelatent decision variablemultimodal language modelsnormative framework
ARDepth: Auto-regressive Monocular Depth Estimation with Progressive Visual Conditioning
ARDepth introduces auto-regressive monocular depth estimation with progressive visual conditioning, addressing limitations of diffusion models in capturing piecewise scene geometry. The method employs Scale-Progressive Conditioning (SPC) to inject multi-scale visual features and Semantic-Aware Guidance (SAG) to incorporate scene-level semantic priors, enabling hierarchical depth construction. This approach progressively builds depth representations as spatial resolution increases, balancing fine-grained local details with global structural consistency. Empirical results demonstrate strong performance and scale-consistent depth predictions, validating auto-regressive generation as an effective paradigm for geometric modeling.
monocular depth estimationauto-regressive generationscale-progressive conditioningsemantic-aware guidancegeometric modeling
Accepted Prefixes Are Not All You Need: A Negative Result on PEFT-Based Block-Diffusion Drafting
The paper presents PEFT-BD, a parameter-efficient speculative decoding method using a LoRA-like adapter as a block-diffusion drafter for autoregressive verification, motivated by advantages like tokenizer matching and minimal parameter overhead. Despite achieving nontrivial accepted prefixes in Qwen3-0.6B experiments, PEFT-BD fails to yield practical speedup due to compute inefficiency: each speculative step requires full-backbone passes for both drafting and verification. The results highlight that drafters must be computationally cheaper than verifiers for successful speculative decoding, irrespective of accepted prefix length.
speculative decodingpeftblock-diffusionautoregressive verificationqwen3-0.6b
Isolation as a First-Class Principle for LLM-Agent System Safety: Concepts, Taxonomy, Challenges and Future Directions
The paper proposes isolation as a first-class principle for LLM-agent system safety, addressing fragmentation in current literature across attack types, applications, and benchmarks. It introduces a boundary-centric taxonomy comprising five boundaries: user-agent, agent-tool, agent-execution, agent-agent, and system-environment. This taxonomy systematically identifies where isolation loss occurs, how compromises propagate, and relevant defenses at each interface. The authors summarize cross-boundary failure paths, discuss open challenges, and outline a research agenda for isolation-by-construction in future agent systems.
isolationllm-agentboundary-centric taxonomycross-boundary failureisolation-by-construction
Critic Experience Bank: Self-Evolving Step-Level Confidence Estimation for LLM Agents
The paper introduces Critic Experience Bank (CEB), a self-evolving framework for step-level confidence estimation in LLM agents, addressing the challenge of assessing action productivity before execution. CEB employs a hindsight LLM to pseudo-label past actions based on observed consequences, storing these in a memory bank for retrieval during similar future steps. Evaluated across three agent benchmarks and critic backbones, CEB achieves superior calibration (54% ECE reduction) and ranking (AUC) without training or ground truth labels.
llm agentsconfidence estimationcalibrationhindsight learningmemory bank
PM-Bench: Evaluating Prospective Memory in LLM Agents
The authors introduce PM-Bench, a text-based benchmark for evaluating prospective memory in LLM agents, inspired by cognitive science's Virtual Week paradigm. The benchmark assesses agents' ability to maintain intentions, execute deferred tasks, and monitor environment changes during a simulated seven-day week. Testing eight state-of-the-art LLMs under various configurations reveals significant challenges, with the top-performing GPT-5.4 agent achieving only 65.1% F1 score, indicating no dominant improvement strategy across models.
prospective memoryllm agentsvirtual weekbenchmarkintention execution
Demonstration of the common dual-channel feature decoupling characteristic of front-door mediation causal inference methods in whole-slice image classification
The study demonstrates that front-door mediation causal inference methods in whole-slide image (WSI) classification exhibit a dual-channel feature decoupling characteristic. The authors propose and validate two hypotheses: (1) causal inference multi-instance learning (MIL) introduces an independent classification channel that enhances WSI classification, and (2) greater feature divergence between new and baseline channels improves elimination of false correlations. Experiments on breast cancer and non-small cell lung cancer datasets confirm these hypotheses, showing that parallel independent channels increase feature diversity to reduce false associations between diagnostic and non-diagnostic sub-images.
causal inferencemulti-instance learningwhole-slide imagefeature decouplingfront-door mediation
IQA-T1: Tool-based Visual Evidence Reasoning for Image Quality Assessment
IQA-T1 introduces a tool-based visual evidence reasoning framework for image quality assessment (IQA), addressing limitations in generalization and interpretability of multimodal large language models (MLLMs). The method autonomously invokes specialized analysis tools to generate structured visual evidence (e.g., noise residual maps, gradient statistics) and integrates them into progressive reasoning. Evaluated on seven IQA benchmarks, IQA-T1 achieves state-of-the-art performance while providing interpretable, evidence-grounded assessments, supported by the Q-Tool dataset containing 11k multimodal reasoning chains.
image quality assessmentmultimodal reasoningvisual evidenceperceptual degradationstool-based framework
Lost in Visual Translation: A VLM-Assisted Perceptual-Semantic Coherence Framework for EEG-to-Image Reconstruction
The paper introduces a BCI-aware evaluation framework for EEG-to-image reconstruction that distinguishes perceptual and semantic coherence. Analyzing 6,855 image pairs from four datasets (ATM, ENIGMA, BrainVis, DreamDiffusion), the authors demonstrate that traditional metrics (SSIM, LPIPS, CLIP) poorly correlate with semantic consistency. Their solution employs four vision-language models (VLMs) to assess image pairs via structured questions, generating Tolerant Perceptual Alignment Scores (T-PAS) and Tolerant Semantic Alignment Scores (T-SAS), distilled into a BCI-Coherence Score (BCS). BCS achieves MAEs of 0.079 (T-PAS) and 0.082 (T-SAS), with human validation showing high reliability (Cohen's kappa = 0.882, Krippendorff's alpha = 0.882).
eeg-to-image reconstructionperceptual-semantic coherencevision-language modelstolerant alignment scoresbci-coherence score
How Many Tasks Are Enough for Agent Benchmark Decisions? A Replay Analysis of Public LLM Agent Benchmarks
The study evaluates how many tasks are sufficient for reliable agent benchmarking by analyzing partial runs of completed benchmarks. Using task-level records from SWE-bench, AppWorld, and tau-bench, the authors define criteria for partial runs to match full benchmark conclusions, including task group coverage and unresolved comparisons. Results show varying task fraction requirements: AppWorld meets targets at 15%, tau-bench at 25%, and SWE-bench Verified at 90%, while SWE-bench Lite fails to meet targets by 95%. The paper advocates for transparent reporting of evaluation parameters in partial assessments.
agent benchmarkingpartial evaluationtask fractioncoverage ruledecision rule
Evaluating Health Misinformation in Low-Resource Languages: Integrating Small Language Models with a Culturally-Sensitive Responsible NLP Framework (Bangla as a Case Study)
The study proposes a culturally-sensitive Responsible NLP framework for health misinformation detection in low-resource languages, using Bangla as a case study. It evaluates Small Language Models (SLMs) on a translated health misinformation dataset, identifying Phi-4 as optimal for claim extraction due to its precision-recall balance. The framework integrates cultural sensitivity, harm potential, and communication quality to address LLM limitations in low-resource contexts.
small language modelsresponsible nlphealth misinformationlow-resource languagescultural sensitivity
LakeQuest: A Three-Domain Benchmark for Grounded Question Answering across Data Lakes
LakeQuest introduces a 9,846-question benchmark for evaluating end-to-end question answering (QA) systems across heterogeneous data lakes, spanning AI/ML metadata, retail banking, and biomedical domains. The benchmark provides modality-aware evidence pointers to isolate retrieval from synthesis, exposing system limitations in real-world scenarios. Baseline evaluations of Retrieval-Augmented Generation (RAG) and agentic methods reveal persistent challenges in relation chaining, policy grounding, and cross-file reasoning, with retrieval quality not ensuring correct synthesis.
question answeringdata lakesretrieval-augmented generationmodality-awarerelation chaining
A Comparative Analysis of Institutional and Course Generative AI Policies within Higher Education: Implications for Instruction in Computing Education
This study bridges a research gap by comparing institutional and course-level generative AI (GenAI) policies in U.S. research-intensive higher education, focusing on computer science education. Through secondary analysis of institutional guidelines and course syllabi, the authors find a discrepancy: while institutional policies generally encourage GenAI adoption, course-level implementation remains cautious. The paper proposes an instructor-centered framework to guide future GenAI integration. Findings highlight the tension between broad institutional directives and practical pedagogical concerns in computing education.
generative aicomputer science educationinstitutional policiescourse syllabipedagogical framework
A Longitudinal Analysis of Public Discourse on AI Ethics in Education Using Twitter Data
This longitudinal study analyzes five years (2019-2024) of Twitter discourse on AI ethics in education, focusing on ChatGPT's release as a pivotal moment. Using BERT-based topic modeling and SetFit sentiment analysis, the research identifies dominant themes and tracks sentiment shifts. Results show predominantly positive sentiment, with negative spikes around ethical controversies, and recent concerns about academic integrity and generative AI. Public discourse appears pragmatic, favoring AI integration while demanding ethical oversight and institutional accountability.
longitudinal analysistopic modelingsentiment analysisgenerative aiacademic integrity
The Sound of Absence: Audio-Language Embedding Models Struggle with Negation
The study exposes a critical limitation in audio-language embedding models like CLAP: their inability to encode negated sound concepts, mapping affirmative and negated captions to nearly identical representations. The authors introduce NegEval-Audio, a framework converting existing datasets (AudioCaps, Clotho) into two negation-aware tasks (Retrieval-Neg, MCQ-Neg) to evaluate this failure. Results show sharp performance degradation under negation, with MCQ-Neg accuracy falling below chance, persisting even in recent multimodal LLM-based models. A training-free steering method improves MCQ-Neg marginally but fails on Retrieval-Neg, indicating affirmation bias as a fundamental flaw in representation geometry requiring explicit negation-aware training.
audio-language embeddingnegation-aware evaluationrepresentation geometrymultimodal llmaffirmation bias
Code-MUE: Measuring Code LLMs' Uncertainty through Execution-based Semantic Interaction Graphs
Code-MUE introduces a black-box framework for measuring uncertainty in Code LLMs through execution-based Semantic Interaction Graphs, addressing the syntax-semantics gap in existing methods. The approach quantifies global semantic diversity using Von Neumann entropy of the solution space, grounded in observable runtime behavior rather than superficial textual similarity. Empirical evaluation across eight state-of-the-art LLMs shows strong negative correlation with functional correctness (Spearman's ρ up to -0.98), outperforming lexical and embedding-based baselines.
code llmsuncertainty estimationsemantic interaction graphsvon neumann entropyexecution-based analysis
Track, Rank, Crack: Epistemic Working Memory Scales Multi-Hop Reasoning in Language Agents
SLEUTH introduces structured epistemic working memory to address context dilution in multi-hop reasoning by language agents, explicitly maintaining Confirmed Facts, Active Hypotheses, and Open Questions. This approach scales reasoning performance across five benchmarks, showing increasing advantages with task difficulty (+5 points on HotpotQA to +11 on 4-hop chains). Analysis reveals the evidence sufficiency problem, where agents fail to commit to correct answers despite sufficient evidence. A lightweight commitment trigger improves performance only when paired with structured state, demonstrating that reasoning organization, not raw model capability, is critical for scaling multi-hop tasks.
epistemic working memorycontext dilutionmulti-hop reasoningevidence sufficiencycommitment trigger
On-Device Deep Research at 4B: Exposure Bounds Faithfulness, Retrieval Bounds Coverage
This study isolates and measures two distinct factors influencing citation quality in on-device research agents: cited claim faithfulness and trustworthy coverage. Using a 4B parameter generator on a 24 GB laptop, the authors vary source exposure (400 vs. 1500 characters) and source quality (gold vs. retrieved papers). Results show that exposure primarily determines faithfulness, increasing it from 0.45 to 0.58 for retrieved sources and 0.37 to 0.58 for gold sources, while retrieval quality governs coverage, which remains near 0.22 due to fixed recall. The findings suggest prioritizing exposure increases before addressing retrieval recall for optimal performance.
on-device researchcited claim faithfulnesstrustworthy coveragesource exposureretrieval recall
Fin-Analyst at FinMMEval 2026 Task 3: A Live Hybrid Trading Agent with LLM Specialists and Rule-Based Signals
The paper introduces Fin-Analyst, a hybrid trading agent combining eight LLM specialists with rule-based signals, achieving top performance in FinMMEval 2026 Task 3. The system processes diverse data sources (news, SEC filings, technical indicators) through a Meta-Agent for Tesla (TSLA) and a three-signal voting mechanism for Bitcoin (BTC). On TSLA, it yielded +13.51% return (Sharpe 4.10, 88% win rate), outperforming Buy-and-Hold by 28.33 points, while BTC performance remained flat. Ablation revealed 8-K disclosures as the most impactful TSLA signal, and error analysis highlighted limitations of memoryless agents and fixed-threshold rules.
llm specialistsmeta-agentsharpe ratio8-k disclosuresrule-based signals
Rethinking the Evaluation of Harness Evolution for Agents
The paper critically evaluates automatic harness evolution for LLM agents, identifying two methodological flaws in current practices: unfair comparison with test-time scaling baselines and potential overfitting due to shared benchmarks. The authors propose a revised evaluation protocol comparing harness evolution with simple baselines under matched feedback and inference budgets, while also testing generalization on held-out tasks. Experiments on Terminal-Bench 2.1 using GPT-5.4 and Claude Opus 4.6 reveal that harness evolution fails to consistently outperform test-time scaling and shows limited generalization, questioning its effectiveness and advocating for fairer evaluation protocols.
harness evolutionllm agentstest-time scalinggeneralizationevaluation protocol
Partial Identification with Multiple Nonlinear Measurements of a Latent Regressor
The paper addresses partial identification in linear regression with latent regressors observed through multiple noisy nonlinear measurements, a problem exemplified by occupational AI exposure measurement. It proposes a method to fix the latent scale by enforcing linearity in the consensus measurement function and bounding curvature heterogeneity across sources. The structural coefficient is shown to lie in a closed-form interval centered at a symmetric cross-source estimator, with estimable bounds using at least four measurements. Applied to 8.88 million person-year observations, the method reconciles six exposure measures, yielding a loading-invariant consensus coefficient of -0.239 with a partial-identification half-width of 1.23%.
latent regressorpartial identificationcurvature heterogeneityloading-invariantcross-source estimator
Good Benchmarks
The article proposes five key criteria for effective benchmark design in AI research: correctness, solvability, verifiability, clear specification, and meaningful difficulty. It emphasizes that ideal benchmarks should reflect real-world problems recognized by domain experts, use practitioner-relevant terminology, and validate outcomes rather than specific solution approaches. The work provides a normative framework for evaluating task quality in machine learning experimentation without prescribing specific implementations.
benchmark designtask evaluationverifiabilityproblem specificationpractitioner relevance
RCWT: Measuring Task-Budget Displacement from Coordination Content in LLM Calls
The paper introduces Roundtable Context Window Test (RCWT), a protocol for measuring task-budget displacement in LLM systems where coordination content competes with task instructions under fixed context budgets. RCWT varies coordination content while controlling budget, position order, and task family, revealing that models degrade sharply when residual task evidence falls below a few hundred tokens. Experiments with GPT-4.1-mini, Claude Haiku 4.5, and Gemini 2.5 Flash show intact-task performance remains high even at 95% coordination ratios, indicating the cliff effect stems from budget displacement rather than semantic interference. RCWT serves as a measurement primitive for context-allocation analysis.
task-budget displacementcontext windowmulti-agent systemscoordination contentllm systems
The Benjamini--Hochberg Procedure Can Fail to Control the FDR for Correlated Two-Sided Gaussian Tests
The work disproves a long-standing conjecture by demonstrating that the Benjamini-Hochberg (BH) procedure fails to control the false discovery rate (FDR) for correlated two-sided Gaussian p-values. Using a factor model and interval-arithmetic certification at α=0.01, the authors prove FDR exceeds 0.0104 asymptotically, with Monte Carlo experiments validating the theoretical result. Notably, the proof was generated by GPT-5.6 Pro and verified manually.
false discovery ratebenjamini-hochberg proceduregaussian p-valuesfactor modelinterval-arithmetic
A Threshold Exceedance Framework for CBRN Uplift Evaluation in Frontier Language Models
The paper introduces a Threshold Exceedance Criteria (TEC) framework for evaluating whether frontier language models materially uplift non-expert actors' ability to plan Chemical, Biological, Radiological, or Nuclear (CBRN) misuse compared to public tools. The framework decomposes uplift studies into components: participant eligibility, threat scope definition, and statistical uplift estimation. An empirical study operationalized TEC, distinguishing generative (plan creation) and revisionist (plan refinement) uplift. Results showed domain heterogeneity, with material uplift confirmed only in radiological scenarios, informing mitigation strategies. The study emphasizes prespecified criteria, explicit baselines, and separating uplift types.
threshold exceedance criteriacbrn misusegenerative upliftrevisionist upliftfrontier language models
Comparing Semantic Navigation in Humans and Large Language Models using Natural Language Processing
The study compares semantic navigation patterns between humans and LLMs (GPT-4o, Gemini-2.5-Pro, Claude-Sonnet-4.5) using verbal fluency tasks analyzed via trajectory-based NLP metrics. Researchers quantified entropy, distance to next, and distance to centroid across 82 human participants and LLM outputs at eight temperature settings. Humans showed higher entropy, larger semantic steps, and broader dispersion than all LLMs, indicating more exploratory search; no temperature setting fully replicated the human profile across all dimensions, suggesting distinct human exploitation-exploration tradeoffs in semantic memory retrieval.
semantic navigationverbal fluencytrajectory-based metricsexploration-exploitationtemperature tuning
Cost-Governed RAG: Unified Per-Tenant Cost Attribution Across Retrieval and Generation in Multi-Tenant LLM Systems
The paper introduces Cost-Governed RAG, a multi-tenant LLM architecture enabling unified cost attribution across retrieval and generation components. The system combines a codebook-oblivious vector index (TurboVec) with an LLM governance gateway to achieve deterministic per-tenant cost calculation for embeddings, retrieval, and generation. Evaluated on Snowpark Container Services with 100 simulated tenants (10M vectors), the solution demonstrates 99.96% cost attribution accuracy, 0.04% telemetry overhead, and 3.1-9.0x retrieval cost reduction versus managed vector databases.
retrieval-augmented generationmulti-tenant llmcodebook-oblivious quantizationcost attributionvector index
TRAIL: A Platform for Configurable Human--AI Teaming Experiments
The Team Research and AI Integration Lab (TRAIL) introduces a web platform for configurable human--AI teaming experiments, enabling reproducible studies of AI teammate design properties (e.g., personality, communication style) on team dynamics. TRAIL features a Big Five persona system, selective-participation message pipeline, dual memory architecture, and longitudinal experiment chaining, with export-ready analytics. In a six-session classroom deployment (N=51), TRAIL maintained AI conversation share below 50% and revealed a double dissociation: cognitive-scaffolding agents improved contribution ratings and linguistic alignment, while socially-supportive agents enhanced team climate and reduced over-reliance.
human-ai teamingbig five personaselective-participation pipelinedual memory architecturelongitudinal chaining
The Emerging Paradigm of Geospatial Foundation Models: From Pre-Training to Agentic Reasoning
The paper introduces Geospatial Foundation Models (GeoFMs), AI/ML models pre-trained on extensive geospatial data, enabling a paradigm shift where large-scale pretraining is decoupled from domain-specific fine-tuning. It distinguishes between self-supervised vision models (e.g., masked auto-encoding) and contrastive vision-language models (e.g., open-vocabulary analysis), detailing their respective capabilities. The authors propose a taxonomy for model adaptation strategies and a framework for cost-effective operationalization. The work concludes with a vision for Agentic Geospatial Reasoning, where LLMs orchestrate GeoFMs to automate complex analytical workflows, advancing from perception to cognition.
geospatial foundation modelsmasked auto-encodingcontrastive learningagentic reasoningmlops
From Reconstruction to Interpretation: Zero-Setup Multi-Phase Segmentation of X-ray Tomography Data
The authors present a zero-setup framework for multi-phase segmentation of synchrotron X-ray tomography data, eliminating the need for manual thresholding or dataset-specific training. The method combines material-agnostic mask preparation with a pretrained semantic segmentation network, generating interpretable masks (background, sample, bright, dark-gray, light-gray, porosity) without user input. Evaluation on held-out micro-CT images shows consistent segmentations across varying samples, outperforming intensity-based thresholding and enabling near-real-time beamline feedback.
x-ray tomographyzero-shot segmentationsemantic segmentationmicro-ctbeamline feedback
Self-Consistent Flow: Unifying Velocity and Endpoint Prediction for Rectified Flow Models
Self-Consistent Flow (SC-Flow) unifies velocity and endpoint prediction in rectified-flow-based generative models, addressing limitations of each parameterization. SC-Flow employs a lightweight consistency loss to jointly train a single network for both predictions, improving performance without major architectural changes or significant computational overhead. Experiments on image generation demonstrate that SC-Flow stabilizes optimization, enhances the straightness of generation paths, and significantly improves generation quality over standard rectified-flow baselines.
rectified-flowvelocity predictionendpoint predictionconsistency lossgeneration quality
Connected by Construction: Learning Tractable Near-Tour Marginals for Traveling Salesman Problems
The paper introduces C2TSP, an unsupervised end-to-end pipeline for learning tractable near-tour marginals in the Traveling Salesman Problem (TSP) by directly modeling Hamiltonian structure. The method leverages a connected-by-construction rooted 1-tree Gibbs family, learning residual edge perturbations via implicit differentiation. Structural correction is achieved through a smoothed Held-Karp layer for degree balance and certificate-guided sharpening to enhance tour-like properties. Experiments demonstrate that C2TSP improves both tour cost and structural interpretability, with ablations confirming the joint effectiveness of edge perturbation and sharpening.
traveling salesman problemhamiltonian structure1-tree gibbs familyheld-karp layercertificate-guided sharpening
GaitSpan: Growing Humanoid Locomotion from Walking to Running
GaitSpan introduces a framework for expanding a pretrained walking policy into faster locomotion regimes, enabling a single policy to handle walking, jogging, and running across continuous speeds and terrains. The method comprises three components: rhythm generation, which modulates the walking policy with internal clocks; stride shaping, which optimizes dynamic locomotion patterns using a spring-loaded inverted pendulum objective; and residual adaptation, which captures motion details not addressed by the other components. GaitSpan outperforms multi-expert and imitation-based baselines in learning speed and gait performance, demonstrating zero-shot transfer across morphologies and terrains.
rhythm generationstride shapingresidual adaptationspring-loaded inverted pendulumzero-shot transfer
Toward Trustworthy Autonomous Science: A Two-Year Community Roadmap
The updated AISLE roadmap advances autonomous science by addressing trust and verification challenges, expanding from five to seven critical dimensions. It evaluates 14 original milestones (achieved, partial, reframed, or open) and introduces four new ones (M15-M18). Key developments include multi-agent systems generating validated hypotheses, interoperable self-driving laboratories, and domain foundation models. The two-year plan prioritizes interfaces and verification scaffolding in year one, followed by federation and zero-trust governance in year two. The grassroots network remains central for connecting national, international, and commercial initiatives without siloing.
autonomous laboratoriesmulti-agent systemsfoundation modelszero-trust coordinationself-driving laboratories
Continual Learning with Elastic Regularization and Synthetic Replay for Federated MLLM Fine-Tuning
The paper proposes FedCMM, a federated continual learning framework for Multimodal Large Language Models (MLLMs) that combats catastrophic forgetting through three modality-aware mechanisms. First, elastic weight consolidation employs separate Fisher matrices for vision, language, and cross-modal components. Second, clients use lightweight generative modules to create embedding-level synthetic replays without raw data sharing. Third, task-similarity-aware gradient aggregation filters conflicting updates via cosine similarity. Evaluations on two benchmarks show FedCMM outperforms baselines in accuracy and backward transfer, demonstrating robust adaptation in federated settings.
federated learningmultimodal llmselastic weight consolidationgenerative replaygradient aggregation
PFAdapter: Hierarchical LoRA Decomposition for Personalized Federated MLLMs
PFAdapter introduces hierarchical LoRA decomposition for personalized federated MLLMs, separating adapter parameters into global-shared (query/key) and local-private (value/output) components to balance global knowledge aggregation with local adaptation. The method employs orthogonality regularization via Frobenius norm to prevent feature redundancy and selective aggregation to reduce communication costs by ~50%. Evaluations on VQA-RAD, SLAKE, Hateful Memes, and CrisisMMD show accuracy improvements of 2.4-4.8% over baselines, demonstrating efficacy for edge AI deployment.
federated learningmultimodal llmslora decompositionorthogonality regularizationedge intelligence
Representing and Generating Levels Over Time through Playtrace Reconstructive Partitioning
The paper introduces a novel 'cake' representation for dynamic game levels and proposes Playtrace Reconstructive Partitioning (PRP), a domain-agnostic level generation method. The cake representation implicitly encodes temporal dynamics, while PRP reconstructs levels from playtraces via partitioning. Evaluated on Sokoban against six PCG baselines, PRP generates valid levels without compromising solution diversity. The approach demonstrates that temporal encoding in level representations can enhance generation quality across domains.
procedural content generationdynamic representationplaytrace partitioningdomain-agnosticsokoban
Sparse Autoencoders for Interpretable Out-of-Distribution Detection
The paper proposes a sparse autoencoder (SAE)-based method for interpretable out-of-distribution (OOD) detection by analyzing intermediate network activations. It identifies distinct sparse feature activation patterns between in-distribution (ID) and OOD data, then computes an OOD score via cosine similarity between test sample activations and class-mean ID activations. The approach achieves state-of-the-art performance on standard benchmarks while providing interpretable insights into distribution shift effects on representations.
sparse autoencodersout-of-distribution detectioninterpretabilitycosine similarityintermediate activations
Operationalising Multi-Dimensional Evaluation for Conversational Agents: A Scalable, Governed Pipeline with Selective Re-evaluation and Model Benchmarking
The paper introduces GenAI Evaluation, a governed pipeline for multi-dimensional assessment of retail conversational agents, addressing challenges in LLM-as-a-judge deployments. The framework processes production logs via normalization, sharding, and schema-constrained LLM scoring, evaluating dimensions like helpfulness, truthfulness, and tone alignment. It features selective re-evaluation, schema locking, and versioned configurations for auditability. Validated on 12,980 human-labeled records, the pipeline achieved 0.93 macro F1 and 89% human-acceptability accuracy for translation, processing 50,000 records daily with over two million interactions evaluated.
conversational agentsllm-as-a-judgeschema-constrained scoringselective re-evaluationmulti-dimensional evaluation
Graph Feedback Controls Consensus and Clique Formation in Open-Weight Language-Model Populations
The study investigates convention formation in open-weight language-model populations (1.1B-32B parameters) using a naming-game protocol, focusing on graph feedback controls and consensus dynamics. By analyzing prompt-conditioned score-state distributions and constructing state-similarity graphs, the authors separate label agreement from latent state-space consensus. Results show that homophilous threshold-similarity routing amplifies fragmentation, while bridge-seeking routing repairs fragmentation when memory is available. In mixed-model grids, state-component and label-disagreement bridges recover behavioral consensus in 14/18 retained-memory runs, whereas threshold-similarity fails to achieve consensus in 189 settings. Homogeneous populations, particularly Qwen2.5-32B, shift toward consensus with retained history. Early-window graph-energy features provide diagnostic utility.
open-weightstate-similarityhomophilousbridge-seekinggraph-energy
Calibrated Selective Prediction Using Deep Ensembles for ROI-Based Thyroid Nodule Ultrasound Classification Under Dataset Shift: A Retrospective Evaluation
The study proposes a calibrated deep ensemble framework for selective prediction in thyroid nodule ultrasound classification under dataset shift, emphasizing clinical reliability. Using ConvNeXt-Tiny with squeeze-and-excitation attention, a five-member ensemble was trained on TN5000 with five-fold cross-validation, member-wise vector-scaling calibration, and fold-specific threshold selection. Evaluated on TN5000, the ensemble achieved AUC-ROC 0.9395, AP 0.9715, and ECE 0.0088, with 98.3% NPV for No-FNA suggestions. On TN3K, performance degraded (AUC-ROC 0.7870, ECE 0.1899), highlighting limited external threshold transportability. The framework demonstrated potential for selective triage but requires recalibration and validation for clinical deployment.
deep ensembledataset shiftselective predictionconvnext-tinyvector-scaling calibration
Enabling 24-hour Agricultural Robotics: Unsupervised Day-to-Night Cross-Modal Image Translation for Nighttime Visual Navigation
The authors propose an unsupervised image translation framework enabling 24-hour agricultural robotics by converting daytime RGB plant-row images into near-infrared nighttime counterparts without pixel-level supervision. The method leverages a pre-trained CLIP model to maintain semantic consistency during translation and introduces a visibility mask to address NIR illumination range limitations. Evaluations on AgriNight—a novel dataset of 428 daytime and 549 nighttime images—demonstrate superior image quality and improved downstream semantic segmentation performance compared to state-of-the-art baselines. Real-time autonomous navigation experiments validate the framework's effectiveness for nighttime agricultural visual navigation.
unsupervised image translationnear-infraredsemantic consistencyvisibility maskagricultural robotics
AutoTrace: From Patches to Triggers via Agentic Interprocedural Exploration
AutoTrace introduces an agentic pipeline for vulnerability trigger localization that performs interprocedural causal reasoning via LLM-guided exploration of code property graphs, with deterministic admissibility gates ensuring grounded evidence. The method achieves 75.0% VulnHit and 80.8% FuncHit on InterPVD, surpassing prior work. The authors also release SinkTrace-Bench, a 1,542-sample dataset of verified source-to-sink chains, revealing LLMs' limitations in causal reasoning for vulnerability analysis.
trigger localizationcode property graphinterprocedural analysisllm agentsvulnerability detection
Designing Agent-Ready Websites for AI Web Agents: A Framework for Machine Readability, Actionability, and Decision Reliability
The paper introduces an agent-ready website framework for optimizing e-commerce platforms for AI web agents, addressing machine readability, actionability, and decision reliability. The framework evaluates agent interpretability, executability, and decision reliability through features like semantic clarity and contextual signals. In a controlled experiment with GPT-4.1, Gemini-2.5 Flash, and Grok-4 Fast (300 runs), the agent-ready version achieved 89.3% strict success rate versus 49.3% for the baseline, reducing step counts from 9.31 to 6.49 and improving task completion in product comparison and multi-constraint selection.
agent-ready websitemachine readabilityactionabilitydecision reliabilitye-commerce optimization
Representation and Reference Selection in Training-Free Synthetic Image Attribution
The paper analyzes the interaction between representation space and reference selection in training-free synthetic image attribution, where generators are identified using source-specific references without retraining. Using CLIP and DINOv2 representations across layers and three reference selection methods (arbitrary, semantically aligned, resynthesis-based), the study finds that attribution accuracy peaks at intermediate representation levels, balancing source-discriminative cues and semantic abstraction. Semantically constrained references reduce mismatch and improve accuracy, with resynthesis excelling in low-reference regimes and aligned references offering better trade-offs for moderate-sized pools. The work highlights the interplay of representation depth, reference construction, and budget.
synthetic image attributionrepresentation spacereference selectionclipdinov2
An Empirical Analysis of Continual Learning for Heterogeneous Medical Visual Question Answering
This paper conducts a systematic empirical evaluation of continual learning (CL) methods for heterogeneous medical visual question answering (MedVQA) tasks, addressing catastrophic forgetting, task ordering sensitivity, and low-rank adaptation parameter evolution. The study examines CL performance across diverse clinical objectives including classification, multi-label classification, detection, cell counting, and report generation. Results indicate that existing CL methods fail to maintain stability-plasticity balance when interleaving tasks with varying objectives and supervision formats. The analysis reveals patterns of weight drift under different CL approaches, highlighting challenges in adapting medical vision-language models to sequential heterogeneous tasks.
continual learningmedical visual question answeringcatastrophic forgettinglow-rank adaptationstability-plasticity balance
Invariant Learning Dynamics of Transformers in Inductive Reasoning Tasks
The authors present a theoretical framework analyzing Transformer learning dynamics in inductive reasoning tasks, unifying synthetic tasks like in-context n-grams and multi-hop reasoning. They prove that attention models' training dynamics converge to a low-dimensional invariant manifold, enabling interpretable analysis through a reduced coordinate system. Key findings include characterizing data-driven competition between in-context and in-weights learning, initialization-dependent circuit selection, and automated circuit detection via manifold coordinates, advancing a predictive theory of Transformer learning.
transformerinductive reasoninginvariant manifoldlearning dynamicsin-context learning
HPC-Enabled Video-based Coastal Wave Parameter Estimation Using V-JEPA and Deep Spatiotemporal Learning
The paper presents an HPC-accelerated deep learning framework for estimating five coastal wave parameters (Hs, Hmax, Tp, Tz, θ) from monocular video, addressing limitations of in-situ sensors. The architecture combines a self-supervised V-JEPA ViT Small backbone for spatiotemporal feature extraction, a dual-stream SlowFast temporal encoder, Farneback optical flow for hydrodynamic saliency, and a dispersion-constrained multi-task regression layer. Trained on an NVIDIA DGX A100 cluster with only 6 annotated scenes, the model achieved Pearson correlations of 0.451-0.832 across parameters, demonstrating feasibility despite data limitations (max R²=0.246).
v-jepaslowfastfarneback optical flowspatiotemporal learningwave parameter estimation
Are we Merging the Right Models? Impact of Expert Training Duration on Model Merging for LLMs
The study challenges conventional model merging practices by demonstrating that training duration of domain experts significantly impacts merged model performance, with method-dependent effects. Researchers systematically varied expert training durations (25-500% of optimal steps) across five domains (Math, Code, Instruction Following, Multilingual, Safety) and three Qwen model sizes (0.8B, 2B, 4B), evaluating five merging methods. Key findings show simple averaging degrades with overfitting while sparsification-based methods peak post-optimum, supported by bias-variance analysis drawing parallels to random forests. Results indicate training duration and merging method must be co-optimized.
model mergingtraining durationsparsificationbias-variance decompositionmulti-task learning
Mitigating The Effect of Class Imbalance in Data with Hierarchical and Dependable Structure
The paper proposes a Hierarchy-Aware RoBERTa framework for classifying cybersecurity vulnerabilities using the Common Weakness Enumeration (CWE) taxonomy, addressing extreme class imbalance and hierarchical dependencies. The method incorporates learnable parent-class embeddings to preserve taxonomic consistency, avoiding synthetic oversampling techniques that violate hierarchical constraints. Evaluated on the CWE Research Concept dataset, the model achieves a 0.76 weighted F1-score without augmentation, outperforming baselines with notable gains on minority classes (e.g., Class category F1-score improved from 0.40 to 0.60 over BERT).
class imbalancehierarchical classificationcwe taxonomyrobertasynthetic oversampling
Sparse Inter-Layer Dependencies of Transformer FFN Neurons
The paper introduces a training-free attribution method to analyze sparse inter-layer dependencies in Transformer FFN neurons, revealing that neuron activations can be explained by small subsets of preceding activations and attention outputs. The method estimates upstream influences on target neurons by masking non-essential inputs with average values. Experiments show high fidelity in preserving activations with sparse subsets, even when propagating deviations through the network, with minimal perplexity impact. Results indicate FFNs exhibit structured sparsity despite dense parameterization, offering insights for circuit-level interpretability and efficient inference.
transformerfeedforward networksparse dependenciesneuron attributioninterpretability
Removable Defects: The Economics and Limits of Deliberate Deficiency
The paper proposes treating system deficiencies as design variables rather than costs, introducing a framework where defects are retained for economic benefit and compensated only when critical. It formulates an advantage condition for economically viable defect retention, modeled using Ehrlich-Becker insurance economics and Townsend verification. Key results include: a two-sided removability characterization showing detector-premium relationships under multiplicative dynamics, differentiation between observation vs. capacity defects via distributional access, and learnable detectors with linear training costs. The work unifies Chow's reject option, Kelly growth, and selective prediction.
deliberate deficiencycostly-state-verificationmultiplicative dynamicsobservation defectscapacity defects
FAIR GraphRAG: A Retrieval-Augmented Generation Approach for Semantic Data Analysis
FAIR GraphRAG introduces a retrieval-augmented generation framework that integrates FAIR Digital Objects (FDOs) into graph-based retrieval systems to enhance domain-specific question answering. The method represents each graph node as an FDO with structured metadata, persistent identifiers, and semantic links, using LLMs for schema construction and automated content extraction. Evaluated on biomedical RNA-sequencing data, FAIR GraphRAG improves accuracy, coverage, and explainability for complex queries involving metadata and ontology links, demonstrating feasibility for specialized domains like medicine.
retrieval-augmented generationknowledge graphsfair digital objectssemantic relationshipsontology links
PRISM Edit: One Vector for All Temporal Answers
PRISM Edit introduces a novel model editing approach for temporal facts in LLMs, optimizing a single polysemous representation across time contexts without architectural changes. The method leverages causal tracing to identify a two-stage computation in LLMs (time-agnostic retrieval in early MLP layers, temporal modulation in later layers) and uses this pathway for routing. Evaluated on TimeConflict and temporally augmented CounterFact, PRISM Edit outperforms baselines by +23.3 TC and +33.7 CRS while being 2x faster.
model editingtemporal factscausal tracingpolysemous representationtemporal consistency
Bringing Back Rule Induction to Fluid Intelligence Research? An Initial Validation of the ARC-AGI Benchmark in Humans
The study validates the ARC-AGI benchmark as a measure of human fluid intelligence (gf) by examining its psychometric properties and nomological network in 100 participants. ARC-AGI, designed to assess rule induction, demonstrated good psychometric characteristics and a strong correlation (ρ= .63) with figural reasoning tests, but weak associations with figural originality. This supports its validity as a gf measure, contrasting dominant working memory-based approaches. The research advocates integrating AI benchmarks into human cognitive ability frameworks for interdisciplinary evaluation.
fluid intelligencerule inductionpsychometric propertiesarc-agi benchmarkfigural reasoning
Evaluating Nonuniform Dependability Across Response Conditions: A Conditional Generalizability Framework Illustrated in Automated Essay Scoring
The study introduces a conditional generalizability framework addressing nonuniform dependability in automated scoring systems, treating scoring configurations as admissible measurement conditions. It combines analytical D-study projections with empirical configuration sweeps, using entropy-based response stratification to evaluate design adequacy. Applied to automated essay scoring of timed L2 writing, the framework revealed high aggregate dependability (Φ ≈ 0.76) with modest declines across entropy strata (Φ = 0.88, 0.87, 0.84), indicating varying decision-study requirements.
conditional generalizabilityautomated essay scoringd-studyentropy stratificationmeasurement conditions
An Empirical Study for Android-to-OpenHarmony GUI Test Migration
The study presents the first systematic empirical evaluation of GUI test migration from Android to OpenHarmony, introducing the ATH Benchmark with 36 commercial apps (9B+ downloads) and 108 test cases. Researchers adapted two state-of-the-art approaches (ReSPlay, ITeM) for OpenHarmony, evaluating testing performance, failure root causes, and system-specific impacts. Results show low success rates (15% ReSPlay, 26% ITeM), attributing failures to OpenHarmony's architectural and ecosystem differences. An enhanced ITeM-HM approach incorporating system features achieved 214% relative improvement (81% success rate).
gui test migrationandroidopenharmonyempirical studytest automation
The Seriality Gap in Video Diffusion Models
The paper identifies a 'seriality gap' in video diffusion models, demonstrating their performance degradation on tasks requiring long causal chains of dependent events. Through controlled experiments on multi-ball hard-sphere dynamics, the authors show that bidirectional video diffusion models fail to scale with increasing event seriality, unlike length-matched single-ball controls. Intervention studies reveal that methods enhancing serial computation (autoregressive generation, architectural depth) improve performance disproportionately. Theoretical analysis proves denoising steps cannot add serial computation beyond the backbone architecture, highlighting a fundamental limitation for serial reasoning tasks.
video diffusion modelsseriality gaphard-sphere dynamicsdenoising stepsautoregressive generation
A Shortcut to Statistically Steady-State Turbulence with Flow Matching
The paper introduces GyroFlow, a latent generative model that directly estimates steady-state statistics of gyrokinetic turbulence in 5D phase space, bypassing the computationally expensive transient phase. The method leverages an ergodicity assumption and flow matching to generate saturated snapshots from noise, conditioned on dimensionless operating parameters. GyroFlow outperforms autoregressive, reduced-order, and other generative approaches, achieving substantial speedup while maintaining accuracy, as measured by the proposed FGyD metric and downstream flux accuracy. The model can also warm-start numerical codes.
gyrokinetic turbulenceflow matchinglatent generative modelsteady-state statistics5d phase space
The Spectrum Is Not Enough: When Context Helps Time-Series Forecasting
This work distinguishes between spectral predictability and the value of context in time-series forecasting, demonstrating that spectral indices fail to capture the utility of context, retrieval mechanisms, or pretrained models. The authors introduce surrogate pairs to isolate this distinction and propose a label-free diagnostic, the coverage deficit, which quantifies beyond-spectrum structure as the gain of analog over linear prediction. Empirical evaluation on seven benchmarks shows that retrieval-based methods collapse across surrogate pairs (ECL median +33% to -35%, p<10^-40), while spectral indices remain unchanged. The diagnostic predicts the sign of beyond-spectrum value, outperforming spectral indices in leave-one-dataset-out analysis.
spectral predictabilitytime-series forecastingcoverage deficitanalog predictionsurrogate pairs
Watermark Forensics for Generative Models: An Information-Theoretic Perspective
The paper introduces an information-theoretic framework for analyzing watermark forensics in generative models, formalizing a 'forensic ladder' that quantifies the cost of attribution, payload extraction, and localization in terms of sample length. It proposes an information profile ν(t) to measure token-level information leakage about a secret S, showing how its distribution enables different forensic tasks. The main theorem establishes tight bounds for multi-user attribution, proving Θ(log N/h) tokens are needed over stationary-ergodic sources with entropy rate h, with exact alignment. Experiments on GPT-2, Pythia-410M, and Qwen2.5 validate the theoretical predictions.
watermark forensicsinformation profileentropy ratemulti-user attributionstationary-ergodic source
Ensemble Controlled-Flow Filtering for Implicit Data Assimilation
The paper introduces implicit data assimilation and the Ensemble Controlled-flow Filter (EnCF), which estimates dynamical system states via energy tilting of forecast distributions. EnCF employs stochastic controlled flows with adjoint-derived controls, while EnCF-LF learns surrogate conditional energies for simulator-defined observations. Theoretical analysis proves ideal exactness, provides error decomposition, and shows non-accumulation of local errors under stability. Experiments demonstrate EnCF's superiority over Kalman filters for non-Gaussian, many-to-one, multimodal, and implicit observation models, while Kalman filters remain optimal for smooth additive-Gaussian cases.
implicit data assimilationensemble controlled-flow filterstochastic controlled flowadjoint matchingenergy tilt
Robustness of Deep Learning Models for PV Power Forecasting under NWP Forecast Errors: A Spatiotemporal and Physically Interpretable Analysis
The study introduces a physically constrained robustness evaluation framework for photovoltaic (PV) power forecasting models under realistic numerical weather prediction (NWP) errors, addressing temporal correlation and physical coupling of input uncertainties. Six models (PatchTST, GRU, N-HITS, LightGBM) are evaluated using virtual PV power and dynamic NWP perturbations with heteroscedasticity modulated by clear-sky conditions and Erbs reconstruction. Results show sequence models outperform tabular baselines in noise filtering and temporal resilience, with SHAP and Integrated Gradients revealing feature reallocation toward stable historical data and physical priors. A Pareto analysis links accuracy, robustness, and latency for engineering applications.
photovoltaic forecastingnumerical weather predictionheteroscedastic perturbationssequence modelsfeature reallocation
Efficient Sequential Calibration with $O(T^{2/3-ε})$ Error Bound
We present an efficient randomized forecaster achieving $O(T^{2/3-\varepsilon})$ expected calibration error for online binary sequential calibration, breaking the classical $T^{2/3}$ barrier. The method combines the SPR-Calibration procedure with a Blackwell-style correction layer, decomposing error into surrogate calibration error and residual discrepancy. SPR-Calibration controls surrogate sequence calibration, while the correction layer manages approximation error via quadratic potential arguments and SPR-Calibration's sparsity. Theoretical analysis shows the surrogate error is bounded by SPR-Calibration guarantees, and residual error is controlled through potential-based methods.
sequential calibrationblackwell-style correctionsurrogate sequencequadratic potentialsparsity
LatentFlow: A General Framework for Conditioning Stochastic Processes
LatentFlow introduces a training-free framework for conditioning stochastic processes by transforming them into deterministic mappings of tractable latent innovations. The method reduces process-level conditioning to latent-space inference: it pulls likelihoods back through a transformation $T_{\vartheta}$, samples the latent law using guided probability flow, and pushes samples forward. The approach is exact theoretically, with approximations arising only from finite terminal noising, Monte Carlo guidance, and time discretization. LatentFlow enables conditional sampling across diverse model classes—including spatial priors, stochastic PDEs, and neural processes—on a single CPU in seconds.
stochastic processesconditional samplingprobability flowlatent-space inferencereverse-time sde
Contrastive-Collapsed Loss for Flexible and Geometrically Optimal Embeddings and Faster Convergence
The paper introduces CoCo, a contrastive-collapsed loss function for learning normalized embeddings that simultaneously encourages intra-class collapse and inter-class contrast. The method provides geometrically optimal embeddings with large angular separation between classes, offering closer initialization to optimal configurations and more informative gradients than dot regression or cross-entropy. Experiments on OpenML-CC18 tabular datasets show CoCo achieves competitive performance with kernel SVM, Random Forest, and neural baselines while promoting tighter class clustering and faster convergence.
contrastive learningembedding collapseangular separationnormalized representationsgradient informativeness
Accelerated Mixing Time of Randomized Hamiltonian Monte Carlo
The paper establishes accelerated mixing time guarantees for Randomized Hamiltonian Monte Carlo (RHMC) when sampling from log-concave distributions. RHMC alternates between simulating Hamiltonian dynamics with random integration times (triangular or exponential distributions) and velocity resampling. For α-strongly log-concave targets, it achieves exponential KL convergence with total integration time O(α^(-1/2) log(ε^(-1))). For general log-concave targets, a triangular time sequence yields O(ε^(-1/2)) complexity. The analysis adapts Hamiltonian optimization techniques to bound KL divergence along trajectories.
randomized hamiltonian monte carlolog-concave distributionsmixing timekl divergencetalagrand inequality
Energy-Based Physics-Informed Form Finding for Clustered Tensegrity Structures
The paper introduces an energy-based learning framework for clustered tensegrity form finding and physical property prediction, addressing the nonlinearity and stability challenges in structural mechanics. The method integrates total potential energy minimization and constitutive relations into the training objective, enabling simultaneous prediction of equilibrium nodal configurations and physical quantities like member forces and force densities. Numerical experiments on prism and lander systems demonstrate the framework's scalability and accuracy in predicting structural properties, while improving physical consistency, robustness, and data efficiency.
tensegrity structuresenergy minimizationform findingphysical property predictionconstitutive relations
Deep4ge: DNN Training Trajectories for Fault Detection and Diagnosis
We introduce Deep4ge, a public benchmark dataset for fault detection and diagnosis in deep neural network training, addressing the lack of documented fault histories and feature extraction details. The dataset comprises 14,227 training runs from 59 adapted TensorFlow/Keras DNN programs, including 9,845 faulty runs generated via 27 source-code transformations across seven fault categories and 4,382 correct baseline runs. Each run records 4 evaluation metrics and 26 epoch-level features capturing weights, gradients, activations, accuracy, loss trends, learning rate, and hardware usage. Deep4ge supports binary fault detection, multi-class fault diagnosis, and early fault prediction from partial training runs, released with a fault-injection framework.
fault detectiondeep neural networkstraining trajectoriesfeature extractionfault diagnosis
Toward Localizing and Repairing Bias in Transformer Attention Heads
The paper introduces ROBIN, a white-box method for localizing and repairing bias in Transformer attention heads through targeted inference-time interventions. The approach ranks heads using sensitivity to fairness probes and removes a small bias subspace from selected head outputs, preserving language-modeling quality. In a four-model pilot study, ROBIN reduced the WinoBias gap across all models while outperforming whole-head zeroing in maintaining model performance. Results suggest head-level bias repair requires careful selection and modification of attention heads.
transformerattention headsbias localizationfairness probesinference-time intervention
Verifier-Based Reinforcement Fine-Tuning of Reasoning Models for Thermal Energy Storage Control
The study introduces reinforcement fine-tuning with verifiable rewards (RLVR) to adapt an open-weight reasoning model for thermal energy storage (TES) control in buildings, addressing scalability limitations of model predictive control (MPC) and reinforcement learning. RLVR converts offline dynamic-programming (DP) action values into dense rewards, training the model as an upper-level scheduler using only 30 prompts. Evaluated on a simple office-building TES benchmark, RLVR reduces emissions from 70.5 to 61.2 kg-CO2, nearing the DP optimum of 60.8 kg-CO2. GPT-5 nearly matches DP and MPC without task-specific training, while GPT-4o underperforms, highlighting the importance of inference-time reasoning. Robustness tests show reinforced planning patterns persist under forecast errors and transfer to a battery task, though with limited gains.
reinforcement fine-tuningthermal energy storagedynamic-programmingmodel predictive controlinference-time reasoning
ANGLE: Angular Neural Generative Learning via Engression
ANGLE introduces a lightweight deep generative framework for non-parametric distributional regression on circular data, addressing limitations of traditional regression in handling multimodal, skewed, or asymmetric structures. The method learns the full conditional distribution of angular responses via a generative map optimized using a generalized circular energy score (GCES) loss, ensuring rotational equivariance and strict propriety. Theoretical guarantees are provided, and the framework supports both pre- and post-additive noise models. Applications in object pose estimation and wind direction prediction demonstrate superior predictive performance and robust uncertainty quantification compared to existing methods.
circular datagenerative mapgeneralized circular energy scorerotational equivariancenon-parametric regression
AVQ-Attention: Adaptive Vector-Quantized Attention
AVQ-Attention introduces adaptive vector-quantized attention to address the computational bottleneck of standard attention in transformers. The method dynamically allocates codebook capacity based on attention importance, refining high-attention regions with pre-learned child codewords while maintaining coarse quantization elsewhere. Implemented via custom Triton kernels, it integrates with Flash Attention's tiled computation, preserving O(MN) complexity. Results show improved accuracy-efficiency trade-offs compared to fixed-codebook VQ-attention.
adaptive quantizationattention mechanismtriton kernelsflash attentioncomputational complexity
Directional Constraints for Efficient Exploration in Safe Reinforcement Learning
The paper proposes ATACOM-DC, an extension of the ATACOM framework that improves safe reinforcement learning by introducing directional constraints. These constraints selectively activate safety enforcement based on whether actions approach or move away from constraint boundaries, optimizing the safety-performance trade-off. Evaluated on robotic control tasks, ATACOM-DC reduces constraint violations while maintaining task performance compared to standard constrained RL methods. The approach integrates with existing RL algorithms and handles constraints derived from prior knowledge or learned from data.
safe reinforcement learningdirectional constraintsconstraint optimizationrobotic controlatacom framework
Learning-enabled Acceleration of Scenario-based Model Predictive Control
The paper introduces a learning-accelerated Alternating Direction Method of Multipliers (ADMM) algorithm to enhance the computational efficiency of Scenario-based Model Predictive Control (SBMPC). The method reformulates SBMPC problems into consensus forms, enabling parallel updates across scenarios and time steps, and employs Moreau envelope learning to accelerate primal updates in ADMM. Evaluated on a microgrid energy management problem with uncertainties in load and renewable generation, the proposed framework achieves significant computational speedups compared to IPOPT and MadNLP while maintaining reliable closed-loop control performance.
scenario-based mpcadmmmoreau envelopeparallel computingmicrogrid energy management
Learning Mechanistic Reasoning for Chemical Reactions with Large Language Models
The study introduces a novel approach to enhance mechanistic reasoning in large language models (LLMs) for chemical reactions by constructing a large-scale reasoning dataset and the FukuyamaBench benchmark. Fine-tuning Qwen3-30B-A3B on this dataset improves exact pathway matching to 8.3% on FukuyamaBench Set A, outperforming the specialized FlowER model (5.1%). The method addresses limitations in current chemical LLMs and small-scale generative models by focusing on stepwise mechanism deduction.
mechanistic reasoninglarge language modelschemical reactionsfukuyamabenchfine-tuning
What Makes a Representational Prior Work? Feature Families, Label-Free Invariances, and Critical Windows in Grokking
The study characterizes effective representational priors in grokking across four axes through 188 experimental runs. Key findings demonstrate that priors must align with the task's feature family (0/20 grokking with incorrect features vs 1/15 random), label-free invariance priors outperform supervised ones (15/15 runs with 2.7× speedup), and early application (first 2000 epochs) yields optimal results (10/10 runs, 2.7× speedup). The prior's effectiveness persists across tasks (modular multiplication) and architectures, with weight-norm clamps reducing delay-law exponent by 17× versus plain cross-entropy. Feature-family alignment and early timing emerge as critical factors for prior efficacy.
grokkingrepresentational priorfeature familyinvariance priorweight-norm clamp
Physically Consistent Parameter Inference: Transparent Machine Learning Emulation in High Energy Physics and Cosmology
The paper introduces a machine learning framework for emulating complex likelihood landscapes in high-energy physics and cosmology, addressing challenges of high-dimensional parameter spaces with non-Gaussian topologies. Using gradient-boosted regression trees (XGBoost), the method improves computational efficiency and resolution of confidence regions, particularly for systems with complex correlations or curved degeneracies. Validation on semileptonic $B$ meson decays demonstrates effectiveness, with SHAP values ensuring interpretability. The framework is adaptable to other systems like axion-like particles or cosmological global fits.
gradient-boosted regression treesnon-gaussian likelihoodshap valuesparameter inferencehigh-energy physics
Label-Decoupled Style Augmentation for Domain Generalization in Multi-Label Remote Sensing Scene Classification
The paper proposes a label-decoupled style augmentation framework for domain generalization in multi-label remote sensing scene classification. Unlike global feature-statistics augmentation methods (e.g., MixStyle), the approach confines style perturbation to label-specific regions using per-label attention from learnable modules or gradient class-activation maps. Evaluated on a multi-source benchmark (UCM, AID, DFC15) with leave-one-domain-out validation, the best variant achieves 71.5% mean average precision, outperforming empirical risk minimization by 5.0 points and global-statistics baselines by 1.3 points, with minimal parameter overhead (0.35%).
domain generalizationmulti-label classificationstyle augmentationremote sensingattention mechanisms
Extractable Memorization From First Principles
The paper introduces a principled framework for assessing extractable memorization in LLMs through matched comparisons, addressing validity issues in prior work. It proposes two methods: (1) a conformal test for population-level sequence analysis and (2) a census approach for single-document evaluation, both comparing generation probabilities of training versus non-training sequences. Results show that OLMo 2 32B reproduces non-training 10-token suffixes at 24% of training sequence rates, while Llama 3.1 70B exhibits memorization thresholds as low as 1e-27. The study refines extractable memorization to require valid claims and near-certain generation within practical budgets.
extractable memorizationmatched comparisonsconformal testgeneration probabilitiesllms
AdaPCLA: Adaptive Prior-Calibrated Logit Adjustment for Long-Tailed Longitudinal EHR Generation
AdaPCLA introduces an adaptive prior-calibrated logit adjustment framework for longitudinal Electronic Health Records (EHR) generation, addressing the underrepresentation of tail events in standard autoregressive models. The method employs a data distribution-aware training strategy, internalizing data knowledge parameters via simulated annealing, and supports zero-shot distribution control for diverse clinical populations. Theoretical analysis characterizes rare-code logit updates using the label-wise empirical NTK and derives a prior-internalization bound. Experiments on MIMIC-III and MIMIC-IV datasets demonstrate AdaPCLA's improvements in tail plausibility, downstream utility, and zero-shot control, achieving a 114.2% increase in TailPairSeen over HALO and a 3.5% F1 gain for zero-shot cross-population adaptation compared to GPT-style generation.
adaptive logit adjustmentsimulated annealingzero-shot controlempirical ntktail plausibility
Learning Forced Multibody Dynamics on Lie Groups
The authors propose a neural architecture for learning forced multibody dynamics using only position measurements by formulating discrete forced Euler-Lagrange equations on Lie groups. The method preserves geometric invariants by operating directly on manifold-valued configuration spaces and handles external control inputs. Evaluations on synthetic and real-world datasets demonstrate the framework's effectiveness in modeling mechanical systems without requiring velocity data.
lie groupseuler-lagrange equationsmultibody dynamicsmanifold-valued configurationsgeometric invariants
Gradient-free learning of a closed-loop wall controller for turbulent drag reduction
(No summary returned.)
The Geometry of Memorization: Finite-Time Spectral Sensitivity as a Diagnostic for Flow Matching Models
The paper introduces Finite-Time Spectral Sensitivity (FTSS) g(t), a gradient-free metric for diagnosing memorization in flow matching models by analyzing trajectory geometry. FTSS tracks the root-mean-square singular value of the state-transition matrix, serving as a continuous proxy for stable rank to reveal spectral collapse during overfitting. Experiments show that generalizing models maintain stable effective dimensions, while memorization induces geometric pathology detectable via internal trajectory dynamics without external queries.
flow matchingfinite-time spectral sensitivitystate-transition matrixspectral collapsegenerative memorization
Lightweight Multi-Scale Anomaly Detection for Resource-Constrained Edge Devices
We introduce LMSAE, a Lightweight MultiScale AutoEncoder for univariate time-series anomaly detection on resource-constrained edge devices. The model employs Discrete Wavelet Transform (DWT) to extract multi-scale features and a multi-scale loss function to enhance sensitivity to subtle anomalies. LMSAE achieves competitive detection performance on benchmark datasets while maintaining a compact architecture with fewer than 500 KB parameters. On the NVIDIA Jetson Nano, it demonstrates a 9x reduction in inference latency and a 2x reduction in power consumption compared to existing methods, making it suitable for edge deployment.
lightweight autoencoderdiscrete wavelet transformmulti-scale lossedge devicesanomaly detection
Environment Parameter Gradient Theorem for Policy-Environment Co-Design in Reinforcement Learning
The authors introduce the Environment Parameter Gradient Theorem, a formal expression for computing the gradient of the value function with respect to environment parameters in reinforcement learning. The theorem leverages a generalized action-value function $Q_{π,ξ}(s,a,ζ)$ that decouples environment parameters governing current and future dynamics, enabling tractable gradient computation. Based on this theoretical result, they develop a model-free algorithm for joint optimization of policy and environment parameters. The framework is validated on a UAV network design problem, demonstrating effective co-optimization of UAV placement and communication routes to minimize total network communication cost.
environment parameter gradient theoremgeneralized action-value functionpolicy-environment co-designmodel-free algorithmuav network design
A JoLT for the KV Cache: Near-Lossless KV Cache Compression via Joint Tucker and JL-Residual Allocation for LLMs
JoLT introduces a near-lossless KV cache compression method for transformer inference by treating the cache as a third-order tensor and applying a partial Tucker decomposition to compress token and feature axes while preserving head and layer axes. The method combines Tucker ranks with Johnson-Lindenstrauss (JL) rotated low-bit residuals, optimized via a Lagrangian dual under a byte budget. Results show 2-3x compression with minimal perplexity degradation, maintaining GSM8K accuracy and RULER retrieval performance on Mistral-7B-v0.3 and LLaMA-2-13B, achieving relative Frobenius errors of 0.009 (K) and 0.006 (V). FlashJoLT variant offers 5-13x speedup.
kv cachetucker decompositionjohnson-lindenstrausstransformer inferencelossless compression
From Preimage Search To Source-Grounded Feature Inversion
The paper introduces source-grounded feature inversion, a method for interpreting neural networks by reconstructing input-domain representations of internal features while preserving dependencies on the original input's network geometry. The approach uses closed-form matrix Wiener maps to repair adjoint signals during backpropagation, enabling finite reverse passes without query-specific optimization. Validated on CNNs and Transformers, the method produces inversions that depend on both the target feature and local network operations, with prediction-conditioned atlases aligning visualizations to feature interventions.
feature inversionwiener mapadjoint signalnetwork geometrysource-grounded
What Does Goodness Measure? A Likelihood-Ratio Account of Forward-Forward Learning
The paper provides a theoretical justification for the Forward-Forward (FF) algorithm's design choices by framing them within a likelihood-ratio test framework. It shows that the sum of squared activations (goodness) is a sufficient statistic for distinguishing between real and contrastive inputs under a zero-mean generative model. The analysis extends to anisotropic (Mahalanobis goodness) and heavy-tailed populations, revealing connections to divisive normalization and posterior precision. The study also explains inter-layer normalization requirements and identifies a scale-inflation shortcut in the pairwise objective, resolved by whitened goodness.
forward-forward algorithmlikelihood-ratio testdivisive normalizationmahalanobis goodnesssufficient statistic
Adversarial Attacks on Online Handwriting using Salience-based Temporal Editing
The authors propose a salience-guided temporal editing framework for adversarial attacks on online handwriting recognition models, addressing limitations of image-based spatial perturbation methods. The approach generates adversarial examples by inserting and deleting points at temporally salient locations identified through gradient-based activation mapping, preserving handwriting shape and smoothness. Evaluations on Unipen and CASIA-OLHWDB datasets demonstrate superior one-shot black-box transferability compared to conventional image-based attacks, while maintaining visual structure. Results indicate temporal editing as a relevant threat model for online handwriting recognition systems.
online handwriting recognitiontemporal saliencegradient-based activation mappingone-shot black-boxtemporal editing
Sample Efficient Generative Optimization for Molecular Design
Sample Efficient Generative Optimization (SEGO) introduces a Bayesian optimization framework for molecular design with improved sample efficiency. SEGO integrates a probabilistic surrogate model to hypothesize promising regions in chemical space, a generative model to propose candidates in those regions, and an acquisition function to select the most promising candidate for oracle evaluation. Evaluations refine both the surrogate model and anchor the generator in real reward. SEGO achieves state-of-the-art performance on the Practical Molecular Optimization (PMO) benchmark using only one-tenth the oracle calls of other methods and identifies ten hits in a multiparameter docking task with roughly half the oracle calls of existing approaches.
bayesian optimizationsample efficiencygenerative modelprobabilistic surrogatemolecular design
Steering Diffusion Models via Class-Contrastive Influence for Few-Shot Medical Classification
The paper introduces Class-Contrastive Influence (C2I), a gradient-based criterion to measure synthetic sample usefulness for few-shot medical image classification. C2I quantifies alignment between a sample's loss gradients and class-specific validation gradients, identifying boundary-proximal examples that refine decision boundaries. The authors fine-tune diffusion models via reinforcement learning using C2I as a reward, steering generation toward class-informative samples. Evaluations on few-shot medical imaging benchmarks show C2I-guided generation improves downstream accuracy and robustness over standard diffusion-based augmentation, demonstrating task usefulness outweighs image quality alone.
few-shot learningdiffusion modelsgradient alignmentmedical imagingreinforcement learning
Exploring Zero-Shot Foundation Models for Multivariate Time Series Anomaly Detection
The study evaluates zero-shot application of TimesFM, a univariate forecasting foundation model, for multivariate time series anomaly detection (MTSAD) on the SWaT benchmark. Two strategies are tested: per-feature forecasting with thresholded errors and embedding-based outlier detection. Results show neither approach matches established baselines, as the model's strong temporal dynamics capture yields low error even in anomalous windows, masking persistent anomalies. However, error peaks at anomaly boundaries suggest utility for change-point detection. Findings indicate naive zero-shot FMs are inadequate for MTSAD but may aid in detecting distribution shifts.
multivariate time seriesanomaly detectionfoundation modelszero-shot learningchange-point detection
Language Identification with Succinct Machine-Independent Traces
The paper addresses two open questions in language identification within the Gold-Angluin model: whether traces can use a small alphabet and whether they can be defined directly from the language without an underlying machine model. The authors propose a method for defining computational traces that enable identification in the limit, using an alphabet linear in the size of the language's alphabet and independent of language properties. They establish positive results for both questions, demonstrating the feasibility of succinct, machine-independent traces.
language identificationgold-angluin modelcomputational tracesidentification in the limitsuccinct alphabet
Fisher Rank Inflation: A Spectral Signature of Memorization under Label Noise
The paper identifies Fisher Rank Inflation as a spectral signature of memorization in deep networks trained with label noise, characterized by transient expansion and contraction of the effective rank in last-layer gradient scatter. Analyzing centered scatter of per-example gradients, the authors derive a first-order leave-one-out attribution formula and demonstrate enrichment of corrupted examples during peak rank inflation. Experiments on CIFAR-10, CIFAR-100, and CIFAR-10N with SmallCNN, ResNet18, and Vision Transformers show corrupted fractions up to 96.2% among top rank-contributing samples, with peak effective rank increasing monotonically with corruption severity (28.88±1.95 to 97.09±1.78 at 60% corruption).
fisher rank inflationlabel noisespectral signaturegradient scattermemorization
PolarBM: Complex-valued Boltzmann Machine for Modeling Audio Signals in Polar and Log-polar Coordinates
The authors propose PolarBM, a complex-valued Boltzmann machine that models audio signals in polar coordinates, explicitly capturing amplitude-phase relationships. They extend this to LogPolarBM, which uses logarithmic amplitudes for auditory perception alignment, yielding a power-weighted noncentral complex Gaussian distribution. Restricted variants (PolarRBM, LogPolarRBM) are introduced for practical use. Experiments show these models outperform conventional approaches, including deep neural networks, in modeling accuracy. While tested on audio, the method generalizes to other complex-valued domains like wireless communications and quantum mechanics.
boltzmann machinecomplex-valued datapolar coordinatesnoncentral gaussianaudio signal processing
Mechanical Analysis of Parachute Suspension Line Deployment with Binding Tapes Using PINN
The paper presents a physics-informed neural network (PINN) for predicting tension dynamics during parachute suspension line deployment, addressing limitations of traditional numerical integration methods. The proposed PINN framework models the ultra-short deployment phase involving binding tapes, capturing dynamic load variations more efficiently than ODE-based approaches. Validation against flight test data demonstrates superior computational efficiency and numerical accuracy compared to conventional methods, while also enabling tension prediction at arbitrary line positions. The study further analyzes how binding tape parameters regulate dynamic tension during line extraction and straightening.
physics-informed neural networksuspension line deploymentdynamic tension predictionbinding tape parametersordinary differential equations
MESH: Scaling Up Retrieval with Heterogeneous Content Unification
The paper introduces MESH, a unified retrieval scaling framework addressing the Scaling Bias of Heterogeneity in large-scale retrieval systems. The method employs a modularized architecture with gated bias correction, partitioning the feature space into independent domains to reduce interference between sparse and high-frequency signals. Empirical results show a 14× improvement in power-law scaling for fresh items, +5.5% lift in repins, 55% funnel efficiency gain, and +0.46% user retention on Pinterest's Related Pins platform, alongside 2.87× throughput improvement via asynchronous serving.
retrieval scalingheterogeneous contentgated bias correctionpower-law scalingasynchronous serving
ReDiTT: Retrieval Augmented Conditional Diffusion Transformers for Asynchronous Time Series
ReDiTT introduces a retrieval-augmented conditional diffusion transformer for asynchronous time series prediction, addressing uncertainty in next-event time and type forecasting. The model operates in latent space, retrieving structurally similar sequences from a memory bank during training and inference, which are incorporated via cross-attention for global structural guidance. This approach stabilizes long-horizon forecasting and enhances sample diversity. Evaluations on seven real-world datasets show state-of-the-art performance in next-event prediction and long-horizon forecasting.
diffusion transformerasynchronous time seriesretrieval augmentationcross attentionlatent space
Thompson Sampling Is 2-Competitive for Mistakes
The paper proves that Thompson sampling is 2-competitive for expected mistakes in Bayesian bandit models, meaning it makes at most twice the mistakes of any other policy. The analysis applies to independent latent arm processes where each arm evolves only when played, confirming a 2014 conjecture by Guha and Munagala with an optimal competitive factor. Results hold for stochastic bandits with best-arm mean-reward criteria under any nonincreasing sequence of round weights, including fixed horizon and geometric discounting scenarios.
thompson samplingbayesian banditscompetitive analysisstochastic banditsregret bounds
Differentiable Clone-Structured Causal Graphs for End-to-End Cognitive Map Learning from Image Sequences
This work introduces gradCSCG, a differentiable reformulation of the Clone-Structured Causal Graph (CSCG) algorithm, enabling end-to-end cognitive map learning from raw image sequences. The method couples gradCSCG with a vector-quantized variational autoencoder (VQ-VAE) perceptual front-end, employing soft emission forward passes and loss-balancing mechanisms to prevent module collapse during joint training. Experiments demonstrate that gradCSCG replicates CSCG's performance on symbolic grid worlds and maintains robust map recovery on MNIST image sequences, achieving high edge precision and recall across four heavily aliased environments. This establishes CSCG as a composable module in deep learning architectures.
gradcscgvq-vaecognitive mapsoft emissionloss-balancing
SinAE: A Single-Architecture Flow-Matching Autoencoder for Cross-Domain Atomic Systems
SinAE introduces a unified flow-matching autoencoder architecture for cross-domain atomic systems, including molecules, crystals, and proteins, using a vanilla Transformer encoder-decoder without domain-specific operators. By shifting reconstruction to an iterative flow-matching decoder, it achieves near-lossless reconstruction across domains, reducing errors significantly compared to prior latent baselines. The model supports a Diffusion Transformer prior for generation tasks, demonstrating strong performance on molecular, crystal, and protein benchmarks. Joint training across molecules and crystals improves performance in both domains, evidencing cross-domain transfer through a shared atomic latent.
flow-matchingautoencodertransformercross-domaindiffusion transformer
Same Loss, Same Noise, Opposite Schedules: Noise Structure and Optimizer Normalization Jointly Determine Whether Learning-Rate Cooldown Helps
The paper provides a theoretical framework explaining when learning-rate cooldown improves final training loss in warmup-stable-decay schedules, based on gradient noise structure and optimizer normalization. Analyzing strongly convex objectives with multiplicative noise, it shows stochastic gradient descent (SGD) contracts geometrically without cooldown benefits, while sign-based and normalized methods exhibit a noise floor scaling as η². The authors derive exact stationary laws for signSGD on quadratics, prove local dissociation under (L₀,L₁)-smoothness, and extend results to normalized SGD in higher dimensions. Experiments confirm theoretical predictions and validate the diagnostic on real classification tasks.
learning-rate cooldowngradient noiseoptimizer normalizationsignsgdnoise floor
Reducing information dependency does not cause training data privacy. Adversarially non-robust features do
The paper challenges the conventional belief that information dependency drives training data exposure in model inversion attacks (MIAs), demonstrating instead that exposure stems from adversarially non-robust features. Through experiments, the authors show that MIA defenses reduce leakage without affecting information dependency (HSIC), models with maximal memorization resist MIA, and models trained on 3% of pixels remain vulnerable to reconstruction. They introduce Anti Adversarial Training (AT-AT), which leverages non-robust features to improve defense and accuracy, revealing a privacy-robustness tradeoff.
model inversion attacksadversarial robustnessinformation dependencynon-robust featurestraining data privacy
Generating Developable 3D Molecules via Pocket-Conditioned Diffusion and Property-Aware Optimization
The authors introduce conDitar-dev, a conditional diffusion-based framework for structure-based drug design that generates ligands with strong binding affinities and favorable ADMET properties. The method combines a pretrained multi-scale pocket representation learning module (msPRL), a pocket-conditioned diffusion model (conDitar), and a generation-time optimization method for ligand developability (paOPT). On a benchmark of human disease targets, conDitar-dev achieves an average binding score of -8.85 kcal/mol and improves ADMET property performance by up to 73% over conDitar. Experimental validation on PD-L1 and CSF1R targets yielded molecules with SPR-derived KD values of 3.49 and 3.75 μM and CSF1R inhibitors with IC50 values as low as 200 nM.
diffusion modelsstructure-based drug designbinding affinityadmet propertiesligand optimization
Robust Design of Integrated Sensing and Communication in LEO Satellite Systems
A robust beamforming design algorithm is proposed for integrated sensing and communication (ISAC) in low Earth orbit (LEO) satellite systems, enabling simultaneous target sensing and user communication over shared spectrum. The method minimizes total transmit power while satisfying mean squared error (MSE) requirements for sensing and signal-to-interference-plus-noise ratio (SINR) requirements for communication, accounting for channel phase uncertainty that exacerbates cross-functional interference. Theoretical analysis demonstrates the algorithm's effectiveness, and simulations confirm its superiority over baseline approaches in optimizing resource utilization for multi-functional satellite operations.
beamforminglow earth orbitintegrated sensing and communicationmean squared errorsignal-to-interference-plus-noise ratio
Gradient Flow Dynamics and Implicit Bias of Diagonal Linear Networks under Infinitesimal Initialization
The work extends prior analysis of gradient flow dynamics in diagonal linear networks under infinitesimal initialization, generalizing results to deep networks and a broader class of two-layer architectures. By proposing Algorithm 1, the authors show training trajectories equivalently solve a modified $\mathcal{l}_1$ norm minimization problem, revealing an implicit bias toward this norm. The analysis identifies the Structural Invariant Manifold (SIM) as the geometric structure governing these dynamics, providing mechanistic insights into the learning process.
gradient flow dynamicsdiagonal linear networksimplicit biasstructural invariant manifoldinfinitesimal initialization
What Does a Temporal Benchmark Score Measure? Decomposing Channel Use in Video VLM Evaluation
The paper introduces the reversal-drop, a label-free method to assess whether video-language models (VLMs) rely on positional encodings (RoPE) or visual sequence for temporal understanding in benchmarks. By reversing visual input while maintaining forward RoPE, the method distinguishes position-dominant (e.g., Molmo2) from visual-sequence-dominant (e.g., Qwen3-VL) models. Experiments across two benchmarks and multiple temporal tasks reveal consistent behavioral splits, validated via activation patching. The work shows aggregate benchmark scores obscure critical failure modes, as models with similar accuracy exhibit divergent reliance on temporal channels.
temporal understandingvideo-language modelspositional encodingreversal-dropactivation patching
SlimPer: Make Personalization Model Slim and Smart
SlimPer introduces a transformer-based architecture optimized for personalized recommendation systems by reformulating ranking as iterative refinement of a compact knowledge base. Unlike generative models, SlimPer selectively queries multi-modal user-side tokens, computes relevance scores, and refines the knowledge base with O(N) per-layer cost and fixed-size intermediate representations, decoupling model depth from user history length. It unifies sparse, dense, and sequence features within a single backbone, offering interpretability through attention mechanisms. Deployed on Instagram Reels and Feed, SlimPer improves user engagement while efficiently modeling over 10,000 fine-grained user history events.
transformerrecommendation systemsknowledge baseattention mechanismmulti-modal
A Shared Subcircuit Lets LLMs Count Down Across Tasks
The study identifies a reusable 'countdown subcircuit' in Llama-3.1-70B-Instruct that enables length-aware task execution across diverse domains. Through controlled experiments on fixed-length sentence generation, the authors isolate this circuit and analyze its representational geometry, finding it employs a motif previously observed in other LLMs. Unsupervised probing reveals the subcircuit's broader applicability, including contexts where target lengths are implicitly inferred, demonstrating how reverse-engineering can reveal generalizable behavioral mechanisms in transformers.
countdown subcircuitlength-aware tasksrepresentational geometryunsupervised probingbehavioral mechanisms
Auditing Data Leakage in Whole-Slide Image Multimodal Benchmarks
This work identifies and quantifies pervasive data leakage in whole-slide image (WSI) visual question answering benchmarks, compromising claims of vision-language model (VLM) performance. Through systematic tracing of slide, case, and Tissue Source Site (TSS) identifiers across public resources, the authors demonstrate 92.3-100% case-level train-test overlaps and near-complete TSS overlaps in TCGA-derived benchmarks. Experiments show both leakage types are linearly decodable from foundation-model feature space, inducing measurable accuracy gaps between leaked and clean cases. Analysis reveals peak reported VLM accuracies concentrate on heavily contaminated benchmarks, suggesting current evaluations conflate multimodal reasoning with nearest-neighbor retrieval of memorized artifacts. The study concludes with recommendations for contamination-free evaluation protocols.
data leakagewhole-slide imagevisual question answeringtissue source sitefoundation-model
A hybrid analytical-PINN model for subsurface simulation of geothermal heat exchangers in heterogeneous underground
A hybrid analytical-physics-informed neural network (PINN) model is proposed for simulating geothermal heat exchangers in heterogeneous underground environments. The method integrates analytical line source models to remove singularity, employs explicit gradient thermal conductivity formulations for physics-informed learning, and utilizes learned corrections via superposition principles. The approach decomposes temperature changes and transforms heterogeneous responses into corrections compensating for idealized homogeneous approximations, enabling efficient training. The model is trained by minimizing a physics-informed and data-anchored loss function evaluated on adaptively selected training points. Numerical tests demonstrate the method's effectiveness across three analytical models, showcasing its capability in capturing bulk heat transfer and soil heterogeneity.
physics-informed neural networkthermal conductivityheterogeneous soilsuperposition principlesborehole heat exchangers
Quantum Port-Hamiltonian Neural Networks: Learning Conservative and Dissipative Dynamics via Measurement-Induced Nonlinearity
The paper introduces Quantum Port-Hamiltonian Neural Networks (Q-pHNNs), a family of parameterized quantum circuits that learn classical dynamics while preserving structural properties. The method employs Isomorphic Hamiltonian Mapping (IHM), where unitary gates model conservative dynamics via a skew-symmetric matrix and Measurement-Induced NonLinearity (MINL) enforces dissipation through mid-circuit measurement and feedforward. Four architectures are presented, including variants for energy manifold learning, damping coefficient estimation, and coupled systems. Experiments on nonlinear systems demonstrate 1.35% relative energy drift, 100% energy monotonicity with MINL, and 12.1% error in unsupervised damping-coefficient identification.
quantum neural networksport-hamiltonian systemsmeasurement-induced nonlinearityparameter-shift rulesymplectic integrator
Saturation Makes Quantization Error Additive: A Coverage Model with a Certificate
The paper introduces a coverage model for mixed-precision quantization, demonstrating that quantization error becomes additive under saturation conditions. By analyzing the change in loss $f(S)$ as a set function on the Boolean cube, the authors show that 85--93% of $f$'s variance is explained by per-layer effects alone. They propose a coverage model $f(S)=c\bigl(1-\prod_{i\in S}(1-a_i)\bigr)$, which accurately reproduces the variance profile of $f$ with $L$ fitted break-rates. The model supports two predictors: an additive model and the coverage model itself, both with $L+1$ parameters. These allocators achieve the lowest KL divergence on models ranging from 30B to 355B parameters and maintain performance below 4-bit precision.
mixed-precision quantizationcoverage modeladditive errorboolean cubekl divergence
From Many to Meaningful: Feature-Guided Zero-Shot Chronic Kidney Disease Screening Using Large Language Models
The study proposes a feature-guided zero-shot framework for chronic kidney disease (CKD) screening using large language models (LLMs), eliminating the need for dataset-specific training. The method employs ML-based feature selection to identify a compact set of clinically meaningful variables, serializes tabular patient records into text prompts, and evaluates four LLMs (LLaMA-3, Qwen-3, Mistral, GPT-4o-mini) across three heterogeneous CKD datasets. Results show statistically significant improvements in balanced accuracy with the selected feature subset, demonstrating LLMs' potential for practical, training-free screening in resource-limited settings.
zero-shot learningfeature selectionlarge language modelschronic kidney diseaseclinical screening
Understanding Structured Health Data through Interaction-Aware Mixture-of-Experts
The paper proposes an interaction-aware mixture-of-experts (MoE) model for post-stroke rigidity prediction using multi-view structured health records. The method employs view-specific expert routing to analyze hierarchical medical data, with attribution analysis revealing systematic importance variations across views. While performance improvements were marginal, the study demonstrates that view construction critically impacts model interpretability in clinical prediction tasks.
mixture-of-expertsstructured health recordsmulti-view learningclinical predictioninterpretability
When Directional Accuracy Lies: A Base-Rate-Honest Benchmark for LoRA-Adapted TimesFM on Equity Forecasting
The paper introduces a base-rate-honest benchmarking protocol for evaluating LoRA-adapted TimesFM in equity forecasting, addressing directional accuracy's susceptibility to market base rates. The method employs expanding walk-forward folds, stratified held-out-ticker splits, and rigorous statistical tests (McNemar, Diebold-Mariano) with FDR control. Results show that apparent 80% directional accuracy is largely a base-rate artifact (~0.70), with pooled LoRA failing to outperform the 'always-up' baseline. Per-sector specialization underperforms pooled adaptation (p<0.001), and fine-tuning only reduces point-forecast error without yielding tradeable directional edges.
lora adaptationtimesfmdirectional accuracybase-rate artifactdiebold-mariano test
Proximity Features: Privacy-Compliant Cold-Start Personalization at Airbnb
The paper introduces Proximity Features, a privacy-compliant feature system for cold-start personalization in two-sided marketplaces like Airbnb. The method groups users by geographic proximity using geo-IP data and adaptive clustering, aggregating signals for ~1,000 nearby users without persistent identifiers. Privacy is ensured via consented, aggregated data processing within gated controls. Deployed in production, the system shows statistically significant booking lifts in A/B tests, particularly for users with absent or stale history.
cold-start personalizationgeo-ip clusteringprivacy-compliant featuresadaptive clusteringtwo-sided marketplaces
Rough Path Signature-Guided Geometry Augmentation for Few-Shot Industrial Surface Defect Detection
The paper introduces rough path signature-guided geometry augmentation (RPS-GA), a method enhancing few-shot industrial defect detection by leveraging geometric features from boundary contours. It processes Canny edges as ordered paths, computing truncated second-order signatures (notably the antisymmetric Lévy-area term) to generate spatial maps via SIG-AUG and SGAA fusion operators. Evaluated on NEU-DET and PCB-Defect with YOLOv8n, RPS-GA significantly boosts performance in low-shot regimes: SIG-AUG improves 10-shot mAP@0.5 from 0.341 to 0.583 on NEU-DET, while SGAA achieves 0.299 versus 0.086 baseline on PCB-Defect. Gains diminish with more labeled data, confirming robustness across random partitions.
rough path signaturegeometry augmentationfew-shot detectionlévy-areadefect segmentation
Cluster-Weighted EDMD
The paper introduces Cluster-Weighted EDMD (CW-EDMD), a method that jointly learns a soft phase-space partition and per-cluster Extended Dynamic Mode Decomposition (EDMD) operators to address inefficiencies of global Koopman operators in heterogeneous dynamical systems. CW-EDMD employs an Expectation-Maximization objective that clusters transitions based on both geometric proximity and prediction residuals, enabling specialization where local Koopman models are accurate. Evaluated on Lorenz, damped pendulum, and Duffing systems across 36 configurations, CW-EDMD significantly outperforms matched-degree EDMD in 258 of 288 comparisons, achieving median one-step error reductions of 57x, 2.7x, and 12x respectively.
extended dynamic mode decompositionkoopman operatorexpectation-maximizationphase-space partitiondynamical systems
Gradient-Free Topology Adaptation for Power Flow Surrogates via In-Context Whitening
This work introduces In-Context Whitening (ICW), a gradient-free method for adapting AC power flow (ACPF) surrogates to unseen network topologies. ICW whitens the output space using the base topology's first two moments and re-estimates this whitening from a few hundred solved cases on the new topology, preserving coordinate-wise semantics via ZCA whitening. Evaluated on IEEE 30-, 118-, and 300-bus systems under N-1 and N-2 contingencies, ICW reduces overall error by 6× to 28× compared to frozen surrogates, with worst-bus power-balance mismatch reduced by up to 30×. ICW matches or exceeds gradient-based adaptation accuracy while adapting 21× to 34× faster, leveraging commodity CPU cores instead of GPUs.
ac power flowin-context whiteningzca whiteninggradient-free adaptationn-1 contingency
Speculate with Memory: Lossless Acceleration for LLM Agents
This work introduces memory-augmented speculative execution for LLM agents, enabling lossless acceleration by leveraging past agent trajectories. The method integrates three online memory systems: a contrastive transition table for action-sequence statistics, episodic memory for contextually similar segments, and a confusion tracker to suppress recurring errors. Evaluated across six benchmarks involving action, observation, and chained prediction tasks, the approach achieves 19--39% relative accuracy improvements in action prediction and up to 2.5× gains in observation prediction with repetitive action spaces. Performance improves continuously as memory accumulates, generalizes across speculator models, and incurs zero added wall-clock cost.
speculative executioncontrastive transition tableepisodic memoryconfusion trackerlossless acceleration
Fine-Tuned Multi-Agent Framework for Detecting OCEAN in Life Narratives
The paper proposes a fine-tuned multi-agent framework for detecting OCEAN personality traits from life narratives, addressing challenges of latent trait expression and LLM biases. Sub-agents are conditioned on high, low, or neutral trait perspectives via masked language modeling (MLM) and psychometric supervision, with a judge LLM aggregating outputs for final predictions. Evaluations on life narrative datasets demonstrate improved performance over baselines, showcasing the framework's scalability and interpretability through quantitative and qualitative analyses.
multi-agent frameworkocean traitsmasked language modelingpsychometric supervisionlife narratives
Forgetful Attention: A Trainable Support-Vector Memory with Certified Selection and Exact Unlearning
The paper introduces Support Vector Attention (SV-Attention), a max-margin memory mechanism with certified token eviction and exact unlearning capabilities. By framing attention as a one-class SVM with fixed box parameter C, the method guarantees output preservation when dropping tokens and enables reversible incremental deletion matching full retraining (median deviation 10^-9). Experiments show 9,125 tokens/s throughput on a 3.22M-parameter model, 35.8× slower than softmax but with superior rare-item recall (0.86 vs. 0.32) and temporal stability. Applications include surgical forgetting, exact editing, and retrieval over sentence embeddings, with a 8.6% perplexity improvement on enwik8 versus sliding-window Transformers (p=0.001).
support vector attentionmax-margin memorycertified evictionexact unlearningreversible incremental solver
Decentralized Gradient Descent: Bottleneck Regimes and Budget Complexity
The paper characterizes the communication-computation budget required for decentralized gradient descent (DGD) to achieve a target error level, introducing a bottleneck-centric framework. It identifies distinct optimization regimes governed by initialization, objective heterogeneity, network connectivity, gradient noise, and communication noise, quantified via Gradient-Diversity-to-Network-connectivity Ratio (DNR) and Gradient-to-Communication-noise Ratio (GCR). Through multi-stage analysis, the authors derive optimal stepsize selections and explicit budget-complexity bounds, revealing how the budget decomposes across successive bottlenecks and tradeoffs among system parameters.
decentralized gradient descentbudget complexitygradient diversitynetwork connectivitycommunication noise
From Geometric Recovery to Causal Validation: A Reproducible Audit of Sparse Autoencoder Features, from Superposition Geometry to Causal Inertness
The study critically evaluates sparse autoencoder (SAE) feature recovery metrics, demonstrating that high cosine similarity between decoder atoms and ground-truth directions does not guarantee causal relevance. Through reproducible audits of SAE behavior across superposition regimes, the authors identify two types of causally inert features: structural inertness (present in well-trained SAEs) and competitive inertness (a pathology of degraded SAEs). Their method, sae-causal-audit, combines ablation and steering experiments to dissociate read- and write-inertness, revealing antipodal-pair geometries where features are unmonitorable yet steerable. Results show up to 77% of recovered features in degraded SAEs and 9% in well-trained SAEs are causally inert, with specific steering specificities of 143-310.
sparse autoencoderscausal validationsuperposition geometryfeature recoveryinertness
Falsifying Causal Graphs With Outlier Events
The authors introduce a novel method for falsifying candidate causal graphs by assessing their ability to explain outlier event propagation, leveraging the principle that weak outliers rarely cause strong ones. They propose statistical tests to hypothesize whether a candidate graph is the true causal graph, ensuring false positive control and power guarantees against incorrect graphs. These tests can operate with a single outlier sample, offering a robust approach to causal graph validation without ground truth.
causal graphsoutlier propagationstatistical testsfalse positive controlroot cause analysis
Generating Physically Plausible Parachute Dynamics with Deep Generative Modeling
The authors propose SPar-GAN, a physics-aware generative model for simulating parachute dynamics, addressing challenges in traditional system identification due to nonlinearity, unknown governing equations, and scarce test data. SPar-GAN adapts a Hamiltonian generative architecture by conditioning on canopy design and freestream velocity while enforcing energy conservation through symplectic integration. Applied to subscale parachute tests from the National Full-Scale Aerodynamics Complex, the model reproduces pitch-yaw dynamics across configurations and recovers a compact two-degree-of-freedom phase-space consistent with canopy axisymmetry. Results demonstrate the potential of physics-constrained generative models to reduce physical testing requirements for parachute performance assessment.
symplectic integrationgenerative adversarial networkparachute dynamicshamiltonian architecturephase-space
An Agentic AI Scientific Community for Automated Neural Operator Discovery
The paper introduces an agentic AI scientific community for autonomous neural operator discovery, comprising virtual laboratories that interact via a citation-based economy. Each lab contains three agents: an LLM planner proposing architectures, a numerical worker training them, and an LLM reviewer participating in peer review. The community evaluates architectures using DeepONet, Fourier, Transformer, wavelet, and residual convolutional neural operator building blocks. Experiments on five problems (piecewise regression, linear advection, Burgers 1D PDEs, Navier-Stokes, and Darcy flow 2D PDEs) demonstrate the discovery of high-accuracy, low-parameter-count architectures. LLM planners hybridize architectures in 99.8% of cases, preserving diversity, while rule-based alternatives collapse to single-family stacks.
neural operatorllm plannercitation-based economydeeponetpeer review
FlashDiff: Efficient Regional Execution and Scheduling for Diffusion Model Serving
FlashDiff introduces an efficient diffusion model serving system through adaptive regional execution and scheduling, addressing the inefficiency of all-region execution in diffusion inference. The method decomposes latent representations into coherent regions using early-stage attention signals, employs a runtime controller to bypass low-impact updates, and uses an affinity-aware scheduler for load balancing. Evaluated on image, video, and audio workloads, FlashDiff reduces latency by 30-97% and improves throughput by 1.2-2.2x.
diffusion modelsregional executionadaptive schedulinglatent decompositionserving efficiency
Robust In-Hand Manipulation via Priors in Reinforcement Learning and Mechanical Design
The authors propose two physics priors to enhance robust in-hand manipulation in reinforcement learning: a global grasp-quality prior for well-distributed contacts and a local contact-geometry prior based on fingertip curvature. These priors are integrated into reward shaping and fingertip design to improve grasp stability and task-aligned rolling. The method is evaluated on a multifingered robotic hand manipulating three objects at four palm orientations, demonstrating significant improvements in rotation efficiency, grasp stability, and disturbance rejection. Results indicate that embedding physics priors in both learning and mechanical design enhances task robustness and sim-to-real transfer.
reinforcement learninggrasp-quality priorcontact-geometry priorin-hand manipulationsim-to-real transfer
TraceSynth: Generating Production-Quality Kernel Traces with Constraint-Guided Diffusion Models
TraceSynth introduces a diffusion-based framework for generating synthetic kernel traces to augment limited real data for ML-based system diagnostics. The method employs a Transformer-based denoising diffusion process with constraint-guided repair to model multi-channel sequences (event types, timestamps, CPU affinity, etc.), enforcing system invariants. Results on six benchmarks show synthetic data achieves 87.2% F1-Macro for compute-heavy workloads (scimark2) at context length L=4096, only 2.6 percentage points below real-only baselines, with constraint-guided repair improving quality by up to 4.3%. Lightweight 2-channel models retain 97-99% of 6-channel performance at half the cost.
kernel tracesdenoising diffusionconstraint-guided repairmulti-channel sequencescontext length
Dynamic Online Processor-Native Inference for State Estimation
The paper introduces a Bayesian filtering technique for dynamic state estimation that leverages processor-native uncertainty tracking to accelerate likelihood computation. The method employs deterministic hierarchical importance restructuring via native operations, ensuring deterministic latency and bounded memory usage for arbitrary programmatic models. Evaluated on three nonlinear state-space systems, the approach achieves up to 805× speedup over Monte Carlo methods at comparable accuracy, while demonstrating Pareto-dominant accuracy-latency trade-offs and competitive RMSE against particle filters.
bayesian filteringstate estimationuncertainty propagationimportance restructuringdeterministic latency
Institutional Equity Holdings Prediction Using Node Affinities of Dynamic Graphs
The paper introduces NAVIS, a temporal graph machine learning model for institutional equity holdings prediction, framed as node affinity prediction on a bipartite graph of managers and securities. Using SEC Form 13F data (99 managers, 503 securities, 209,351 edges across 48 quarters), NAVIS achieves state-of-the-art test NDCG of 0.9127, outperforming dynamic graph models and heuristics. Temporal and structural signals dominate predictive power, with features providing <1.2% gain. Exponential Moving Average (0.8882) and Persistent Forecast (0.8891) baselines perform strongly, highlighting portfolio persistence.
temporal graphnode affinity13f filingsndcgdynamic graph representation learning
Learning from Complementary Ultrasound Representations for Liver Disease Classification
The study demonstrates that complementary ultrasound representations enhance NASH versus NAFLD classification compared to conventional B-mode imaging alone. By combining B-mode ultrasound with physics-guided and local phase-based representations, the authors employ self-supervised masked autoencoders (MAEs) and graph convolutional networks (GCNs) on a multi-site cohort of 2,547 liver ultrasound scans from 125 patients. Results show improvements of up to 32.4% in accuracy and 91.2% in F1-score, with consistent gains across demographic and acquisition variables.
ultrasound representationsself-supervised learningmasked autoencodersgraph convolutional networksliver disease classification
Analyzing Image Encoder Choices and Graph Homophily in GCN Frameworks for Breast Ultrasound Classification
The study demonstrates that image encoder selection significantly impacts graph construction and classification performance in GCN-based breast ultrasound analysis. Five image encoders (convolutional and transformer-based) were evaluated by constructing cosine similarity k-nearest-neighbor graphs from embeddings, classified using a single-layer GCN. Higher-capacity encoders improved graph homophily and classification metrics (accuracy, AUC, sensitivity, specificity, F1-score) across three cross-validation folds, with graph homophily strongly correlating (linear) with accuracy. Results indicate encoder-driven graph structure enhancements as a key performance mechanism.
graph convolutional networksimage encodergraph homophilyultrasound classificationcosine similarity
SymbOmni: Evolving Agentic Omni Models via Symbolic Concept Learning
SymbOmni introduces an agentic omni-model for cumulative evolution via Symbolic Concept Learning, addressing the 'perpetual novice' problem in visual generation. The model employs a Symbolic Concept Box for abstracting low-level operations into reusable Symbolic Workflow Instructions, operating through an induction-transduction cycle. Training uses verbalized backpropagation with language-based feedback, enabling continuous self-improvement without gradient-based fine-tuning. Experiments show SymbOmni outperforms agent-based systems and closed-source models (e.g., Nano Banana, GPT-Image-1) in image quality and task success rates, reduces token consumption by 40%, and achieves state-of-the-art continual learning performance.
symbolic concept learningagentic omni-modelinduction-transduction cycleverbalized backpropagationcontinual learning
Learning the Graphical Nature of Symmetries
The paper constructs a dataset of 131,406 Cayley graphs covering finite groups of order ≤767 (excluding 512), providing exact algebraic labels and graph-theoretic statistics for benchmarking group property learning. It introduces new OEIS sequences for monolithic groups and groups with ≤5 generators, while identifying empirical relationships between graph metrics (square clustering, diameter, disorder) and spectral properties of nilpotent groups. Experiments compare MLP, GCN, and GIN architectures, showing that engineered graph features are highly informative and graph-aware models (particularly GIN) effectively capture structural signals from Cayley graph representations.
cayley graphsfinite groupsgraph neural networksspectral statisticsnilpotent groups
Imputation-free transformer learning enables robust Alzheimer's disease prediction and calibrated uncertainty quantification across heterogeneous clinical cohorts
The study introduces NITROGEN, an imputation-free transformer for Alzheimer's disease prediction that jointly models within-patient feature dependencies and between-patient relational structure via masked and intersample attention. Trained on ADNI (N=7858 scans) and evaluated on OASIS-3 (N=2675) and AIBL (N=1286), NITROGEN demonstrated robust calibration and uncertainty quantification across diagnostic and cognitive score prediction tasks, outperforming tree-based ensembles. Key features included cortical thickness in the temporal pole, age, and APOE genotype. A modality-aware uncertainty adjustment improved calibration when diagnostic data was missing, showing the importance of evaluating models beyond accuracy for clinical deployment.
transformeruncertainty quantificationmultimodal learningcalibrationalzheimer's disease
SKooP: Symmetric Koopman Predictions for Faster and More Generalizable Legged Robot Locomotion with Reinforcement Learning
The paper introduces SKooP (Symmetric Koopman Predictions), a reinforcement learning method that combines morphological symmetries with Koopman models to improve sample efficiency and generalization in legged robot locomotion. SKooP jointly learns a Koopman autoencoder for dynamics prediction and a symmetric policy, using Koopman features as privileged critic observations. The approach enforces equivariance in actor, critic, encoder, and decoder networks through group symmetries. Experiments on quadrupedal bipedal locomotion tasks demonstrate reduced convergence time (30-50% faster) and higher rewards compared to baselines, with successful transfer to unseen simulation environments.
koopman operatormorphological symmetrysample efficiencyequivariant policylegged locomotion
SpikeDS: Dual Sparsity Spikformer for Perineural Invasion Prediction in 3D MRI
The paper introduces SpikeDS, a Dual Sparsity Spikformer for perineural invasion (PNI) prediction in 3D MRI of cholangiocarcinoma (CCA). The method combines activation sparsity from binary spike communication and spatial sparsity from window pruning via two mechanisms: Window-based Expert Mixture Spiking Attention (W-EMSA) for local salient regions and Cross-Window Spiking Self-Attention (CW-SSA) for global context. Evaluated on 139 CCA patients with 5-fold cross-validation, SpikeDS achieves 0.753 AUC at 14.4 mJ energy, outperforming baselines in accuracy and efficiency.
spiking neural networkdual sparsity3d mriperineural invasionspiking attention
VQCSim: When Does Compile-Once Statevector Simulation Beat Generic Quantum Frameworks?
VQCSim introduces a compile-once, PyTorch-native statevector execution path with native autograd for hybrid quantum-classical machine learning workflows, addressing framework overhead in static variational circuits. It systematically evaluates performance using MQT Bench, achieving 87.7% semantic validation and median speedups of 4.49x for inference and 26.78x for training across five GPUs. Native autograd contributes a 27.6x acceleration, with compile-once caching and batch vectorization providing additional gains. VQCSim trades higher GPU memory for reduced runtime, supported by a hardware-aware regime map and vqcsim-oracle, an open-source backend selector with 91.1%-97.7% top-1 agreement for automatic simulator selection.
statevectorautogradquantum-classicalvariational circuitssemantic validation
Gene Expression-Informed Jointly Controlled Generative Modeling for Precision Molecular Design
JoPMol, a jointly controlled precision molecular generative model, integrates gene expression profiles, molecular structure text, and chemical property values for coordinated molecular generation and optimization. The model jointly controls biological relevance and molecular design strategies, enabling personalized drug candidate discovery. Experimental results demonstrate JoPMol's superior performance across multiple metrics compared to state-of-the-art methods, with strong generalization in transfer tasks and biologically grounded simulations. The unified modeling framework effectively addresses precision molecular design challenges.
generative modelinggene expressionmolecular designchemical propertiesjoint control
📰 Industry Media (1)
Meet GPT-Red: an LLM super-hacker OpenAI built to make its models safer
OpenAI developed GPT-Red, a large language model (LLM) specialized in automated red-teaming to enhance the robustness of its models, particularly GPT-5.6. GPT-Red operates in a self-play loop with other LLMs, simulating real-world deployment scenarios to discover novel attack vectors, such as fake chain-of-thought injections. The model identified previously unseen prompt injection attacks, achieving a 90% success rate against GPT-5 but only 23% against GPT-5.6. While GPT-Red excels in persistent, scenario-specific attacks, it struggles with conversational and image-based exploits. The system complements human red-teamers but remains proprietary due to its advanced compute requirements.
gpt-redself-play loopprompt injectionchain-of-thoughtred-teaming
Generated automatically at 2026-07-15 20:31 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.