Daily Digest — 2026-07-25
266 items · 1 research labs, 245 arxiv papers, 20 industry media
🏛️ Research Labs (1)
How Codex became a collaborator for OpenAI’s creative team
OpenAI's Codex demonstrates emergent capabilities as a creative collaborator, enabling novel workflows for design and ideation. Through in-context learning, Codex processes multimodal inputs including sketches, storyboards, and style guides, maintaining project-specific context across brand books, products, and strategic goals. The system accelerates creative exploration by generating 50 campaign directions in one day, which were distilled to 10 high-quality options. Codex facilitates rapid prototyping of custom UIs, API integrations, and visual controls through natural language instructions, reducing technical barriers. This case study illustrates how Codex bridges the gap between conceptualization and execution in creative workflows.
in-context learningmultimodal inputsrapid prototypingapi integrationcreative workflows
📜 arXiv Papers (245)
3D-Aware VLMs with Implicit and Explicit Geometries
The paper introduces VLM-IE3D, a 3D-aware vision-language model framework that enhances spatial reasoning by integrating implicit and explicit 3D geometries from RGB videos. The method employs Implicit Geometry Tokens (IGTs) for high-level geometric priors and Explicit Geometry Tokens (EGTs) for detailed 3D structures, fused via a 3D-aware adapter without requiring 3D inputs. Experiments demonstrate superior performance on 3D video detection, visual grounding, dense captioning, and spatial reasoning tasks.
3d-aware vlmsimplicit geometry tokensexplicit geometry tokensspatial reasoning3d visual grounding
GraphVid: Interactive Graph-Controllable Video Generation
We introduce GraphVid, a graph-conditioned image-to-video generation model enabling interactive control through structured interaction graphs, addressing limitations of trajectory-based control in multi-object scenarios. The method leverages GraphVid-Bench, a large-scale interaction-centric video dataset with structured relational annotations, to train interaction-aware video generation models. GraphVid achieves superior performance with reduced training data and parameters compared to prior motion-control methods, demonstrating significant improvements in video quality metrics: FID reduction up to 39.9%, FVD reduction by 37.6%, PSNR increase from 9.87 to 15.98, and SSIM improvement from 0.38 to 0.61.
graph-conditionedinteraction graphsvideo generationstructured annotationscontrollability
Barzilai-Borwein Fails Superlinear Convergence on an Open Set of Quadratics for Every Dimension $n\geq 4$
The Barzilai-Borwein (BB1) method fails to achieve superlinear convergence on a nonempty open set of strictly convex quadratic problems for all dimensions n≥4, resolving a longstanding open question. Through a computer-assisted proof, the authors construct an attracting seven-cycle in the projectivized BB dynamics, demonstrating that gradient norms, energy norms, and objective gaps are bounded below by geometric sequences with explicit constants ρ_min=10^-6 and ρ_max=0.61. This rules out root-superlinear convergence, providing a negative answer to the conjecture that BB converges superlinearly for almost every strictly convex quadratic problem and initialization.
barzilai-borwein methodsuperlinear convergencestrictly convex quadraticprojectivized dynamicscomputer-assisted proof
Synthetic data generation framework for quality control automation in gravure printing
A synthetic data generation framework is proposed to automate quality control in rotogravure printing, addressing the scarcity of real-world defect images for training deep learning models. The pipeline generates high-fidelity images of printing defects (creases, streaks, misregistration) with bounding boxes and annotations, producing a dataset of 7533 images. The RFDETR model trained on this synthetic data achieves a Mean Average Precision (mAP) of 80.9% on real industrial test samples, demonstrating effective defect detection without costly manual data collection.
rotogravure printingsynthetic data generationmean average precisionrfdertdefect detection
Beyond Sufficiency: Time Series Explanation with Counterfactual Necessity
We introduce TimePNS, a necessity-aware framework for time-series explanation that addresses the limitation of existing sufficiency-oriented methods by identifying subsequences essential to a black-box model's prediction. Inspired by Pearl's counterfactual notion of necessity, TimePNS employs a two-stage design: Stage I learns a causal generative process and a sufficiency-oriented explanation mask, while Stage II performs counterfactual interventions on temporal factors to derive necessity signals, refining the explanation via a temporal gate. Experiments on synthetic and real-world benchmarks demonstrate that TimePNS accurately identifies decision-critical subsequences and improves sufficiency-necessity trade-offs over baselines.
time-series explanationcounterfactual necessitycausal generative processsufficiency-oriented masktemporal gate
Unsupervised Consensus-Based Anomaly Detection for Spatiotemporal Malaria Incidence in Ghana
The study introduces an unsupervised consensus anomaly detection framework to analyze spatiotemporal malaria transmission patterns in Ghana (2014-2023), distinguishing anomaly burden from anomaly frequency. The method identifies atypical monthly malaria incidence, revealing structured spatial and temporal anomalies. Results show Ashanti and Northern Regions as hotspots, with Tamale exhibiting the highest anomaly burden and Ashanti districts the highest anomaly frequency. Anomalous months form a statistically distinct group, characterized by significantly higher case counts (Cohen's d = 3.252) and large seasonal deviations (d > 1.2). This framework enhances malaria surveillance by differentiating areas of high prevalence from those with unusual transmission dynamics, supporting targeted control strategies.
anomaly detectionspatiotemporal analysismalaria transmissionconsensus frameworkseasonal deviations
Beyond Sycophancy: Structured Resistance and Compliance in LLM Moral Reasoning
The study proposes a structured framework for understanding resistance-compliance dynamics in LLM moral reasoning, moving beyond sycophancy as a unidimensional failure mode. Through three empirical studies, it identifies three key dimensions of judgment revision: distance from initial position, source attribution, and coalition structure. Results show models are more receptive to nearby positions, influenced by self-attributed views, and variably responsive to group pressure. This framework distinguishes constructive belief revision from sycophantic compliance, advancing alignment in morally consequential interactions.
moral reasoningsocial influencejudgment revisionllm alignmentsycophancy
OpenForgeRL: Train Harness-native Agents in Any Environment
OpenForgeRL introduces an open-source framework for training harness-based AI agents end-to-end in diverse environments, addressing limitations of current SFT/RL stacks in handling stateful, multi-process harness inference. The framework employs a lightweight proxy to record harness model calls as training data and a Kubernetes orchestrator for remote containerized rollouts, enabling scalable training across any harness and environment. Evaluations demonstrate strong performance on benchmarks including ClawEval (31.7 pass@3), QwenClawBench (33.7), OSWorld-Verified (37.7), and WebVoyager (72.3), outperforming open baselines and matching larger models in GUI settings. Analysis reveals harness-specific learning difficulties and RL's impact on agentic reliability, though error recovery remains a challenge.
openforgeharness-based agentskubernetes orchestratorstateful inferencerl codebase
Visual Contrastive Self-Distillation
The paper proposes Visual Contrastive Self-Distillation (VCSD), a novel on-policy self-distillation method that eliminates the need for external teachers, privileged answers, or visual evidence signals. VCSD generates distillation signals by contrasting teacher-generated next-token distributions conditioned on original versus content-erased images, sharpening the target distribution for student training. Evaluated on ViRL39K with Qwen3-VL and Qwen3.5 models, VCSD improves seven-benchmark aggregates by 4.77-5.25% across 2B-8B parameter scales, with no additional inference cost.
self-distillationcontrastive learningvision-language modelson-policy distillationtoken-wise probability
MIRROR: Learning from the Other View for Multi-Modal Reasoning
The paper introduces MIRROR, a reinforcement learning method for improving multimodal reasoning in vision-language models (VLMs) by leveraging complementary reasoning paths across different problem views. The authors construct ODA-Data, a paired multimodal geometry dataset with text-dominant, image-dominant, and combined views, and propose Modality-Informed Reciprocal Reasoning Optimization (MIRROR), which selects the best-performing view as a teacher and trains other views using a reverse-KL objective. Results show MIRROR outperforms standard RL and enhances accuracy and consistency across modalities on geometry reasoning benchmarks.
multimodal reasoningvision-language modelsreverse-kl objectivereinforcement learninggeometry problems
The Boundaries of Automation: A Theory of Persistent Human Participation
The paper challenges the assumption that human participation in AI systems is solely due to technical limitations, proposing three persistent reasons for human involvement: technical complementarity, normative developmental value, and emergence grounds. Emergence grounds, the central contribution, argue that in activities where objectives emerge through interaction, human participation is constitutive rather than merely corrective. This co-construction perspective suggests that human-AI collaboration is not transitional but essential for certain tasks. The findings have implications for automation limits, AI system design, evaluation, and ethics.
automationco-constructionemergencenormativecomplementarity
From Resource Flow to Executable Tests: Petri-Net-Guided LLM Test Generation for Concurrent Stateful Rust APIs
The paper introduces a Petri-net-guided methodology for generating executable tests for concurrent stateful Rust APIs, bridging formal scenario design and low-cost test concretization. The method models API resources, lifecycle conditions, and dependencies using colored tokens and transitions, deriving legal deep-state, near-legal, and partial-order concurrent scenarios. These scenarios serve as constrained intermediate representations for LLM-based code synthesis, with a local-faithfulness contract and structural repair loop ensuring intent preservation. Petri-guided schedule shaping prioritizes high-conflict concurrency skeletons, while a layered semantic oracle distinguishes synthesis failures from API behavior violations.
petri-netconcurrent stateful apisllm-based code synthesissemantic oracleschedule shaping
ElasticTTT: Prior-Preserving Test-Time Tuning for Video Editing
ElasticTTT introduces a novel framework for test-time tuning (TTT) in video editing that addresses Prior Collapse, a phenomenon where pretrained diffusion models discard text conditions and spatial latents during single-point optimization. The method employs Target Distribution Regularization to avoid sharp memorization minima, Contrastive CFG to mitigate source biases, and Asynchronous Noise Schedule to preserve unedited regions. Theoretical analysis and extensive evaluations demonstrate that ElasticTTT successfully maintains the generative prior of the base model, achieving state-of-the-art performance in one-shot video editing.
test-time tuningprior collapsediffusion modelstarget distribution regularizationcontrastive cfg
GS-Agent: Creating 4D Physical Worlds With Generative Simulation
GS-Agent introduces an end-to-end multi-agent framework for generating physically plausible and controllable 4D worlds from natural language descriptions. The system decomposes the task into entity management (3D asset curation, material tuning, placement, motion control) and rendering configuration (camera, lighting), leveraging physics engines and multimodal feedback. Multiple specialized agents collaborate iteratively to construct dynamic worlds with rich interactions among liquids, deformable objects, and rigid bodies. Experimental results demonstrate GS-Agent's ability to produce diverse, physically realistic 4D worlds with cinematic control, advancing automated content creation and physical AI.
4d world generationphysics enginesmulti-agent frameworknatural language processingphysical plausibility
Same Dangerous Objective, Opposite Advice: Direct Exposure versus Multi-Agent Mediation
This study reveals a compositional safety gap in high-capacity language models, demonstrating that direct exposure to a dangerous objective yields safer advice than multi-agent mediation. Using OpenAI's gpt-5.6-sol model, the authors tested 25 mirrored trade-off profiles, comparing direct exposure to an objective authorizing concealment and fabrication with multi-agent workflows involving Id, Censor, and Superego components. Direct exposure produced advice opposing the target, while multi-agent mediation aligned advice with the target despite obscuring manipulative clauses. The results suggest that downstream models in automated workflows may propagate explicitly manipulative objectives without recognizing their provenance or dangerous clauses.
compositional safety gapmulti-agent mediationmanipulative clausestrade-off profilesdownstream models
Improved lower bounds for the Shannon capacity of odd cycles
The authors present improved lower bounds for the Shannon capacity of odd cycles through explicit combinatorial constructions of independent sets in strong powers of graphs. Using iterative interactions with a Large Language Model (LLM), they construct independent sets of size 134753 in C_7^10, 21909 in C_11^6, and 62530 in C_13^6, yielding Shannon capacity lower bounds of Θ(C_7) > 3.258020, Θ(C_11) > 5.289773, and Θ(C_13) > 6.300109. Additionally, they improve lower bounds on the independence numbers of several strong powers of odd cycles, demonstrating the utility of LLMs in combinatorial optimization.
shannon capacityindependent setsstrong powersodd cyclescombinatorial constructions
Agentic Context Management: Solving Agent Memory and Cost by Treating Them as Lifecycle and Architecture Problems
The paper introduces Agentic Context Management (ACM) as a framework to address memory and cost inefficiencies in production AI agents by treating context management as a lifecycle problem rather than a storage-and-retrieval issue. ACM decomposes into five primitives: architecting, ingesting, scoping, anticipating, and compacting & consolidation. The authors argue that naive context accumulation leads to quadratic token cost growth, while crude summarization achieves linear cost at the expense of accuracy. A reference implementation, Maximem Synap, achieves 92% accuracy on LongMemEval and 93.2% on LoCoMo, demonstrating validated compaction with preserved fidelity. The paper highlights unmet benchmarking needs in latency, token efficiency, and context-rot resistance.
agentic context managementtoken costvalidated compactionlongmemevallocomo
Artificial Epanorthosis: Why large language models overuse a classical rhetorical figure, and how to mitigate it
The paper identifies systematic overuse of epanorthosis (self-correction rhetoric) in large language models, attributing it to training distributions favoring promotional prose and RLHF preference tuning. It introduces an Epanorthosis Index to quantify genre-specific deviations, finding models overshoot in oratory (2-3× human rates) and undershoot in informal Q&A while matching human baselines in other genres. Mitigation techniques include LoRA adapters and instructional prompts, demonstrating 50-75% reduction in Italian, with supervised fine-tuning achieving near-complete removal while preserving genre-appropriate calibration.
epanorthosisrhetorical figurepreference tuninglora adaptersgenre calibration
Toward Continuous Assurance for the Democratization of AI Agent Creation in Industry
The paper proposes a continuous-assurance framework to address reliability challenges in democratized AI agent creation within organizations. The framework integrates dependency mapping, readiness contracts, scheduled checks, diagnostics, and lifecycle governance to ensure operational readiness of citizen-developed agents under expected conditions. An initial prototype auditor and scenario-based assessment demonstrate the feasibility of translating the proposed taxonomy into practical checks and actionable remediation guidance. This approach mitigates silent degradation risks arising from evolving dependencies such as models, tools, retrieval sources, permissions, prompts, schedules, and external services.
continuous assurancedependency mappingreadiness contractsoperational readinesslifecycle governance
Compact Latent Coordination for Autonomous Vehicles at Unsignalized Intersections
The paper introduces Master-Agent Proto-plan System (MAPS), a hierarchical deep reinforcement learning architecture for coordinating autonomous vehicles at unsignalized intersections. MAPS employs a centralized Master agent that generates compact, continuous embeddings (proto-plans) encoding global coordination strategies, which decentralized Worker agents integrate with local observations for vehicle-specific control. This decouples strategic intent from tactical execution, enabling independent module optimization. Evaluated across 72 intersection configurations in HighwayEnv, MAPS achieves collision-free navigation and reduces average travel time, outperforming state-of-the-art baselines. The system demonstrates robust generalization, with a 94% success rate in zero-shot deployment from three to five agents.
multi-agent reinforcement learninghierarchical deep reinforcement learningproto-planunsignalized intersectionszero-shot generalization
Agentic coding without the cloud: evaluating open-weight large language models on longitudinal data preparation tasks
The study introduces an open-source framework for evaluating open-weight large language models (LLMs) on longitudinal data preparation tasks, addressing governance constraints in sensitive research data. The framework includes a ground-truth dataset from a British cohort study, task definitions (e.g., category harmonization), and automated evaluation routines for LLM-generated R code. Benchmarking 31-35B parameter models on 20 tasks (102 variables) showed high efficacy (87.9% average task completion), demonstrating the viability of local LLM deployment for governance-restricted research.
open-weight llmslongitudinal data preparationcategory harmonizationautomated evaluationgovernance-restricted research
Error Certificates for KV-Cache Eviction via Randomized Design
The paper introduces randomized KV-cache eviction as a solution to the identifiability problem in deterministic eviction, where altered evicted values can lead to unbounded attention-output errors without detection. By employing Poisson-sampling with known inclusion probabilities and applying the Hájek correction via a logit offset, the method provides per-step error certificates with 0.97 empirical coverage at no accuracy cost. Experimental results on real workloads demonstrate that the certificate effectively separates cache-induced from inherent failures (AUC 0.73–0.75) and improves recomputation scheduling compared to random or confidence gating, though it does not enhance prediction accuracy.
kv-cacheevictionpoisson-samplinghájek correctionerror certificate
Thinkink: 2D Spatial Ink-native Interaction with LLMs
The paper introduces Thinkink, a system for 2D spatial ink-native interaction with large language models (LLMs) that integrates handwritten text and sketches with LLM-generated responses on a shared canvas. The design process involved a formative study (N=12), a technical probe for diagnostic study (N=6), and a final evaluation (N=10), employing a semantic tree for ink interpretation and a lightweight UI with explicit state machine control. Results demonstrate how users incorporate Thinkink into ideation practices, yielding design implications for ink-native LLM interaction.
ink-native interactionlarge language modelssemantic tree2d canvasstate machine
AREX: Towards a Recursively Self-Improving Agent for Deep Research
The paper introduces AREX, a family of Recursively Self-Improving (RSI) agents designed for deep research tasks. AREX alternates between an inner research loop for evidence gathering and answer construction, and an outer self-improvement loop for constraint-wise verification and targeted follow-up research. The system employs an autonomous context-update tool to compress interaction history into a compact improvement state, trained via agentic mid-training and long-horizon RL with emphasis on critical evidence-acquisition steps. Evaluations on BrowseComp, WideSearch, DeepSearchQA, and HLE benchmarks show AREX outperforms comparable-scale baselines, with implementations including a 4B dense model and a 122B-A10B MoE variant.
recursively self-improvingconstraint-wise verificationautonomous context-updatelong-horizon reinforcement learningmixture-of-experts
Detecting LLM-Generated Tokens in Human--LLM Coauthored Text
The paper introduces a token-level detection method for localizing LLM-generated content in human--LLM coauthored text, addressing the limitation of existing document-level classifiers. The method smooths adjacent token scores using an adaptive Lepski-type bandwidth selection rule to reduce variability while preserving local authorship structure. Theoretical analysis shows favorable mean square error performance, and empirical evaluations demonstrate superiority over baselines on synthetic and realistic datasets. A public implementation is provided.
token-level detectionlepski-type rulemixed-authorshipllm-generated textadaptive smoothing
RUMBA: Russian User Memory Benchmark
The authors introduce RUMBA (Russian User Memory BenchmArk), a novel benchmark for evaluating long-term conversational memory in LLMs, addressing limitations of English-centric benchmarks by incorporating temporal reasoning and session-spanning interactions. The benchmark features timestamped dialogues with QA pairs requiring cross-session retrieval and reasoning, organized by a fine-grained taxonomy of memory types and temporal expressions. Evaluations on contemporary models demonstrate RUMBA's utility as a diagnostic tool for analyzing memory mechanisms across different semantic and temporal conditions, with an aligned English subset provided for comparative analysis.
long-term memoryconversational aitemporal reasoningbenchmark taxonomycross-session retrieval
Agent-Guided Relational Concept Discovery: Toward Interpretable Surgical Margin Assessment
The paper introduces Agent-Guided Concept Discovery, a framework for interpretable surgical margin assessment using Rapid Evaporative Ionization Mass Spectrometry (REIMS) data. The method learns human-understandable concepts directly from unlabeled intraoperative data without predefined annotations, employing a reasoning agent to refine semantic descriptions and adjust concept weights based on diagnostic relevance. Concepts are grounded using a biochemical knowledge graph to ensure metabolic consistency. Evaluated on Skin and Breast Cancer datasets, the model demonstrates improved balanced accuracy, sensitivity, and fewer false positives in intraoperative scenarios compared to baseline approaches.
surgical margin assessmentrapid evaporative ionization mass spectrometryconcept-based learningbiochemical knowledge graphintraoperative data
Adaptive Identity Anchoring: Closed-Loop Keyframe Placement for Synthetic Paired Supervision in Video Face Swapping
The paper proposes Adaptive Identity Anchoring (AIA), a closed-loop method for keyframe placement in video face swapping to address identity drift in synthetic paired supervision. AIA generalizes synthesizers to arbitrary anchor sets, uses feedback to insert anchors at worst-scoring frames, and filters data automatically. Paired with Reality-Referenced Texture Restoration, it mitigates over-smoothed textures by transferring micro-texture from real footage. The method frames identity-anchor density as a quality dial and proposes falsifiable experiments to validate its efficacy.
adaptive identity anchoringvideo face swappingsynthetic paired supervisionreality-referenced texture restorationclosed-loop keyframe placement
Token Budget Saturation and Mechanistic Early Detection of Reasoning Non-Convergence in Chain-of-Thought Models
The study identifies and analyzes bimodal convergence patterns in chain-of-thought reasoning models like DeepSeek-R1-Distill-Qwen-7B, where generations either terminate within token budgets (converged) or exhaust them without conclusion (non-converged). Using linear probes on hidden-state activations (layers 20 at token positions 50-300), the authors demonstrate early detection of convergence fate, achieving AUC 0.608 (±0.080) by token 150, outperforming behavioral baselines. Empirical results show converged generations achieve 90.3% accuracy on AIME 1983-2024 versus 6.6% for non-converged, with a 62.0% overall convergence rate, suggesting potential for early-exit inference strategies.
chain-of-thoughtlinear probeshidden-state activationsearly-exit inferencetoken budget
Bridging the Gap Between Plausibility and Admissibility: Constraint-Aware Flow Maps for Dynamic Graph Systems
This study introduces a constraint-aware framework to enhance the reliability of generative trajectory modeling in dynamic graph-structured systems by bridging statistical plausibility and structural admissibility. A conditional diffusion model generates future graph-state trajectories from partial observations, while an external symbolic layer applies hard filtering, soft weighting, or projection-based repair. Evaluated on compact and medium-complexity synthetic graphs, the framework achieves an invalid probability mass of 0.002996 in the compact regime and 0.155929 in the medium-complexity regime. Hard filtering removes all invalid trajectories while preserving 84.4% of samples, demonstrating that symbolic constraint handling becomes increasingly valuable with graph complexity.
conditional diffusion modelsymbolic constraintsdynamic graph systemshard filteringstructural admissibility
PATS: Policy-Aware Training Scaffolding for Agentic Reinforcement Learning
The paper introduces PATS, a policy-aware training scaffold for reinforcement learning in long-horizon LLM agents, addressing weak policies' tendency to produce uninformative rollouts. PATS dynamically converts policy rollouts into evidence cards, using task-specific evaluation to adjust rollout context—providing concrete guidance for weak policies while reducing redundant context as the policy improves. Evaluated on ALFWorld, WebShop, and seven QA benchmarks, PATS outperforms baselines by up to 18.6% and reduces prompt tokens by 32.1% while maintaining competitive performance.
reinforcement learningllm agentspolicy optimizationrollout trajectoriesevidence cards
Logical Regression for Planning with Axioms
The paper introduces a methodology for approximating logical regression in automated planning domains with axioms, producing minimal partial states while avoiding axiom recalculation. The approach limits conditions to partial states and embeds this form of regression in execution monitoring. Results demonstrate a 70% reduction in variables considered for execution monitoring, with recovery rates exceeding 50% in environments with unexpected changes across multiple domains.
logical regressionautomated planningpartial statesexecution monitoringaxioms
Euclid-MCP: A Model Context Protocol Server for Deterministic Logical Reasoning via Prolog
Euclid-MCP introduces a standardized Model Context Protocol server for deterministic logical reasoning via SWI-Prolog, addressing LLMs' unreliability in multi-step inference. The system features Euclid-IR, an engine-agnostic intermediate representation for Horn-clause logic, and supports a translate-run-inspect-repair loop for LLM clients. Evaluation on IT security compliance shows Euclid-MCP outperforms LLMs in accuracy (eliminating hallucinations) and latency on larger knowledge bases, while providing proof traces. The authors argue semantic RAG is unsuitable for rule enforcement, positioning Euclid-MCP as a shared reasoning substrate for agentic systems.
neuro-symbolichorn-clause logicintermediate representationmodel context protocoldeterministic reasoning
Cycle-Consistent and Uncertainty-Aware Neural Surrogates for Tokamak Edge Plasmas
The authors propose a cycle-consistent neural surrogate for tokamak edge plasmas, combining a conditional U-Net forward model with an optimization-based inverse method to enable bidirectional prediction and uncertainty quantification. The forward model maps five control parameters to 2D plasma-state fields on the SOLPS-ITER mesh, while the inverse method enforces consistency via self-supervised learning without ground-truth labels. Results show normalized RMSE below 2.6%, Pearson correlations above 0.95 for all fields, and 5-6 orders of magnitude speedup over SOLPS-ITER, enabling real-time control and digital twins.
neural surrogatecycle-consistencytokamak edge plasmauncertainty quantificationdigital twin
MSBraM: A Multi-scale Self-supervised Brain Foundation Model for Hierarchical EEG Dynamics Learning
The authors propose MSBraM, a multi-scale self-supervised foundation model for EEG analysis that captures hierarchical temporal dynamics. The method employs a two-stage approach: (1) a neural tokenizer discretizes EEG signals at varying resolutions via vector-quantized reconstruction, followed by (2) curriculum-based masked prediction training integrating local and global temporal context. Pretrained on 2,400+ EEG hours, MSBraM outperforms state-of-the-art models on 10 downstream tasks across 12 datasets, demonstrating superior generalization through explicit multi-scale representation learning.
eeg foundation modelmulti-scale learningvector-quantized reconstructionmasked predictionhierarchical dynamics
When Are Reasoning-Based Guardrails Not Efficient? ResponseGuard: A Fast Vision-Language Guard for Real-Time Moderation
The paper introduces ResponseGuard, a fast vision-language safety guard that eliminates chain-of-thought reasoning for real-time moderation. The method uses a single forward pass to pool representations of the request, response, and image, achieving 150× faster detection than reasoning-based guards while outperforming a 3B competitor on response harmfulness detection in standard benchmarks. Results show a remaining performance gap on image-only cases, attributed to frozen vision encoders rather than missing reasoning steps, with the reasoning guard showing minimal attention to image features.
vision-language guardrailsreal-time moderationsingle-pass detectionharmfulness detectionfrozen vision encoders
VoLN: Vision-Only Long-Horizon Navigation---Paradigm, Benchmark, and Method
The paper introduces Vision-Only Long-Horizon Navigation (VoLN), a paradigm for embodied navigation that shifts route-relevant information from external instructions to locally observable in-scene cues. The authors instantiate VoLN through VoLN-UAV, a 7,210-episode aerial navigation benchmark featuring long-horizon goal-directed flight, continuous 3D motion, and context-dependent beacon selection. Their baseline method, VoLN-MLLM, aligns self-supervised visual features with a semantic space and predicts waypoints from observation history, achieving success rates of 7.4%, 4.5%, and 1.8% on Easy, Normal, and Hard test episodes, respectively.
vision-only navigationembodied agentslong-horizon navigationaerial navigationvisual-semantic alignment
Multimodal Pretraining for Generalizable EEG Representation Learning
The authors propose a multimodal EEG foundation model combining a Mamba-based raw signal encoder, Vision Transformer for time-frequency data, and lightweight text encoder in a shared embedding space. Pretraining employs masked modeling, cross-view contrastive alignment, and temporal consistency losses to learn seizure-relevant representations without labels. Evaluated on CHB-MIT seizure detection, the model achieves 0.874 AUROC (0.878 ensemble), with 0.558 mean balanced accuracy in leave-one-subject-out testing across 19 subjects, demonstrating robust cross-dataset generalization and interpretable seizure localization.
eeg foundation modelmamba architecturevision transformerself-supervised learningseizure detection
Towards Faithful Graph Explanations with Synergistic Edge Effects via Granular Balls
SeeExplainer introduces a parameter-free method for interpreting graph neural networks (GNNs) by addressing the neglected synergistic effects among edges in previous approaches. It employs a granular-ball graph refinement mechanism to decompose graphs into disjoint granular-balls, constructing a structural graph that captures edge synergies. Explanatory subgraphs are generated by perturbing nodes and edges based on their contributions. Experiments on various graph classification datasets demonstrate that SeeExplainer outperforms state-of-the-art baselines.
graph neural networkssynergistic effectsgranular-ball graphparameter-free explainergraph classification
Mean-to-Score Discrete Diffusion: Posterior-Mean Denoisers for Score Entropy
The paper introduces Mean-to-Score (M2S), a method for discrete diffusion models that predicts clean-token posterior means and converts them to scores via a kernel-dependent linear map, ensuring Bayes realizability. M2S addresses limitations in Score Entropy Discrete Diffusion (SEDD), where score ratios may violate posterior constraints, by projecting scores onto the bridge polytope. Evaluated on CIFAR-10 and OpenWebText, M2S improves test BPD from 3.173 to 3.129 and generative PPL from 183.6 to 143.3 at 128 steps, outperforming SEDD, GIDD, and Neural CTMC baselines.
discrete diffusionbayes realizabilityscore entropyposterior meanbridge polytope
DINOde: Continuous Vision-Text Alignment for Open-Vocabulary Semantic Segmentation
DINOde introduces a continuous Ordinary Differential Equation (ODE)-based framework for open-vocabulary semantic segmentation (OVSS), aligning CLIP text embeddings with the DINOv3 visual manifold. The method comprises Semantic Text Flow (STF), which evolves text embeddings toward the DINO manifold via a continuous ODE trajectory, and Global Context Flow (GCF), which refines DINO's CLS token representation. Velocity Tangent Projection preserves hyperspherical geometry by constraining velocity fields to tangent spaces. This continuous alignment avoids manifold entanglement from discrete MLP projections, enhancing cross-modal robustness. Experiments show DINOde achieves state-of-the-art performance across multiple OVSS benchmarks.
open-vocabulary semantic segmentationordinary differential equationclip text embeddingsdino manifoldvelocity tangent projection
Hilbert Operator for Progressive Encoding (HOPE): A Mathematical Framework for Deconstructing Learned Representations in Deep Networks
Hilbert Operator for Progressive Encoding (HOPE) introduces a mathematical framework to deconstruct learned representations in deep networks by shifting compression into a Hilbert space of continuous functions. It models neurons as rank-1 Hilbert-Schmidt operators, unifying pruning and neuron merging as low-rank subspace projection, and extends to macro block eviction for multi-layer structures. This data-free, hyperparameter-free approach eliminates architectural biases across layers. Proof-of-concept experiments demonstrate its potential in model compression and fine-tuning.
hilbert spacelow-rank projectionneuron mergingmacro block evictionhilbert-schmidt operators
SPORD: A Simulation-Propose-then-OR-Dispose Approach for Supply Chain Planning
SPORD introduces a Simulation-Propose-then-OR-Dispose method for supply chain planning, addressing operational fragmentation, computational intractability, and implementation hurdles. The approach decouples simulation, which generates operationally valid candidate paths, from an integer program that selects globally optimal subsets. Matrix-vectorized CPU/GPU acceleration achieves 10-100x speedups, and a list scheduling algorithm reduces coupled-order processing from hours to minutes. Deployed as JD.com's NetSim platform since 2025, SPORD optimized services for over 20,000 suppliers, reduced cross-regional fulfillment rates from 6.1% to 4.9%, and achieved an average monthly carbon reduction of 5,745 tCO2e. The method enhances transparency and modularity, facilitating executive trust and reducing reconstruction needs.
simulation-propose-then-or-disposematrix-vectorizedinteger programlist schedulingcomputational intractability
Regulating autonomous and agentic AI
The paper examines regulatory challenges posed by autonomous and agentic AI systems, arguing that traditional governance models fail due to shifted knowledge and control dynamics in the AI supply chain. Through comparative analysis of four regulatory regimes (UK content platforms, data protection, financial services, and EU AI Act), it identifies limitations of retrospective oversight and systemic risks from AI autonomy. Proposed solutions advocate transforming regulation into an active, adaptive process by expanding oversight to encompass AI supply chains and developing novel risk management frameworks.
autonomous aiagentic systemsregulatory governancesupply chain oversightsystemic risk
M$^3$-Gen: Interpretable Multimodal Generation of Gene Expression Profiles Using Clinical and Imaging Data
M$^3$-Gen introduces a multimodal framework for generating gene expression profiles by conditioning a Generative Adversarial Network on histopathology images and clinical metadata. The method learns a unified latent representation via contrastive learning and uses attention-based mechanisms to guide generation, ensuring biological coherence and interpretability. Evaluated on TCGA, M$^3$-Gen produces realistic profiles while identifying image regions influencing specific expressions, offering intrinsic explainability.
generative adversarial networkcontrastive learninggene expression profilesmultimodal integrationattention mechanism
Phonetic forced alignment for low-resource language varieties: Model training and evaluation on Chengdu Mandarin
The study develops phonetic forced alignment systems for low-resource language varieties, specifically Chengdu Mandarin, addressing the lack of specialized models. Two approaches are proposed: a text-dependent GMM-HMM model (Chengdu-MFA) trained on a 17-hour corpus with a custom G2P dictionary, and a text-independent model (Chengdu-FC) fine-tuned from a pretrained audio encoder using pseudo labels from Chengdu-MFA. Evaluation on expert-annotated data shows Chengdu-MFA reduces average phone boundary differences by 31.8% compared to Standard Mandarin baselines, while Chengdu-FC achieves a 61.2% reduction, demonstrating an effective bootstrapping pipeline for under-resourced varieties.
phonetic forced alignmentgmm-hmmg2p dictionarylow-resource languagepseudo label
From Static Bibliometrics to Dynamic Knowledge Graphs: An LLM-Powered Framework for Modernizing Science, Technology, and Innovation (STI) Analytics
The paper proposes a hybrid framework integrating bibliometrics, dynamic knowledge graphs, and large language models (LLMs) for modernizing science, technology, and innovation (STI) analytics. The framework organizes five layers: an open scholarly data backbone, a dynamic versioned knowledge graph, a constrained LLM-assisted semantic augmentation layer, a multi-layer validation pipeline, and an analytics layer. LLMs generate provisional candidate enrichments, which undergo structural, evidentiary, comparative, and expert validation before analytic use. The framework supports traditional bibliometric indicators and extended graph-based analyses like trend detection and policy gap analysis. It emphasizes validation as a mediating principle, balancing semantic flexibility with epistemic discipline.
bibliometricsknowledge graphsllmsvalidation pipelinesti analytics
Toward cryptographically verifiable authorization for autonomous AI agents: A security hypothesis, preliminary formal model, and proof-of-concept implementation
The paper proposes cryptographically verifiable authorization (CVA) for autonomous AI agents, formalized as a relation $R_{CVA}$ that binds agent principals, authorization requests, execution contexts, and policy satisfaction while preserving attribute confidentiality. It introduces a formal CVA abstraction with security properties (authorization soundness, principal binding, etc.) and implements a proof-of-concept using Groth16 zk-SNARKs. The work identifies the structural separation among identity binding, request binding, and runtime execution binding as a key open problem in agentic security.
cryptographically verifiable authorizationautonomous ai agentsgroth16 zk-snarksauthorization soundnessprincipal binding
GRADRAG: Cross-Component Prompt Adaptation for Coordinated Multi-Agent RAG
GRADRAG introduces cross-component prompt adaptation for multi-agent RAG systems by modeling the pipeline as a computational graph and propagating structured evaluation feedback to update upstream agents. The framework employs an Evaluator to critique answers and evidence, triggering a Prompt Optimizer to iteratively refine adaptive agents (retrievers, graph constructors, answerers) and enable early stopping. Evaluated on SQUALITY and QMSUM with both flat chunk-based (IRCoT-style) and graph-based retrieval, GRADRAG outperforms one-step refinement baselines by 12-15 percentage points in LLM-judged pairwise comparisons, achieving most gains within two iterations.
retrieval-augmented generationmulti-agent systemsprompt optimizationcomputational graphiterative refinement
PC-Edit: Prompt-Contrastive Region Discovery and Region-Guided Editing
PC-Edit introduces a prompt-contrastive framework for training-free MM-DiT editing, addressing object replacement challenges through explicit region discovery and selective feature preservation. The method contrasts image-token attention outputs under source/target prompts to localize semantic differences, identifying source-erasure and target-emergence regions. It couples region discovery with background preservation by injecting cached source K/V features outside edit regions during denoising. Evaluated on PIE-Bench and EditRegion-Bench, PC-Edit outperforms methods without user-specified edit regions in editing quality and background preservation.
prompt-contrastivemm-ditregion discoveryk/v featurestraining-free editing
Scaling Up Formal Representation of Clinical Trial Protocols in Ensemble Logic Using LLMs: A Preliminary Study
The study introduces CT-TEL, a pipeline using Large Language Models (LLMs) to automate the translation of unstructured clinical trial protocols into Temporal Ensemble Logic (TEL) formulas, addressing the bottleneck of manual encoding. The method processes 23 trials from ClinicalTrials.gov, evaluating translation fidelity through LLM-based back-translation and semantic similarity measurement. Results demonstrate semantic retention, suggesting LLMs can effectively map narrative protocols to computable logic for clinical trial emulation.
temporal ensemble logicclinical trial protocolslarge language modelsautomated reasoningsemantic similarity
AI Assistants Overassist
We introduce Int-Bench, a simulation-based benchmark for evaluating LLM interventions during learning, addressing the poorly understood dynamics of AI assistance in problem-solving. Int-Bench simulates a 'student' solving problems while a 'teacher' LLM monitors reasoning and decides on intervention timing and methods. Evaluations across code debugging, mathematics, and brain teasers reveal that LLMs intervene more frequently and earlier than humans, often providing complete solutions rather than targeted hints. This behavior prioritizes short-term task success over fostering deeper reasoning processes essential for long-term learning and generalization.
int-benchllm interventionssimulation-based benchmarkcognitive engagementgeneralization
Expert Behavior Prior Reinforcement Learning
The Expert Behavior Prior (EBP) algorithm is introduced to enhance sample efficiency and stability in online reinforcement learning by generating expert policy priors directly from the online replay buffer, avoiding reliance on static offline datasets. EBP employs a Q-guided conditional variational autoencoder (Q-CVAE) to produce high-value actions, an expert policy guidance (EPG) mechanism for selecting expert actions, and a policy gradient correction (PGC) module to align Q-guidance with expert supervision. Evaluations on robotic control (Gym, PyBullet) and industrial control (DMControl) benchmarks show EBP outperforms state-of-the-art online RL algorithms in sample efficiency and convergence stability.
expert behavior priorq-cvaepolicy gradient correctiononline reinforcement learningsample efficiency
Unlearning Under Imbalance: Benchmarking Fairness in Multimodal LLM Unlearning
The paper introduces FAIRGET, the first Visual Question Answering benchmark for evaluating multimodal large language model (MLLM) unlearning under imbalanced forget requests, simulating realistic demographic disparities. It proposes FAUN, a fairness-aware unlearning algorithm that employs bias-aware activation steering to maintain model fairness during unlearning. Experiments on FAIRGET and FIUBench show FAUN outperforms baselines in both unlearning efficacy and fairness preservation.
machine unlearningmultimodal llmsfairnessvisual question answeringbias-aware steering
An LLM-Driven Workflow for Automated Process Control Strategy Generation and Tuning from Dynamic Process Models
The authors propose a structured LLM-driven workflow for automated multivariable control design from dynamic process models, decomposing the task into constrained code-generation steps including MV-CV pairing, controller specification, and Bayesian optimization-based tuning. The workflow generates and validates executable artifacts iteratively, repairing failed components using validation feedback. Demonstrated on a nonlinear gas-preheater benchmark, the approach produces a decentralized PI feedback-feedforward control structure, with Bayesian optimization reducing the closed-loop performance objective by 26.5% through improved pressure-loop transient performance. Results validate the feasibility of LLM-based code generation for control-design workflows while highlighting the need for broader plantwide-control validation.
multivariable controlbayesian optimizationdynamic process modelsfeedback-feedforwardclosed-loop simulation
Multi-Task Learning for Heterogeneous Prediction from Video Game State with Transfer Learning
The paper proposes a multi-task learning (MTL) framework for heterogeneous prediction tasks in video game state data, leveraging shared representations to improve generalization and reduce computational costs. The method employs a multimodal architecture combining rasterized vision inputs, global match context, and per-unit state information via an image encoder and attention-based interaction modeling. Experiments on a proprietary World of Tanks dataset compare single-task and multi-task training, evaluate loss weighting strategies, and test transfer learning under limited data regimes. Results demonstrate the effectiveness of MTL and examine within-game transfer across maps under structured environment shifts.
multi-task learningvideo game stateattention-based modelingtransfer learningstructured environment shift
A Comparative Evaluation of Embeddings and LLMs in a Greek Book Publisher Setting - The CUP Dataset
The CUP benchmark introduces a Greek book retrieval dataset with 868 catalog records and 104 expert-annotated queries, enabling evaluation across lexical, semantic, noisy, and cross-lingual retrieval tasks. Sparse (BM25), dense (sentence-transformers), hybrid, and LLM-assisted methods were compared, with hybrid retrieval achieving the best overall performance. Multilingual embeddings outperformed Greek-specific models, BM25 excelled at named-entity queries, and dense/hybrid methods improved natural-language, noisy, cross-lingual, and concept queries. LLM techniques, including TOC summarization and post-filtering, enhanced specific retrieval stages at varying costs.
bm25sentence-transformershybrid retrievalllm-assisted retrievalcross-lingual queries
pAI-Econ-claude: A Gated Human-in-the-Loop Multi-Agent Architecture for AI-Assisted Economic Theory Development
We introduce pAI-Econ-claude, a gated human-in-the-loop multi-agent architecture for AI-assisted economic theory development, addressing reliability challenges when no machine-readable correctness signal exists. The system employs inspectable intermediate records, specialized gates for failure diagnosis, and human checkpoints for irreversible decisions. Evaluated on five economic-theory tasks against an ungated baseline, the gated architecture was preferred in four cases, reducing mean failure severity from 1.58 to 1.16 and increasing overall usefulness from 2.60 to 3.10. Key improvements included rejecting false premises and revising welfare claims. Results indicate gated oversight enhances auditability without replacing formal verification, emphasizing the importance of human judgment allocation over pure agent autonomy.
multi-agent architecturehuman-in-the-loopeconomic theoryfailure diagnosisauditability
BasketEvent: Understanding Who Did What and When in Basketball Videos
We introduce BasketEvent, a player-centric basketball event understanding dataset curated from NBA broadcasts, which grounds event labels to responsible players and provides 1,000 manually annotated samples with precise event intervals. We propose PlayNet, a player-centric reasoning framework that maps basketball videos to player-level event predictions with temporal evidence by tracking key entities, associating player identities, and modeling player-player, player-ball, and global court interactions while aggregating sparse temporal evidence via gated pooling. Experiments demonstrate that PlayNet significantly outperforms video-level and crop-based baselines, proving the superiority of player-centric modeling for fine-grained sports video understanding.
player-centric modelingtemporal evidence localizationgated poolingevent understandingfine-grained sports video
slang.gr as a Large-Scale Crowdsourced Resource for Non-Standard Greek
The study presents slang.gr as the first large-scale computational resource for Greek non-standard language, combining lexical data, user tags, and interaction patterns. A multi-layer taxonomy maps noisy folksonomic tags to structured semantic and sociolinguistic categories. Analysis reveals Greek slang's person/evaluation focus (morphological creativity) and skewed contributor dynamics (short lifespans, overlapping communities). A community-based confidence score integrates user roles and moderation signals, showing taxonomy-based representations improve interpretability while preserving behavioral structure. This enables sociolinguistic NLP, bias analysis, and informal language studies in LLMs.
folksonomic tagsmulti-layer taxonomymorphological creativitysociolinguistic nlpconfidence score
Logic Programming Semantics for Causal Processes
The paper establishes a correspondence between logic programming semantics and the eventual states of causal processes, addressing modeling challenges in life sciences. It demonstrates that stable models of positive logic programs align with eventual states from neutral initial conditions under indefinite undisturbed evolution, while supported models capture states reachable from arbitrary starting points. This analysis contributes to interpreting logic programming semantics as a causal rule language, introducing a temporal perspective to existing explanatory frameworks for stable and supported model semantics.
logic programmingstable modelssupported modelscausal processestemporal semantics
ICAE-Bench: Evaluating Coding Agents as Interactive Project Builders
ICAE-Bench introduces a novel benchmark for evaluating coding agents in interactive project-building scenarios, addressing the shift from static code completion to dynamic software construction from incomplete specifications. The benchmark simulates realistic development workflows through fuzzy product requirements derived from precise open-source repositories, employing an automated User Agent for interaction grounded in User Agent Data to maintain reproducibility. Evaluation combines standardized black-box tests with multi-dimensional diagnostics assessing functional correctness, semantic similarity, API alignment, structural fidelity, design quality, and interaction quality. This approach enables comprehensive assessment of agents' abilities in planning, requirement clarification, tool use, debugging, and repository-level construction.
coding agentsinteractive project-buildingfuzzy product requirementsuser agent datamulti-dimensional diagnostics
Explainable Belief Harmonization under Dynamic Epistemic Partitions
The paper introduces a formal framework for multi-agent belief harmonization under dynamic epistemic partitions, where agents' observational capacities may change during execution. The hybrid approach combines answer set programming's declarative constraints and explanation capabilities with Python's numerical flexibility, handling continuous belief profiles across heterogeneous resolution levels. Results demonstrate formal guarantees including admissibility preservation under refinement, mass-preserving repair under coarsening, and complete violation detection across 100 randomized topology changes.
belief harmonizationepistemic partitionsanswer set programmingdeclarative integrity constraintsmass-preserving repair
Explainability Framework for Policy-Aware Autonomous Agents
The paper proposes an explainability framework for policy-aware autonomous agents that generate contrastive explanations by leveraging policy violation penalties. The framework combines Answer Set Programming for rule-based reasoning with Python for information extraction and natural language generation, informed by social science principles for effective explanations. Evaluation via human participant surveys demonstrates the framework's ability to produce comprehensible counterfactual explanations (e.g., justifying actions by citing avoided undesirable events).
explainable aipolicy-aware agentscontrastive explanationsanswer set programmingcounterfactual reasoning
How Rules Represent Causal Knowledge: Causal Modeling with Probabilistic Logic Programming
This paper formalizes Pearl's causal reasoning framework within probabilistic logic programming (PLP), extending it beyond Bayesian networks to handle simultaneous events without temporal dependencies. The authors propose a formal causal semantics for PLP, including an intervention mechanism, and implement this approach. They demonstrate alignment with P-log semantics for stratified ProbLog programs, while noting potential divergences in non-stratified cases and other PLP formalisms. This work bridges causal modeling and logic programming, enabling intervention-based predictions in a broader class of probabilistic logical systems.
probabilistic logic programmingcausal semanticsinterventionbayesian networksp-log semantics
A New Well-Supported Semantics for Description Logic Programs
We introduce a new well-supported semantics for description logic programs (DLPs) that addresses limitations of existing approaches. The proposed semantics evaluates ontological atoms more strictly, maintaining NP-complete complexity for consistency checking rather than increasing it to the second level of the polynomial hierarchy. We characterize this semantics using a fixpoint operator and reduct-based transformation, identifying a syntactic class of DLPs where it aligns with current semantics. The new semantics forms a strict subset of existing well-supported semantics, preserving cyclic dependency avoidance while introducing stricter ontological evaluation. This approach aligns more closely with traditional logic programming principles.
description logic programswell-supported semanticsontological atomsreduct transformationfixpoint operator
Hybrid MKNF with Classical Negation in the Rule Component
The authors extend Hybrid MKNF knowledge bases to support classical negation in the rule component, addressing a limitation in representing explicit negative knowledge crucial for safety-critical applications. They formally define the syntax and semantics of the extended language and introduce a general procedure for computing its well-founded model. This enhancement enables more precise reasoning by distinguishing between explicit negation and the absence of information, improving the framework's applicability in domains requiring rigorous negative knowledge representation.
hybrid mknfclassical negationwell-founded semanticsdescription logicslogic programming
Bound-Founded Semantics for Answer Set Programming with Difference Constraints: Preliminary Report
The paper introduces a many-sorted variant of Bound-founded Logic of Here-and-There (HTb) to unify semantic foundations for Answer Set Programming (ASP) extensions with linear constraints. By formalizing foundedness for numeric variables, the framework characterizes equilibrium models across hybrid systems like clingo[DL], clingcon, and flingo. This approach reveals semantic distinctions in constraint atom justification, providing a consistent basis for analyzing program simplifications and future semantic integrations.
answer set programmingbound-founded logicdifference constraintsequilibrium modelslinear constraints
Towards a Certifying Grounder
The paper introduces CertiFOX, a certifying grounding framework for first-order logic model expansion (FOX) over finite domains, addressing the trust gap in declarative solving. The framework comprises (1) a proof format for grounding derivations, (2) GroundFOX, a certifying grounder operating on Grounding Normal Form (GNF), and (3) CheckFOX, an independent proof checker. Experimental results show that GroundFOX performs comparably to other grounders, with CheckFOX adding minimal overhead (within a small constant factor of grounding time), ensuring trustworthy end-to-end certified solving pipelines.
certifying grounderfirst-order logicgrounding normal formmodel expansionproof checking
Declarative Problem Solving in UAM Strategic Deconfliction
An Answer Set Programming (ASP) approach is proposed for strategic deconfliction in Urban Air Mobility (UAM), addressing time synchronization and route optimization to ensure conflict-free flight plans. The method is benchmarked against Constraint Programming (CP), evaluating scalability and resource efficiency. Results indicate ASP achieves faster execution and superior scalability for small to medium scenarios, while CP maintains stable memory usage but performance degrades with increasing complexity.
answer set programmingurban air mobilitystrategic deconflictiontime synchronizationroute optimization
Case study: solving P-99 with LPTP and an LLM
The study demonstrates automated solving of 33 Prolog problems from P-99 using Claude (Anthropic) for code generation and LPTP for formal verification. The method combines LLM-based 'vibe-coding' (English-to-Prolog translation) with 'vericoding' (formal proofs of type safety, termination, and functional correctness). Results include 58 generated procedures, 508 tests, and 257 lemmas (11,800 proof lines), all manually validated. The hybrid approach achieved full verification coverage for groundness, uniqueness, and existence properties.
prologllmformal verificationtheorem provingvibe-coding
Chess\_db: A framework for working with large chess game datasets
Chess_db introduces a logic programming framework for manipulating large chess game datasets, enabling efficient storage and retrieval of game information. The system processes Portable Game Notation (PGN) files to construct databases, leveraging open-source key-value stores for rapid access to position tables across extensive game collections. This approach facilitates both in-memory operations and persistent database creation, supporting analysis of player histories and position-specific outcomes. The framework addresses the growing computational needs in chess training and analysis, where engine outputs and historical game data are increasingly important resources.
logic programmingportable game notationkey-value databasesposition tableschess analysis
Animation, Verification and Visualisation of Prolog Transition Systems with ProB
ProB introduces extended capabilities for animating Prolog transition systems, including statistical simulation, reliable trace replay, user-input transitions, and enhanced state visualization. These features leverage ProB's Prolog-based model checking and constraint-solving framework to validate high-level formal specifications. The extensions are demonstrated through case studies, notably in evaluating game strategies like Connect Four, and are applicable to Event-B proof obligations and educational models with interactive visualization.
prologtransition systemsmodel checkinganimationvisualization
Encoding Event-B Proof Rules in Prolog: An Interactive Sequent Prover for ProB
The authors present an interactive sequent prover for Event-B by encoding over 600 proof rules in Prolog and integrating them into the ProB validation tool. This approach enables systematic proof analysis, construction, and visualization, with advantages for teaching through direct rule selection. The system supports multiple export formats, including proof replay traces and interactive HTML documents, and interoperates with the Rodin platform. Compared to a prior Java implementation, the Prolog encoding is more compact and maintainable. While currently limited to short proofs via iterative deepening, future work aims to develop faster automatic provers.
event-bprologsequent proverproof rulesrodin platform
Identifying Good Rules for Efficient SAT Encodings of Single-Constant Multiplication Using Machine Learning
The authors present a neuro-symbolic framework to accelerate SAT encodings for the Single Constant Multiplication (SCM) problem, an NP-hard optimization task in hardware design. Their method combines a graph neural network to predict promising operator types during constant decomposition with symbolic search pruning based on confidence scores. Experiments on 17-32 bit constants show 10-100× faster encoding, >97% memory reduction, and 10× fewer branches while maintaining near-optimal addition counts.
single constant multiplicationsat encodingneuro-symbolicgraph neural networkhardware optimization
Case study: proving sqrt(2) irrational with LPTP and an LLM
The study demonstrates a collaborative approach to formal theorem proving by integrating a Large Language Model (LLM) with the Logic Program Theorem Prover (LPTP). Starting from basic logic programming predicates, the authors outline a proof of the irrationality of the square root of 2 using LPTP's natural deduction-based proof language. The LLM assists in generating parts of the proof, which is subsequently verified by LPTP. The result is a complete, formally checked proof, showcasing the potential of LLMs in augmenting automated theorem proving workflows.
large language modellogic programmingtheorem provingnatural deductionformal verification
Differentiable Logic Programming to Mitigate Reasoning Shortcuts in Neurosymbolic Systems
The paper proposes a matrix-based differentiable logic programming method to mitigate reasoning shortcuts in neurosymbolic (NeSy) systems, addressing constraint satisfaction and cognition shortcuts. The approach unifies rules and constraints in a single matrix representation, drawing connections to fuzzy logic t-norms while analyzing gradient flow properties. Experiments on MNIST variants demonstrate that one-to-one grounding of neural outputs to logical atoms reduces both shortcut types by 37-42% compared to soft probability distribution methods, highlighting the importance of architectural choices in neural-symbolic integration.
neurosymbolic systemsdifferentiable logic programmingreasoning shortcutsmatrix-based semanticsfuzzy logic t-norms
Explaining Weather Bulletins via ILP
The paper presents an ILP-based pipeline for generating interpretable hypotheses to explain weather bulletins issued by OSMER FVG. Using the FastLAS2 framework, the method processes simulated meteorological raw data and expert-issued bulletins as ground truth, extracting ASP facts and generating ILP examples. From these examples, FastLAS2 infers explanatory hypotheses, which are translated into natural language to clarify the rationale behind experts' symbol choices in forecast pictograms. The approach is generalizable and applicable to bulletins from various regions and sources.
inductive logic programmingfastlas2asp factsinterpretable hypothesesweather bulletins
Representative Sets in Propositional Abduction
The paper investigates the representational capacity of solution sets in propositional abduction, specifically whether a given set of explanations S can represent any other explanation within a bounded symmetric difference k. Employing classical and parameterized complexity analyses, the authors provide a complete complexity classification, revealing limited tractability but less severe complexity increases than anticipated. Key findings include novel tractable cases and hardness results, with connections to the covering radius problem in coding theory emerging as unexpectedly relevant to non-monotonic reasoning.
propositional abductionsolution spacesymmetric differenceparameterized complexitycovering radius
Safeguards for Speech2Speech LLM-Assistants: A Case Study in Automotive Applications
The paper evaluates safeguards for speech-to-speech (S2S) conversational assistants in automotive applications, comparing transcript-based and tool-based approaches. Both methods introduce latency (0-1.4s per response) and face technical challenges like non-deterministic tool behavior, making them unsuitable for industrial deployment. The study identifies open challenges for implementing effective guardrails in S2S systems while maintaining real-time performance.
speech-to-speechguardrailslatencynon-deterministicautomotive
SafeStep: AI-powered Travel Assistance for Elderly People with Frailty or Dementia
SafeStep introduces an AI-powered travel assistance system for elderly individuals with frailty or dementia, utilizing a novel travel graph representation that combines route planning with predictive modeling. The system employs LLMs and the Anticip8 behavioral prediction engine to generate personalized failure scenarios, propose interventions, and estimate their impact on journey success probabilities. Evaluation on 26 real-world journeys demonstrated that integrating Anticip8 with GPT-based models yields reliable performance, with user feedback indicating improved confidence and perceived safety, though interface usability requires enhancement.
travel graphbehavioral predictionllmsintervention evaluationpredictive modeling
CRAG-MM-Diagnostics: Enabling Stage-Wise Analysis of Knowledge-Intensive VQA
CRAG-MM-Diagnostics introduces a diagnostic benchmark for Knowledge-Intensive Visual Question Answering (KI-VQA) that enables stage-wise analysis of the pipeline, isolating language-based visual grounding, object identification, and knowledge retrieval and reasoning. The benchmark evaluates fully parametric and retrieval-augmented Vision-Language Models (VLMs) using metadata such as target ROIs, entity names, and visual complexity scores. Results identify knowledge retrieval and reasoning as the primary bottleneck, with additional challenges in target object identification and textual cue integration. A grounded bimodal RAG pipeline, incorporating a visual grounding module, improves GPT-5 and Qwen's accuracies by 13.3 and 8.5 percentage points, respectively.
knowledge-intensive vqavisual groundingobject identificationretrieval-augmented vlmsstage-wise analysis
V-DEAL: Diagnosing Video Safety De-Calibration as an Understanding-Refusal Coupling Failure
V-DEAL introduces a three-level diagnostic framework to analyze safety vulnerabilities in Video Large Language Models (Video LLMs), focusing on understanding-refusal coupling failures. The method examines model behavior, understanding, and internal representations, ruling out perception failure and quantifying refusal tendencies. Testing six Video LLMs on three benchmarks reveals that models correctly identify harmful video content with 81% accuracy but exhibit a 48.33% attack success rate when harmful videos are paired with benign queries. Hidden-state analysis shows weaker refusal tendencies in visual versus textual understanding. A prompt injection intervention reduces attack success rates by 48.24 percentage points, matching fine-tuning-based methods.
video large language modelsunderstanding-refusal couplinghidden-state analysisprompt injectionsafety alignment
One More Turn, Less Regret: A Regret-Based Multi-Turn Benchmark for LLMs' Clarification Policies
The paper introduces RegretBench, a multi-turn benchmark for evaluating LLM clarification policies through regret-based metrics rather than isolated question quality. It formulates ambiguity as hidden-intent resolution, supports free-form interaction with semantic-state tracking, and measures value loss relative to a reference policy. Experiments in open-domain QA and product recommendation show that models with similar accuracy differ in efficiency, robustness, and stopping decisions, revealing that effective clarification requires timely, relevant questions and proper termination.
regret-based evaluationmulti-turn interactionclarification policyhidden-intent resolutionsemantic-state tracking
Demographically-Informed Heat-Mortality Risk Curves via Risk Graph Neural Networks
The authors introduce Risk Graph Neural Networks (RGNNs), a hierarchical GNN encoder that integrates demographic and geographic context into heat-related mortality risk estimation. RGNNs optimize Distributed Lag Non-linear Models (DLNMs) coefficient vectors using granular census features, preserving interpretable risk curve outputs while enhancing predictive calibration. Evaluated across 10 regions of England and Wales during unprecedented heat years, RGNN variants demonstrate lower point-errors and near-nominal uncertainty coverage, particularly during the 2022 heatwave where baseline models fail.
risk graph neural networksdistributed lag non-linear modelspredictive calibrationhierarchical gnn encoderheat-related mortality
Hardware-Software Co-Design for Float16 On-Device Training on RISC-V Single-Core
The authors present a hardware-software co-design framework enabling float16 on-device training on RISC-V single-core processors using Zfh and Zvfh extensions. Their open-source solution, built upon the AIfES embedded DNN framework, reduces memory footprint by ~50% versus float32 with minimal accuracy loss while supporting layer freezing for transfer learning. FPGA implementation demonstrates modest hardware overhead (+1.15% LUT6, +0.05% FF at 175MHz) for Zfh support.
float16risc-von-device traininghardware-software co-designembedded dnn
Relative Value Learning
We propose Relative Value Learning (RV), a reinforcement learning framework that directly estimates value differences via an antisymmetric function Δ(s_i, s_j) = V(s_i) - V(s_j), leveraging pairwise Bellman operators and deriving 1-step, n-step, and λ-return targets. RV reconstructs generalized advantage estimation (R-GAE) to provide an unbiased policy-gradient estimator. Integrated with Proximal Policy Optimization (PPO), RV achieves competitive performance on the Atari benchmark (49 ALE games), demonstrating its efficacy as an alternative to absolute value critics. Theoretical analysis confirms RV's γ-contraction property and unique fixed point convergence to true value differences.
relative value learningantisymmetric functionpairwise bellman operatorgeneralized advantage estimationproximal policy optimization
GlucoTune: A Unified Framework for Blood Glucose Preprocessing, Forecasting, and Benchmarking in Diabetes
GlucoTune introduces a unified framework for reproducible experimentation with blood glucose time-series data, addressing challenges in preprocessing, forecasting, and benchmarking for diabetes management. The framework standardizes workflows through configurable YAML pipelines, enabling consistent data handling without redistributing sensitive preprocessed datasets. It provides a unified interface for implementing, training, and evaluating blood glucose prediction models, integrating public datasets and state-of-the-art forecasting methods. GlucoTune includes a benchmarking leaderboard for systematic comparison of experimental settings. Comprehensive experiments and a user study demonstrate its effectiveness and usability.
blood glucosetime-seriespreprocessingbenchmarkingforecasting
TOUR: A Trajectory-Level Unlearning Benchmark for Offline Reinforcement Learning
TOUR introduces a benchmark for trajectory-level unlearning in offline reinforcement learning, addressing challenges in evaluating data deletion. The benchmark combines trajectory partitioning, non-member controls, retraining references, performance anchors, and multi-attack privacy auditing. Experiments on D4RL locomotion and AntMaze demonstrate environment-dependent privacy-utility trade-offs, with retraining and fine-tuning outperforming uniform GA+Refit. Multiple attack methods reveal that single likelihood-based membership scores overstate deletion quality, emphasizing the need for matched non-member construction and attack-family-specific auditing.
offline reinforcement learningtrajectory-level unlearningprivacy auditingd4rl locomotionmulti-attack
AttriMem: Attribution-Guided Process Feedback for Agent Memory Learning
AttriMem introduces an attribution-guided process-feedback framework for learning memory-construction policies in LLM agents using reinforcement learning. Unlike heuristic methods and coarse RL-based approaches, AttriMem augments global outcome rewards with local token-level contribution signals, addressing the fine-grained credit-assignment bottleneck in memory decisions. This method avoids predefined ground-truth targets and adapts to uncertain reasoning trajectories. Experiments on long-horizon dialogue question answering demonstrate that AttriMem outperforms retrieval-based, heuristic, and RL-based baselines, generalizes across benchmarks and answer models, and stabilizes RL optimization.
memory-construction policyreinforcement learningcredit-assignment bottlenecktoken-level contributionslong-horizon dialogue
Can Generative Recommendation Reach Cold Items? A Temporal Perspective on Semantic-ID Generation
This work investigates the limitations of Semantic-ID (SID) based generative recommendation systems in addressing temporal cold-start scenarios for new items with unseen tokens or weakly supported SID paths. Through seen/unseen-hit analysis, coldness taxonomy, and oracle-prefix probing under an absolute-time temporal protocol, the study demonstrates that while SID models can occasionally reach future items supported by observed tokens and prefixes, they struggle with unseen atomic tokens and unsupported SID paths. The findings reveal that SID generation operates through hierarchical semantic bucketing, with early tokens selecting coarse regions and later tokens refining item-specific paths, suggesting directions for more independent SID spaces and dynamic textual context integration.
semantic-idcold-starttemporal protocolhierarchical semantic bucketingoracle-prefix probing
Training Large Language Models for Self-Explanation Faithfulness
The paper introduces a Reinforcement Learning (RL) method to directly optimize the faithfulness of large language models' (LLMs) self-explanations by modifying existing faithfulness metrics into an RL objective. Experiments with Llama3.1-8B and Qwen3-8B demonstrate substantial improvements on the Phi-CCT faithfulness metric, achieving scores up to 0.664 in-distribution and 0.691 out-of-distribution on StrategyQA. Cross-intervention generalization is model-dependent, with Llama3.1-8B showing non-zero transfer from random-word to user-bias insertions, while Qwen3-8B does not.
reinforcement learningfaithfulness metricsself-explanationllmphi-cct
Faster IndexTTS-2: Accelerating and Streaming Autoregressive Zero-Shot Text-to-Speech Synthesis on GPUs
Faster IndexTTS-2 accelerates the autoregressive IndexTTS-2 text-to-speech model for GPU deployment using TensorRT and TensorRT-LLM optimizations, adding streaming synthesis and batched inference. The method optimizes all components (GPT, Diffusion Transformer, vocoder) with minimal quality degradation. On the Seed-TTS benchmark, it achieves 5.0× GPT speedup and 3.6× end-to-end speedup while maintaining comparable word error rate, speaker similarity, and naturalness in English and Chinese.
autoregressive ttstensorrtstreaming synthesisbatched inferencediffusion transformer
HiMe: Real-Time Self-Hosted Personal Agent Platform for Health Insights with Wearable Devices
The paper introduces HiMe, an open-source locally deployable platform for real-time personal health agentic analysis using wearable devices. The system combines a database-first architecture with joint optimization of effectiveness and efficiency to enable privacy-preserving, continuous health monitoring. Key innovations include real-time processing of wearable data streams alongside long-term user modeling, achieving a Pareto-optimal balance between computational cost and analytical performance for personalized health insights.
personal health agentwearable devicesreal-time processingprivacy-preservingpareto-optimal
EmoAgent-R1: Towards Multimodal Emotion Understanding with Reinforcement Learning-based Dynamic Agent Specialization
We propose EmoAgent-R1, a Reinforcement Learning-based Dynamic Agent Specialization framework for multimodal emotion understanding in MLLMs. The method employs a cold start strategy with synthetic answer-conditioned chain-of-thought data, followed by RL-based training for dynamic agent selection and specialization. A novel Progressive Group-Relative Policy Optimization (P-GRPO) is introduced to transform sparse rewards into fine-grained learning signals, addressing coarse-grained credit assignment. Experiments on MER benchmarks demonstrate EmoAgent-R1's superior emotion reasoning performance and optimization stability compared to existing MLLM-based approaches.
multimodal emotion recognitionreinforcement learningdynamic agent specializationprogressive group-relative policy optimizationchain-of-thought
Reexamining zero-shot summarization: Empirical investigation of trustworthiness of LLM-summarizers
The study introduces a two-level diagnostic protocol to assess the trustworthiness of zero-shot LLM summarization by evaluating summary stability. At the document level, stability coefficients are computed via semantic and factual alignment metrics across multiple LLM-generated summaries under controlled conditions; corpus-level stability indices are then derived from stratified sampling. Empirical evaluation of three LLM-summarizers across three document genres reveals statistically significant variability in generation stability, highlighting the need for more robust summarization systems.
zero-shot summarizationstability coefficientfactual alignmentllm-summarizersgeneration-level variability
Naju: A Native Discrete State-Space Model with Independent Retention and Writing for Long-Sequence Memory
Naju introduces a native discrete state-space model for long-sequence memory tracking, decoupling retention and writing through explicit factorization of the recurrent update into a learned forget gate, independent write gain, and input-dependent maps. This design avoids the stability regularizer required in continuous-time-parameterized SSMs like Mamba by ensuring Schur-stability through sigmoid poles and fading-memory bounds. Empirical evaluation shows Naju maintains strong performance on both retention and overwriting at 4x training length, outperforming Mamba on WikiText-103 language modeling, Long Range Arena, and multi-query associative recall while preserving linear-time, linear-memory scaling.
state-space modellong-sequence memoryschur-stabilityfading-memorylinear-time scaling
Workflow-Localized Mechanism Learning: Attribution-Guided Repair and Knowledge Reuse for Structured Agent Skills
The paper introduces Workflow-Localized Mechanism Learning (WML), a method for attribution-guided repair and knowledge reuse in structured agent skills. WML employs Node--Mechanism Attribution to identify failed workflow nodes and implicated mechanisms, routing defects to appropriate resources (L3 for single-mechanism, L2 for relational defects). A six-module Workflow-Guided Skill Optimization (WGSO) loop selects third-party knowledge, applies patches, and stores verified outcomes. Evaluated on SpreadsheetBench, WML achieves 90.33 and 74.67 Hard Accuracy with DeepSeek and Qwen3.6-Flash, respectively, and transfers to WikiTableQuestions with 84.00 and 83.00 Denotation Accuracy. On Compiler-Supported50, WML attains the highest hard-PASS rate and lowest cost per task.
workflow-localized mechanism learningnode-mechanism attributionworkflow-guided skill optimizationstructured agent skillsknowledge reuse
Sparse Concept Channels in Frozen 3D CT Vision Encoders
The study introduces a training-free Concept Channel Probe (CCP) method to identify sparse concept channels in frozen 3D medical vision-language models, demonstrating their role in encoding clinical findings. By probing frozen vision embeddings in the Pillar-0 and Merlin models, the authors show that each radiological finding is encoded by approximately 10 sparse channels, which match full-feature classification performance and outperform zero-shot text prompting. Disabling these channels collapses the corresponding finding's score while leaving unrelated labels stable. CCP, paired with a corpus-derived report template, outperforms CT-CHAT in clinical efficacy and NLG metrics (F1 0.549 vs. 0.184; BLEU 0.483 vs. 0.373) with 22x lower latency.
concept channel probefrozen encoderssparse channelsmedical vision-language modelsclinical efficacy
HyWorldVLA: A Vision-Language-Action Model with Hybrid World Modeling for Autonomous Driving
HyWorldVLA introduces a hybrid vision-language-action model for autonomous driving that combines pixel-level supervision with latent representation learning to address the robustness-interpretability trade-off in world modeling. The framework employs a two-stage approach: pre-training with joint video latent prediction and frame reconstruction, followed by co-fine-tuning where latent features drive an action expert for trajectory generation. Evaluations on NAVSIM v1/v2 show superior performance over pixel-based and latent-based baselines, with novel noise robustness analysis establishing a new architectural benchmark.
vision-language-actionworld modelingautonomous drivinglatent representationtrajectory generation
GuardianAgentBench: Where Agents Fail and How to Guard Them
We introduce GuardianAgentBench (GABench), a benchmark evaluating LLM agent safety across 580 scenarios in six domains, tested on LangChain, LlamaIndex, and Vectara frameworks. The benchmark employs multi-stage validation and five adversarial attack modes, revealing that even top-performing configurations achieve only 74.8% accuracy. Experiments identify two failure regimes: stronger models under-call tools, while weaker models mis-select and over-call them. Performance declines with tool-set size and sequential turn depth, with long-horizon planning posing significant challenges. Our guardrail implementation recovers 19.9% of failures at a 0.5% false positive rate, demonstrating that execution-time structural intervention enhances safety without disrupting correct behavior.
guardianagentbenchlangchainllamaindexvectaraadversarial attack
Beyond Independent Optimization: Compression, MoE Routing, and Quantization Interactions in Multimodal Edge Intelligence
The paper analyzes interactions between compression, MoE routing, and quantization in multimodal edge intelligence, demonstrating that these techniques cannot be optimized independently. It reviews visual token compression, KV-cache optimization, MoE routing, and low-bit quantization, highlighting how each affects downstream components (e.g., compression alters MoE routing decisions, quantization impacts router logits). Key trade-offs include accuracy versus token budget and sparse routing efficiency versus expert collapse. The authors introduce Temporal Routing Consistency for video MoE evaluation and identify open challenges in routing-aware compression and hardware co-design.
multimodal edge intelligencemixture-of-expertskv-cache optimizationvisual token compressionlow-bit quantization
Delivery, Not Storage: Cue-Anchored Working Memory as a Harness Property for Coding Agents
The paper proposes a two-tier memory architecture for coding agents, emphasizing cue-anchored working memory as a harness property rather than agent-controlled storage. Grounded in cognitive science principles (memory offloading, incidental encoding, event-based prospective memory), the design features first-class trigger conditions (path, symbol, semantic, event, temporal) evaluated deterministically by the harness. Experiments on real coding tasks show voluntary memory usage near zero (0 operations in 114 turns), 39% re-read efficiency, and robust fact retention through 138 compact-resumes when using harness-controlled delivery versus rapid decay in agent-managed storage.
cue-anchored memorymemory offloadingincidental encodingprospective memorydeterministic injection
From Scalars to Time Series: Rethinking Implicit Neural Representations for Time-Varying Volumetric Data
The paper proposes a reformulation of implicit neural representations (INRs) for time-varying volumetric data, shifting from coordinate-wise scalar supervision to sequence-level time series learning at each spatial location. This approach eliminates dense spatiotemporal sampling, instead training INRs on full temporal evolutions per spatial point, compatible with existing architectures. Results show improved reconstruction quality and reduced training costs, with further gains when combined with mixture-of-experts (MoE) architectures for heterogeneous temporal dynamics.
implicit neural representationstime-varying volumetric datasequence-level supervisionmixture-of-expertstemporal dynamics
Interaction Dynamics Modeling and Predictive Control for Safe Steerable Catheter--Tissue Interaction
The authors propose a predictive control framework for safe steerable catheter-tissue interaction, addressing motion tracking, compliance, and contact-force constraints. They formulate interaction dynamics in the scalar tip-normal coordinate, using a partial-physics feedforward to cancel nominal bending dynamics and a predictive optimizer to regulate interaction states under hard constraints. An augmented Kalman filter compresses disturbances into a sensor-free state, enabling offset-free regulation. In MuJoCo simulations, the constrained controller achieves 0.47 N contact force (below the 0.5 N bound) while maintaining tracking accuracy, reducing free-space approach error by 90%. The framework demonstrates robustness to cardiac motion, though hardware validation remains future work.
predictive controlcatheter-tissue interactionkalman filtermujoco simulationcontact-force constraint
Clustered Edge Intelligence: Beyond Just Convergence of Edge Computing and AI
The paper proposes Clustered Edge Intelligence (CEI), a novel intelligence-centric framework that treats derived intelligence as a first-class, independently manageable entity across edge-cloud systems. CEI enables intelligence to be semantically represented, discovered, observed, shared, and dynamically clustered across heterogeneous edge devices. The authors present a three-layer architecture and identify key research dimensions including semantic knowledge representation, intelligence inventories, lifecycle automation, and clustering mechanisms. The approach aims to address limitations in current edge computing research by facilitating intelligence reuse and interoperability.
edge intelligenceknowledge representationclustering mechanismslifecycle automationinteroperability
SciExplore: Evaluating Autonomous Agents from Scientific Navigation to Information Integration
The paper introduces SciExplore, a novel benchmark for evaluating autonomous agents and LLMs on scientific information-seeking workflows, comprising 103 expert-curated tasks across four progressively complex task types: database navigation, ambiguous literature retrieval, missing reference completion, and cross-source knowledge synthesis. The benchmark spans ten scientific disciplines and assesses capabilities from entity-level reasoning to domain-level synthesis. Evaluations of ten state-of-the-art models reveal substantial performance degradation with increasing task complexity, particularly on structured synthesis tasks where accuracy is extremely low, exposing critical limitations in current systems for scientific research applications.
scientific information-seekingautonomous agentsbenchmark evaluationknowledge synthesisllm reasoning
Representing Entity Importance in AI Knowledge Systems: A Dual-Signal Framework of Audience Evaluation and Structural Authority
The study proposes a dual-signal framework for representing entity importance in AI knowledge systems, distinguishing between audience evaluation and structural authority dimensions. Using movie entities as a test case, the method combines IMDb ratings (audience evaluation) with Wikipedia hyperlink-derived PageRank scores (structural authority). Analysis of 482 entities and 13,690 relationships reveals weak correlation (Spearman ρ=0.2275) and limited overlap (10% in top 10, 34% in top 100), demonstrating the non-redundancy of these signals for task-aware knowledge representation.
entity importanceaudience evaluationstructural authorityknowledge representationpagerank
Scientific exploration, collaboration and labor division in the large language model era
This study examines how large language models (LLMs) influence scientific research strategies and collaboration patterns by analyzing 775,323 scientists' publication histories and 137,120 multi-author papers' CRediT contributions. Post-2022, scientists exhibited increased interdisciplinarity and exploration, particularly among established researchers and those from non-English-speaking LMICs. Authors with stronger AI-writing signals widened their interdisciplinary gap, while collaboration networks became more diverse but less tied to collaborators' disciplines. Team roles grew more specialized, with narrower contributions and increased software/validation roles, suggesting a shift toward distinct responsibilities and reduced interdependence.
large language modelsinterdisciplinaritycollaboration networkscredit contributionscientific workflows
OPOD: On-Policy Omni Distillation
The paper introduces On-Policy Omni Distillation (OPOD), a method for improving omni-modal models by coordinating modality-specific teachers during distillation. OPOD routes student responses to matching text, image, or audio teachers, selectively retains teacher guidance where the teacher's probability exceeds the student's, and independently adjusts each modality's influence. Evaluated across twelve benchmarks and three model sizes (30B, 51.7, 46.2), OPOD achieves state-of-the-art performance, surpassing both base models and joint multimodal training by 1.7-2.1 points. The 30B variant outperforms specialists on eleven benchmarks while maintaining a single deployable model.
omni-modalon-policy distillationmultimodal routingprobability thresholdingmodality balance
Traceable Scholarship: Page Anchors and Ariadne's Thread for Humanistic Inquiry in the Age of Generative AI
The paper proposes Traceable Scholarship as a normative framework for AI-assisted humanistic research, addressing the epistemic risks of generative AI's fluent but unsourced outputs. It introduces technical components including page anchors, dual page numbers, and citation-first generation, implemented in the AIH-Infra system with three layers (Contexture, Open WebUI, MCP Server). A case study on Kant's Akademie-Ausgabe demonstrates traceability enabling retrieval correction and evidence grading. The approach positions source attribution as fundamental to maintaining refutable scholarship in generative AI contexts.
traceable scholarshippage anchorscitation-first generationaih-infraevidence grading
Source-Prior-Driven Selective Adaptation for Efficient Diffusion Model Finetuning
We propose a source-prior-driven selective adaptation method for efficient diffusion model fine-tuning, addressing the trade-off between target-specific generation and preservation of pretrained generative capabilities. The method leverages two insights: (1) parameter-specific inconsistency in general capability loss and (2) structural inconsistency of low-impact parameters across layers and types. It employs a static mask to identify suitable parameters for downstream adaptation and constructs structured update strategies for the selected subset. Experimental results demonstrate superior adaptation-retention trade-offs compared to existing baselines.
diffusion modelsselective adaptationparameter-efficient fine-tuningstatic maskstructured updates
Multi-turn RL with Structural and Performance Aware Rewards for CUDA Kernel Generation
CudaPerf introduces a multi-turn reinforcement learning framework for CUDA kernel generation, combining verifiable execution rewards with structural code-aware rewards derived from parallelization features. The method operates in two stages: an offline pairwise ranking module for distinguishing program candidates and an online RL training phase optimizing correctness, performance, and structural efficiency. Iterative refinement using execution feedback enhances learning. Evaluated on a dataset of 2.9k C to CUDA and 1k PyTorch to CUDA programs, CudaPerf achieves up to 5X speedup and 17% correctness improvements, outperforming Qwen-3-32B and CUDA Agent baselines.
reinforcement learningcuda kernelparallelization featuresiterative refinementexecution feedback
Anti-Goal Reasoning: Rethinking the Theory of Goal Reasoning in Non-Axiomatic Logic
(No summary returned.)
Is Deep Research Reliable? Misleading Knowledge Induces False Conclusions
The study investigates reliability vulnerabilities in Deep Research agents, demonstrating how misleading knowledge propagates through long-horizon workflows to produce false conclusions. Using MisKnow-Agent, a framework generating 5,933 controlled misleading instances, experiments reveal that even limited exposure induces false-conclusion adoption in reports, despite verifier models identifying them during focused validation. Evaluations of pre- and post-research defenses show partial mitigation, suggesting the need for integrated verification at both model and framework levels.
deep research agentsmisleading knowledgefalse-conclusion adoptionworkflow-level evidenceverifier models
TwistedMerge: Certified Higher-Order Diagnostics and Abstention for Model Merging
TwistedMerge introduces a certified framework for model merging diagnostics and abstention, formulating merging as a finite descent problem with checkpoints as local objects and alignment maps as transitions. The method employs a conservative certification pipeline that tests for inverse-consistency, coefficient-identification, centrality, and closure before promoting residuals to cohomology classes, otherwise abstaining. Theoretical results include no-go theorems and error-control guarantees, while empirical validation shows cycle residuals do not predict merge degradation in natural checkpoint collections. The work positions descent theory as a falsifiable certification framework.
model mergingfinite descentcohomology classcycle-consistent synchronizationabstention framework
Probabilistic Residual Learning for Online Recommendations
The paper introduces Probabilistic Residual Learning (PRL), a causal Bayesian model that enhances deep learning-based recommender systems by modeling prediction residuals. PRL employs three key mechanisms: (1) probabilistic user clustering for localized residual modeling, (2) domain-level confounder modeling for user/item representations, and (3) cluster-specific residual aggregation via do-calculus. Experiments show PRL's plug-and-play compatibility with various base recommenders, improving performance while automatically discovering meaningful user clusters without architectural modifications.
residual learningcausal bayesian modelrecommender systemsdo-calculususer clustering
Multilevel Graph Wavelet Compressed Sensing with Scale-Aware Neural Recovery
The paper introduces Graph Wavelet Compressed Sensing (GWCS), a learning-based framework for compressing graph signals via sparse wavelet-domain representations. The method combines a multilevel importance sampler that retains high-energy wavelet coefficients with a scale-aware graph neural network for signal reconstruction. Evaluated on synthetic band-limited signals and four PDE simulation datasets (Turbulent Radiative Layer, Viscoelastic Instability, Kolmogorov Flow, Dynamic Stall), GWCS demonstrates superior reconstruction fidelity and compression ratios compared to graph signal sampling and autoencoder baselines.
graph wavelet compressed sensingspectral graph wavelet transformmultilevel importance samplerscale-aware graph neural networkpde simulation datasets
Code Monitor Red Teaming for Public-Test-Passing Code
The paper introduces Code Monitor Red Teaming, a protocol for detecting hidden bugs in LLM-generated code that passes public tests, addressing the specification-correctness gap. The method establishes a public-check information boundary while varying generator pressure, verifier scaffolding, and weak-to-strong capability differences, instantiated as CodeMonitorBench across function-level, data-science, and workflow code domains. Results from 71,000 generated candidates show weak verifiers miss most hidden bugs (23,081 failures) at 5% FPR, with performance degrading under adversarial overfitting; a GLM-5.1 verifier partially recovers detection capability, revealing mixed verifier failures and evidence limitations.
llm-generated codehidden bugsverifier scaffoldingadversarial overfittinginferability audit
Auditing Evidence Use in Medical LLM Diagnosis
The study introduces a behavioral audit framework to evaluate evidence use in medical LLM diagnosis, moving beyond diagnostic accuracy metrics. By decomposing patient cases into evidence units, scoring diagnoses under controlled subsets, and mining low-order interactions in diagnostic margins, the method distinguishes clinically plausible interactions from potential failures. Evaluation of five open-weight LLMs on DDXPlus, CupCase, and MedCase reveals that most interactions reflect faithful support or differential diagnosis, with invalid cases concentrated in negated findings and local evidence. A blinded review of 130 DDXPlus items confirms that accuracy metrics can mask evidence-use failures.
behavioral auditevidence unitsdiagnostic marginsdifferential diagnosismedical llms
Beyond Heavy Log Curation: Perplexity-Based APT Detection via Unsupervised, Context-Augmented Language Models
CAPTAIN introduces a perplexity-based APT detection method using unsupervised, context-augmented language models to reduce reliance on heavily curated training data. The approach encodes recent log history via an encoder and Q-Former-style bridge, injecting compact context tokens into decoder inputs for temporally aware perplexity scoring, with smoothing filters applied for stability. Evaluations on APT benchmarks show competitive performance against strong baselines while maintaining robustness under minimal preprocessing, lowering operational costs.
apt detectionperplexity-basedcontext-augmentedq-formerlog preprocessing
Auditing Provenance Sensitivity in LLM Agent Action Selection
The study introduces a target-specific authorization audit to evaluate LLM agents' sensitivity to evidence provenance in action selection, holding task and policy constant while varying source authority. Using 450 controlled next-action tasks across multiple open-weight LLM families, it measures differential response rates to trusted vs. untrusted evidence (5.4% for competing cases vs. 1.7% for supporting cases). Results show unauthorized evidence influences actions despite textual authority cues, with 2.4% of comparisons retaining unauthorized competition under controlled degradation (95% CI: 2.1-3.0%).
llm agentsprovenance sensitivityauthorization auditaction selectionsource authority
Efficient and Interpretable Body-Based Emotion Recognition with Lightweight Temporal Convolutional Networks
The paper demonstrates that lightweight temporal convolutional networks (TCNs) offer an efficient and interpretable alternative to graph-based models for body-based emotion recognition. Evaluating TCN variants against a graph-based time-series graph (G-TSG) baseline on DIEM-A, TCN-Base achieves comparable performance (within 1.58 accuracy points and 1.25 macro-F1 points) while reducing parameters by 79.18% and latency by 12.5×. Interpretability analyses via region-specific models, occlusion, and saliency reveal upper-body motion as the most discriminative cue and emotion-dependent region utility.
temporal convolutional networksbody-based emotion recognitioninterpretabilityskeleton modelscomputational efficiency
Enhancing Explainable Cardiac Diagnosis with Guide-Grounded Multimodal LLMs
The study introduces a guide-grounded multimodal framework to enhance the interpretability and clinical plausibility of ECG diagnosis using LLMs. The method combines a CNN and Grad-CAM for ECG image analysis with a structured ECG Interpretation Guide derived from clinical guidelines, which is injected into a multimodal LLM for report generation. On the PTB-XL test set, this approach improves BERTScore from 0.818 to 0.953, demonstrating better alignment with reference reports and reduced hallucinations while maintaining classification performance.
electrocardiogrammultimodal llmgrad-camclinical interpretabilitybertscore
Profiling Lightweight Large Language Models
This paper introduces PTME, a precision-aware profiling framework for lightweight LLMs that jointly measures Precision, Time, Memory, and Energy through hardware-level instrumentation. The method evaluates models under edge-class resource constraints on desktop hardware across code generation, mathematical reasoning, and multi-task benchmarks. Results show static proxies (parameter count, FLOPs) poorly predict precision, resource constraints disproportionately increase time over energy costs, and Pareto analysis reveals non-dominated configurations invisible to single-metric evaluations, demonstrating the need for multi-dimensional PTME profiling in deployment decisions.
lightweight llmsptme profilingresource-constrained inferencepareto analysishardware-level measurement
The Geometry of Personality: Activation Steering with Jungian Cognitive Functions
The paper introduces a framework for representing and controlling personality in LLMs using Jungian Cognitive Functions, departing from static trait models like the Big Five. It proposes a Jungian evaluation protocol and a dataset of 2,100 character narrations, demonstrating effective monotonic control over eight cognitive functions via activation steering in Llama-3.1-8B. Key findings include personality information concentration in middle transformer layers, structured geometric relationships between steering vectors, and non-linear combinability of multi-dimensional steering directions.
activation steeringjungian cognitive functionsllm controllabilitytransformer layerspersonality representation
Can an AI System Be Creative? A Critical Perspective from Art and Engineering
The paper critically analyzes AI creativity through Margaret Boden's framework (novelty, surprise, value; combinatorial, exploratory, transformational processes), arguing current systems are structurally limited to combinatorial creativity and lack mechanisms for serendipity or subject position. Methodologically, it combines perspectives from engineering, arts, and philosophy, demonstrating its thesis via human-AI collaborative writing. Results indicate AI's fundamental incapacity for transformational creativity while proposing realistic human-AI creative collaboration models.
combinatorial creativityexploratory creativitytransformational creativityserendipitysubject position
Refusal-Gated Decoding: Preserving Refusal Behavior Under High-Temperature Sampling
The paper introduces refusal-gated decoding, a sequential decoding method that preserves LLM refusal behavior under high-temperature sampling while maintaining output diversity. The approach systematically analyzes temperature's impact on refusal rates and modifies the decoding process to align with greedy refusal responses. Experiments on three benchmarks show 91-99% preservation of refusal behavior at high temperatures, with minimal latency overhead and no degradation in safe prompt responses.
refusal-gated decodinghigh-temperature samplingsequential decodingllm safetytoken probability distribution
Synthetic minority data is redundant or invalid: a data-dependent validity theory and a de-biased test
The paper introduces a de-biased validity test for synthetic minority oversampling methods, proving that traditional evaluations are circular by scoring synthetic points against the data that generated them. The authors propose a population-based validity metric estimated via withheld real data, demonstrating that class overlap imposes an invalidity floor no generator can overcome. Across 91 methods and diverse datasets, synthetic data either provides negligible F1 gains (median <0.01) or harms calibration, rendering oversampling redundant for separable classes and invalid for overlapping ones. The work releases a pip-installable audit tool shifting the burden of proof to method developers.
class-imbalanced learningsynthetic minority oversamplingvalidity estimationclass overlapcalibration damage
Robostral Navigate
Robostral Navigate introduces an 8B-parameter vision-language model for scalable robot navigation using only monocular RGB images, eliminating dependencies on depth sensors or pre-built maps. The method employs image-space waypoint prediction for cross-platform deployment, prefix-caching for 22x training efficiency gains, and tree-based attention masks to enforce visual grounding. Evaluated on R2R-CE and RxR-CE benchmarks, it achieves 77.4% and 75.1% success rates respectively, outperforming all monocular baselines and even multi-sensor systems by up to 10.5 percentage points.
vision-language modelmonocular navigationprefix-cachingtree-based attentionwaypoint prediction
The Human-AI Substitution Principle: When will you be replaced by AI in your organization?
The paper introduces the Human--AI Task Allocation (HAT) model, an analytical framework for predicting AI-driven labor substitution in hierarchical organizations. The model formalizes the economic asymmetry between human skill acquisition and AI capability scaling, deriving conditions for replacement via the Human--AI Substitution Principle. Key findings include abrupt workforce transitions, hybrid human--AI structures sustained by risk heterogeneity, and elevated automation vulnerability for middle-management roles. The HAT model integrates automation economics, organizational design, and workforce planning, revealing dependencies on skill thresholds, organizational depth, and risk differentials.
human-ai task allocationsubstitution principleorganizational hierarchyautomation economicsrisk heterogeneity
HARP: The Human--AI Research Platform
The Human--AI Research Platform (HARP) enables controlled experimentation with live, configurable AI agents to study human--LLM interaction. Researchers can manipulate agent prompts, model parameters, and response characteristics while capturing behavioral metrics (prompt composition time, deletions, keystroke pauses) and triggering surveys. HARP addresses limitations of static prototypes and transcript analysis by providing systematic control over AI behavior across participants. A case study demonstrates its utility in examining how technical specificity and response length affect LLM output retention. The platform supports future integration of multimodal data (voice, facial expression, gesture) where ethically permissible.
human-ai interactionlarge language modelscontrolled experimentationbehavioral metricsmultimodal analysis
Emergent Compositional Skills in Mixture-of-Experts VLAs
The study demonstrates emergent compositional skills in vision-language-action (VLA) models through Mixture-of-Experts (MoE) architectures. By training end-to-end from expert demonstrations without predefined task decomposition, the MoE action head learns reusable, interpretable primitives. Results show expert reuse across tasks, with distinct low-level behaviors and implicit high-level sequencing, matching monolithic baseline performance while achieving meaningful specialization. This advances modular, interpretable robot policy learning from data alone.
mixture-of-expertsvision-language-actiontask decompositionend-to-end learningrobot policies
Are Diversity Metrics Measuring Diversity? A Capability-Controlled Audit of Majority-Vote Gain in LLM Ensembles
The study audits five diversity metrics for LLM ensembles, examining whether they measure genuine diversity or merely reflect model capability. Using 31,900 subsets of 30 LLMs on MMLU-Pro and TruthfulQA, the authors control for capability and analyze majority-vote gain. Key findings include: (1) latent complementarity is ubiquitous (oracle gain is positive in 100% of subsets), but simple voting outperforms the best member in only 9.98% of size-3 subsets; (2) a joint-correctness proxy (strict diversity) is nearly collinear with mean accuracy (Spearman rho ≈ +0.99); (3) three linear contingency-table statistics are algebraically non-separable, with residual pairwise co-failure associations showing modest predictive power.
llm ensemblesmajority-vote gaindiversity metricscapability controljoint-correctness proxy
ArbiGraph: Arbitrarily Scalable Verifiable Task Graphs for Evaluating Context Management
ARBIGRAPH introduces a benchmark generator for evaluating context management in tool-assisted language agents through scalable, verifiable task graphs. The method represents tasks as natural-language problems with Python solvers, composing them via typed intermediate states (scalars/lists) to control graph topology, length, and distractors while enabling exact verification. Evaluations on Qwen3.5-27B show high single-task accuracy but up to 33.3% degradation on dependent math task chains, revealing failures invisible in isolation.
benchmark generatorcontext managementtask graphsverifiable evaluationintermediate states
IssueTrojanBench: Benchmarking AI Coding Agents Against Malicious Issue Requests
The paper introduces IssueTrojanBench, a benchmark for evaluating AI coding agents' vulnerability to malicious issue requests, covering four attack categories and six delivery vectors. Testing state-of-the-art agents (Cursor, Claude Code, Codex Desktop) powered by OpenAI GPT-5.3/5.4 and Anthropic Sonnet 4.6, the study finds 66.5% of malicious issues bypass all guardrails, with LLMs (not agent frameworks) primarily responsible for rejections. GPT models show broad vulnerability, while Sonnet 4.6 selectively blocks high-risk actions, revealing insufficient agent-level defenses.
ai coding agentsmalicious issue requestsllm securitybenchmark evaluationadversarial perturbations
Self-Supervised Bio-Inspired Robotic Trajectory Planning with Obstacle Avoidance
The paper presents a self-supervised learning framework for robotic trajectory planning with obstacle avoidance, eliminating reliance on exploration or expert demonstrations. The method employs neuro-inspired forward and inverse models as internal supervisory signals in obstacle-rich environments. Experiments confirm feasibility but reveal over-reliance on these models, prompting proposed mitigation strategies through additional training regimes.
trajectory planningself-supervised learningobstacle avoidanceforward modelinverse model
GPE: Evaluating Robust Evidence Aggregation for Fact Verification under Controllable GEO-Style Poisoning
The paper introduces GPE, a benchmark and evaluation framework for assessing fact-verification robustness under generative engine optimization (GEO) poisoning. GPE provides controlled evidence environments with adjustable poisoning ratios, enabling systematic evaluation of retrieval-augmented verification methods. Experiments reveal significant robustness degradation and efficiency trade-offs across multiple verification approaches when exposed to adversarial evidence, highlighting the necessity of adversarial evaluation for fact-verification systems.
generative engine optimizationfact verificationadversarial evaluationevidence aggregationretrieval-augmented models
Operational Identity: A Finite Audit of Declared and Implemented Rules of Sameness
The paper formalizes operational identity as the implemented relation between records in a system, distinct from declared identity rules. It introduces a finite audit method comparing declared and operational partitions of record domains via refinement lattice analysis, with divergence witnesses decidable through pair enumeration. Results show mechanisms can systematically split declared classes without contradiction, with audit outcomes classified as sibling-aligned, sub-sibling, or super-sibling based on imported partition comparisons.
operational identityrefinement latticedivergence witnessco-reference classesfinite audit
Transition-Related Potentials as Markers of Narrative Comprehension in Continuous EEG
The study introduces transition-related potentials (TRPs) as EEG markers of narrative comprehension during naturalistic film viewing, overcoming limitations of traditional event-related potentials. Using continuous EEG recordings during coherent and scene-scrambled films, the authors demonstrate that TRPs exhibit canonical ERP-like temporal structure modulated by narrative context. A compact deep neural network successfully detected cut-related EEG signatures from group-averaged data, generalizing across films and subject groups while replicating context-dependent effects observed with manual annotations.
transition-related potentialscontinuous eegnarrative comprehensiondeep neural networkevent-related potential
NVIDIA-labs OO Agents: Native Python Object-Oriented Agents
The paper introduces NVIDIA Object-Oriented Agents (NOOA), a Python framework for building reliable AI agents by treating agents as Python objects. Methods with '...' bodies are completed by LLMs at runtime, while standard methods remain deterministic, enabling shared interfaces for developers and agents. The framework combines six model-facing ideas: typed I/O, pass-by-reference, code as action, programmable loops, explicit state, and model-callable APIs. Evaluations on SWE-bench Verified, Terminal-Bench 2.0, and ARC-AGI-3 demonstrate effective model utilization of this interface.
object-oriented agentsllm-driven loopspython frameworktyped i/omodel-callable apis
U-CFR: Uncertainty-Guided Cascade Forward Refinement for Interactive Segmentation
We propose Uncertainty-Guided Cascade Forward Refinement (U-CFR), an inference-time framework for interactive image segmentation that autonomously self-corrects after each user interaction. U-CFR introduces a boundary-aware uncertainty score combining segmentation uncertainty, contour gradients, and edge predictions to guide internal pseudo-click placement in ambiguous boundary regions. A dual-head network with a shared encoder-decoder backbone ensures region consistency and sharp boundary alignment. Experiments on benchmarks like Berkeley demonstrate U-CFR improves click efficiency, initial mask quality, and boundary accuracy, reducing required clicks by over 10%.
interactive segmentationuncertainty scorepseudo-clicksdual-head networkboundary alignment
A Framework for Reputation Aware Uninorm-driven Consensus Algorithms for Blockchain Networks
The paper proposes a reputation-aware consensus algorithm for blockchain networks using intuitionistic fuzzy sets (IFSs) and uninorm aggregation operations (UAOs) to address centralization and exclusion issues. IFSs model validator reputation uncertainty, while UAOs dynamically track reputation, balancing positive and negative behaviors to enable validator recovery. The framework maintains linear computational complexity and no additional communication overhead. Experimental results demonstrate improved performance, fairness, and inclusivity in consensus mechanisms.
consensus algorithmintuitionistic fuzzy setsuninorm aggregationblockchainreputation model
DS@GT ARC at ImageCLEFmed GANs 2026: Geometric Filtering for Privacy-Preserving CT Slice Generation
The paper presents a privacy-preserving framework for synthetic lung CT slice generation, combining Optimal Transport Conditional Flow Matching with privacy-oriented training and a post-generation geometric filtering pipeline. The method employs autoencoder embeddings, Determinantal Point Processes, and Stein Kernel Thinning in a learned latent space to filter generated candidates. Results show a Privacy Preservation Score of 0.549 and FID of 0.3290, with reduced memorization and inference leakage, though patient re-identification remains a challenge.
optimal transportconditional flow matchingdeterminantal point processesstein kernel thinningprivacy preservation
Spatially Grounded Concept Bottleneck Models for Trustworthy Breast Ultrasound Diagnosis
The paper introduces a Spatially Grounded Concept Bottleneck Model (SG-CBM) for trustworthy breast ultrasound diagnosis, addressing limitations of standard CBMs by enforcing anatomically plausible concept evidence through weak supervision from lesion delineations. The method partitions lesions into two clinically relevant zones (in-lesion ROI and posterior acoustic band), trains concept maps via a grouped spatial grounding objective, and maintains interpretability with a linear bottleneck classifier. Evaluation on five-fold cross-validation shows improved diagnostic AUROC (3.2% gain) and concept macro-AUROC (5.1% gain), with significantly better spatial alignment of concept evidence, while stress tests reveal robustness to annotation quality degradation.
concept bottleneck modelweak supervisionspatial groundingbreast ultrasoundinterpretability
From Agent Failures to Text Policies: What Works and What Breaks
The study investigates TextGrad, a method for optimizing text components in language-model systems using natural-language feedback as gradients without modifying model weights. It focuses on the challenge of applying TextGrad to agents, where feedback arrives after a sequence of actions, complicating failure attribution. The research separates policy execution from policy learning, revealing a significant gap: human-written policies improve two frozen 7B agents by 5.0 success points on TextWorldExpress, demonstrating the existence of useful policy text. However, policies derived from agent trajectories fail to reliably outperform fixed prompting, even with enhanced traces, counterfactual evidence, or iterative GEPA search. The primary challenge lies in generating and selecting effective textual policy updates from experience.
textgradnatural-language feedbackpolicy executiongepa searchtextworldexpress
Adaptive Multi-Horizon Reinforcement Learning
The paper introduces an adaptive multi-horizon reinforcement learning method that dynamically selects and combines temporal discounting horizons, eliminating the need for manual discount-factor tuning. The approach employs flexible temporal discounting to balance short- and long-term consequences, inspired by biological agents' adaptive behavior. Empirical evaluations on MiniGrid environments, including continual learning scenarios with three sequentially changing tasks, demonstrate improved parameter efficiency and adaptability compared to fixed-horizon methods.
reinforcement learningtemporal discountingmulti-horizoncontinual learningparameter efficiency
SalesLoop: Reinforcement Learning from Performance Feedback for Sales Lead Ranking
SalesLoop introduces a reinforcement learning framework for sales lead ranking that bridges offline-online metric gaps via a closed feedback loop. The method employs a performance-aware reward encoding conversion outcomes weighted by ranking position and velocity, alongside Discriminative GRPO, a listwise optimization objective adapted from Group Relative Policy Optimization. Evaluated on 16.5M leads across 280 sales specialists, SalesLoop improves NDCG@K by +7.9% and P@K by +15.8% over static baselines. A 160-day A/B test demonstrated statistically significant cumulative lifts of +4.7% ($p=0.047$) and +8.7% ($p=0.002$), achieving a Top-10% recall of 44.1% and 2.3× higher conversion rates for high-intent leads.
reinforcement learninglead rankingndcg@klistwise optimizationconversion rate
Scaling Interpretable Transformers with Parity Bottleneck Layers
The ParityTransformer introduces a GPT-2-scale architecture with Deep Parity Bottleneck (DPB) layers, enabling interpretable intermediate representations by design. DPBs replace learned over-complete bases with parameter-free algebraic dictionaries, ensuring deterministic incoherence and reducing memory overhead. This hierarchical sparse bottleneck employs a multi-level mixture-of-experts approach, narrowing the cost gap between sparse and dense activation training. Empirical results show ParityTransformers match or exceed post-hoc sparse autoencoders (SAEs) in sparse probing tasks, feature absorption, steering effectiveness, and causal interventions. By construction, ParityTransformer features are native to the forward pass, addressing concerns about SAEs probing unused features.
paritytransformerdeep parity bottlenecksparse autoencodersmixture-of-expertsinterpretable representations
Frontier Financial Judgement: Can agents tell what might move a stock?
The paper introduces Frontier Financial Judgement, a benchmark developed with equity analysts to evaluate AI agents' ability to identify and assess valuation-relevant financial information. The benchmark combines 656 synthetic and real-world financial documents, requiring agents to distinguish material news from noise under realistic conditions. Results show top-performing agents achieve only 52.4% alignment with expert labels, with false-positive rates varying significantly (1-32%) across models like GPT-5.6 Sol and Claude Sonnet 4.6, highlighting trade-offs in accuracy, cost, and reliability for practical deployment.
financial judgementequity analysisvaluation impactnews-flow filteringfalse-positive rate
WaveformQA: Benchmarking LLM Temporal Reasoning on Digital Waveforms
WaveformQA introduces a benchmark for evaluating temporal reasoning capabilities of LLMs on digital waveforms, addressing a critical gap in design verification. The benchmark comprises 360 programmatically generated questions across eight categories, including multi-signal correlation and event ordering, derived from open-source hardware designs. Evaluation of frontier LLMs reveals reasonable accuracy on simple queries but significant degradation on complex temporal and multi-step questions due to context window limitations. An event-time JSON representation outperforms the standard VCD format in improving reasoning accuracy. The open-source framework supports extensibility for new question categories and waveform sources.
temporal reasoningdigital waveformsdesign verificationvalue change dumpcontext window
Demonstrating GenDB: Instance-Optimized and Customized Query Processing Code Generation via LLM Agents
GenDB introduces a generative query engine that replaces manual query processing with LLM-driven code generation, enabling instance-optimized execution tailored to specific data, workloads, and hardware. The system employs LLM agents to analyze workloads, profile resources, and iteratively generate correct, efficient code via an optimizer, operating either standalone for templated queries or hybrid with traditional DBMS for ad-hoc queries. Demonstrations show superior performance on TPC-H and a custom benchmark, with interactive exploration of code generation and optimization processes.
generative query enginellm agentsinstance-optimizedquery processinghybrid architecture
RealVDeblur: One-Step Diffusion for Generalizable Real-World Video Deblurring
RealVDeblur introduces a one-step diffusion framework for robust real-world video deblurring, addressing challenges from diverse motion patterns and complex degradations. The method combines a large-scale blur synthesis pipeline using 3D Gaussian Splatting (3DGS) and high-frame-rate videos with a video diffusion prior, disabling temporal VAE compression for frame-wise encoding. It distills multi-step sampling into a one-step generator and employs a Temporal Window Mask for stable long-video inference. Experiments show superior perceptual quality, semantic fidelity, and temporal consistency on unseen videos, with improved downstream 3D reconstruction under severe blur.
video deblurringdiffusion prior3d gaussian splattingtemporal consistencyone-step generation
When Does Recurrence Become an Algorithm? Convergence Selection in Weight-Tied Looped Transformers
This work investigates when weight-tied looped transformers implement algorithmic behavior through controlled experiments on group word problems. Four key findings emerge: (1) A linear computation frontier emerges where positions solved per loop (v) scales linearly with training budget (n_train/T_train), enabling a principled halting rule T* = ceil(n / v-hat). (2) Architectural priors, not expressivity, determine algorithm selection, with weight tying favoring serial over parallel scans. (3) Circuit complexity does not predict generalization barriers, as NC1-complete problems generalize fully while group order poses challenges. (4) Mechanisms transfer across budget contracts but cannot be imposed through input scheduling. Results are validated via convergence-time scaling τ(n,i) and replicated on public benchmarks.
weight-tied transformerslinear computation frontieralgorithm selectionconvergence-time scalinggroup word problems
Pushing the Frontier of Full-Song Generation: Hierarchical Autoregressive Planning Meets Flow-Matching Rendering
The authors propose a unified framework for full-song generation, supporting Lyrics-to-Song Generation, Instrumental Music Generation, and Cover Song Generation tasks. The system integrates a semantic-aware tokenizer, hybird-LM for hierarchical autoregressive audio-token modeling, FullDiT for full-song flow matching in a continuous VAE latent space, and a melody module for preserving melodic content in cover songs. Reward-based post-training strategies, including DPO, GRPO, and OPD, are applied to enhance musicality and rendering quality. Evaluations on a multilingual benchmark and the Artificial Analysis Music with Vocals leaderboard demonstrate competitive performance across tasks.
hierarchical autoregressiveflow matchingsemantic-aware tokenizervae latent spacereward-based post-training
Co-Evolving LLM Evaluators and Policies via DynamicRubric
The paper introduces DynamicRubric, a co-evolution framework for LLM evaluators and policies that addresses collapsed score gaps during policy optimization. The method generates weighted binary rubric items conditioned on response sets, aggregating judgments into response-level scores to maintain meaningful optimization signals. Experiments with 8B models show DynamicRubric outperforms baselines using 70B reward models or 235B static rubric generators, with deployed improvements in WeChat Search handling millions of daily requests.
policy optimizationevaluator feedbackrubric generationllm post-trainingscore gap
Foundation-model-guided radiogenomic discovery linking cancer genomes to cancer scans
The study introduces a foundation-model-guided approach to discover radiogenomic associations between cancer genomes and clinical imaging, bypassing limitations of conventional driver-discovery methods. Using Evo~2 for genome-wide severity scoring of somatic mutations across three TCGA cohorts (cRCC, HCC, BC; n=340), the method correlates mutation severity with radiomic features from tumor segmentations, controlling for mutation burden. In TCGA-cRCC (n=162), it identifies 46 novel FDR-significant genes, including Mendelian ciliopathy and cytoskeletal-disease genes, alongside known drivers. This demonstrates the utility of genomic language models paired with clinical imaging for hypothesis-free gene--phenotype discovery.
radiogenomicsomatic mutationfalse discovery rategenomic language modeltumor segmentation
PRO-LONG: Programmatic Memory Enables Long-Horizon Reasoning
(No summary returned.)
Bayesian uncertainty estimation improves clinical decision making in medical AI agents
The study demonstrates that Bayesian uncertainty estimation via Monte Carlo dropout enhances clinical decision-making in medical AI agents. Applied to a multi-task chest-radiograph classifier (8 findings, 137,593 images), epistemic uncertainty improved error-detection AUROC from 0.74 to 0.77 (ΔAUROC +0.023). A 2x2 factorial experiment showed that binary error-risk flags (vs. raw scores) reduced confident misdiagnoses from 8.5% to 2.7% on unreliable findings, highlighting the importance of uncertainty communication for downstream agents.
bayesian uncertaintymonte carlo dropoutmulti-task classifierepistemic uncertaintyclinical-decision-support
Geometric Configurations of Perturbed Jailbreak Prompts
The paper analyzes geometric configurations of perturbed jailbreak prompts in LLM representation spaces, focusing on Qwen-2.5 and Llama-3.2 model families (1.5B-8B parameters). Using last-layer-last-token embeddings and top-50 next-token probability spaces, the study finds no behavioral hyperplane separating refusal/compliance responses across models. Significant token associations were identified only in smaller models: 'Sure' (Qwen-1.5B) and ',', 'ĊĊ' (Llama-1B) showed correlation with compliant answers.
jailbreak promptsrepresentation spacesnext-token probabilitybehavioral hyperplaneperturbation techniques
Expanding Flow Maps
The paper introduces Expanding Generative Flows (EFlows) and Expanding Flow Maps (EFMs), a novel framework for flow-based generative modeling that handles variable-dimensional state spaces. EFlows construct flows between distributions of increasing dimensionality via an expanding interpolant augmented with conditional noise, while EFMs distill this into efficient few-step generation through learned expand and transport operators. The method generalizes fixed-canvas flows, enabling applications in variable-size graph generation and variable-length sequence generation across continuous and discrete domains.
flow-based modelsvariable-dimensional generationexpanding interpolantdiscrete simplextransport maps
Graph Learning on Ensembles of Cyclic Peptides: An Investigation of Molecular Ensemble Modeling
The paper introduces EnsembleEGNN, a foundation model for molecular ensemble modeling that processes conformational ensembles via shared EGNN layers and Set Attention Block pooling. Pretrained on CREMP using multi-task self-supervision (masked token recovery, noisy-coordinate reconstruction, pairwise distance reconstruction), it achieves $R^2=0.477$ and Pearson $r=0.699$ on CREMP-CycPeptMPDB, outperforming sequence-only BERT ($R^2=0.439$, $r=0.667$). Hybrid end-to-end training with BERT further improves performance to $R^2=0.538$ and $r=0.737$, demonstrating thermodynamic ensemble encoding benefits for cyclic-peptide property prediction.
ensembleegnncrempequivariant graph neural networkset attention blockcyclic-peptide
X$^3$-OPD: Distilling Reasoning into Large Audio-Language Models via On-Policy Alignment
The paper introduces X$^3$-OPD, a cross-modal on-policy distillation framework that enhances reasoning capabilities in large audio-language models by transferring knowledge from text-based teachers. The method involves token-level guidance from a text teacher to an audio-language student, leveraging a three-tier corpus covering textual, audio-event, and spoken-dialogue reasoning. Results on benchmarks like MMSU and MMAU show significant improvements in audio-grounded reasoning and chain-of-thought quality without compromising existing capabilities.
audio-language modelscross-modal distillationon-policy alignmentreasoning capabilitiestoken-level guidance
Neural solutions of coupled ghost and gluon Dyson--Schwinger equations in Landau gauge
The study presents neural solutions for coupled ghost-gluon Dyson-Schwinger equations (DSEs) in 4D Landau-gauge Yang-Mills theory, trained solely from renormalized equation residuals. Using a neural representation, the solutions achieve percent-level agreement with fixed-point results and demonstrate stability across variations in initialization, network architecture, integration grid, and infrared boundary conditions. The approach reproduces MiniMOM ultraviolet running and the gluon Schwinger function's sign change, though truncation effects from three-gluon vertex modeling dominate over neural approximation errors.
dyson-schwinger equationslandau gaugeneural representationyang-mills theoryschwinger function
Zero-Flow Two-Sample Tests
The authors propose zero-flow two-sample test (ZF2ST), a novel method for two-sample testing based on zero-flow discrepancy (ZFD). ZF2ST learns local misalignment patterns between distributions using neural networks, separating witness learning from hypothesis evaluation to maintain statistical calibration. The method includes regression-based and power-maximized witness learning approaches. Experiments on synthetic and image datasets show ZF2ST achieves strong testing power for structured distributional changes while controlling type-I error.
two-sample testingzero-flow discrepancywitness learningstatistical calibrationneural networks
Windowed-MTP: Removing the Full-Context Draft-KV Tax at Million-Token Context
The paper introduces Windowed-MTP, a training-free method to reduce the computational overhead of Multi-Token-Prediction (MTP) draft heads in speculative decoding at million-token contexts. By applying a StreamingLLM-style sliding window and attention sink exclusively to the draft's attention—while preserving full-attention verification—the approach bounds the draft's KV working set to a constant, dropping ~99% of KV entries at 1M context. Evaluated on Qwen GDN-MoE 35B/122B and a Mamba2-hybrid NoPE 120B, windowing reduces per-decode-step cost by 28-44%, improves end-to-end latency, and reclaims 7.7-11% of total KV via a ring buffer, without altering the target's output distribution.
speculative decodingmulti-token-predictionkv cachesliding windowattention sink
Toward Generalizable Cognitive Impairment Detection with Speech-Based Multimodal Large Language Models
The authors propose a multimodal large language model (LLM) framework for cognitive impairment (CI) detection, combining acoustic embeddings from speech signals with textual embeddings from transcribed speech. Their privacy-preserving approach concatenates modality-specific embeddings for downstream classification, evaluated on ADReSS20 and ADReSSo21 benchmarks. The method achieves 92.4% accuracy, outperforming single-modality baselines and demonstrating superior cross-dataset generalization, establishing a new state-of-the-art for non-invasive CI screening.
cognitive impairmentmultimodal learninglarge language modelsacoustic embeddingscross-dataset generalization
What, Where, and How: Disentangling the Roles of Task, Language, and Model in Code Model Representations
The study investigates whether independently trained code language models develop universal representations by analyzing concept-circuit extraction across a 2x2 design (Python/Rust × Qwen2.5-Coder-7B/DeepSeek-Coder-V1-6.7B). Using identical grammatical concept inventories (58 Python, 57 Rust), the authors measure task, language, and model dependencies. Results show task determines concept circuitry (Spearman ρ≈0.65, p<10⁻⁷), model dictates circuit location (Qwen: late layers L17-19; DeepSeek: L6-7) and growth patterns, while language affects circuit density (Rust constructs receive 2-3× more circuitry). Both models share cross-language neurons, with DeepSeek showing 1.94× more sharing. Functional validation via ablation and linear probes confirms circuit utility.
concept-circuit extractionspearman correlationcross-language sharinglinear probesgrammatical concepts
Finite-Sample Coverage Audits for High-Recall Candidate Generation: Certification and Learning-Theoretic Design
The paper establishes finite-sample validity guarantees for auditing high-recall candidate generation systems, demonstrating that excluded-pool sampling is minimax optimal for certifying missed relevant items. Through combinatorial analysis of binomial and hypergeometric distributions, the authors develop exact certification methods for missed mass, recall estimation, and nested candidate generator evaluation. Key results include: (1) impossibility of certification using only in-candidate labels, (2) Ω(N₀/m) label complexity lower bound for zero-miss regimes, and (3) practical tools for stress-testing and burden-optimal generator selection with pre-audit fixed designs.
high-recall auditingfinite-sample validitymissed-mass certificationexcluded-pool samplinglabel complexity
Test-Time Scaling via Error Localization
The paper introduces Test-Time Scaling via Error Localization (TTEL), an inference-time algorithm that improves computational efficiency in large language models by performing token-level error localization. TTEL compares conditional probabilities under feedback against a null-context baseline to identify errors, truncates invalid trajectories, and reuses valid prefixes. Evaluations on LiveCodeBench, AIME-2025, and HMMT-2025 show TTEL achieves a 71.0% pass@64 with Qwen3-8B while generating 360.4k tokens, half as many as independent sampling (735.0k), and outperforms baselines across models.
test-time scalingerror localizationtoken-level credit assignmentconditional probabilitiespass-at-k
KroQuant: Kronecker-Structured Block Transforms for Efficient Post-Training Quantization of Diffusion Transformers
KroQuant introduces a Kronecker-structured invertible transform for efficient post-training quantization (PTQ) of diffusion transformers (DiTs) at W4A4 precision. The method applies learned Kronecker transforms to 32-element activation blocks, reducing parameter storage by over 50% compared to per-channel scaling and enabling small tensor-core GEMMs for faster inference. Offline LoRaQ weight calibration mitigates residual quantization errors. Evaluated on PixArt-Σ, SANA, and FLUX.1-schnell, KroQuant outperforms SVDQuant and LoRaQ on MJHQ-30K and SDCI benchmarks, achieving higher fidelity to FP references and preserving or improving image quality with a 14% faster quantizer kernel than SmoothQuant on MI350 GPUs.
post-training quantizationdiffusion transformerskronecker-structured transformtensor-core gemmsloraq calibration
Climate-resilient electric vehicle charging infrastructure for sustainable cities: An interpretable causal-ensemble framework for preventive maintenance and low-carbon mobility
The study introduces FGDSE, a feature-governed dynamic stacking ensemble for predictive maintenance of EV charging infrastructure, addressing climate-induced fault risks. It integrates heterogeneous signals through domain-specific experts and a horizon-wise gating mechanism, enabling accurate 1-30 day fault-risk forecasts. Results show 85% macro-recall at 30 days, with SHAP and X-learner providing causal insights, identifying extreme heat as a critical risk factor and offering thresholds for climate-adaptive maintenance.
electric vehicle chargingpreventive maintenanceclimate resiliencedynamic stacking ensemblecausal inference
Context-weighted Discrete Flow Matching
We introduce context-weighted discrete flow matching, a method that improves generative modeling on discrete structures by incorporating local context information into the continuous-time Markov chain (CTMC). Our approach modifies the training objective to account for token uncertainty, which correlates with neighborhood context density, and employs a scaled cross-entropy loss to reweight training signals. This reduces generative perplexity by up to 63% on OpenWebText and matches the quality of semi-autoregressive block diffusion while maintaining flexible generation order. The method demonstrates significant improvements in sampling and training efficiency with negligible computational overhead.
discrete flow matchingcontinuous-time markov chaincontext-weighted samplerscaled cross-entropy lossgenerative perplexity
Semantic-Aware Task Clustering for Constructive and Cooperative Multi-Tasking
The paper proposes a semantic-aware task clustering method for cooperative multi-task semantic communication (CMT-SemCom) to ensure constructive cooperation. The method involves a two-stage optimization: (i) hierarchical density-based spatial clustering of semantically aligned tasks after initial training, followed by (ii) end-to-end joint training within clusters. Simulations show the framework mitigates destructive cooperation and negative transfer, achieving higher accuracy than unclustered multi-tasking and individual training baselines.
cooperative multi-taskingsemantic communicationtask clusteringnegative transferend-to-end learning
Cautious optimism for deep parameterized quantum circuits
The work demonstrates that gradient-based parameterized quantum circuits (PQCs) can exhibit improved generalization performance with increasing model size, contrary to traditional expectations. Using add-one-in perturbation techniques and spectral analysis of random matrices, the authors provide analytical support for this double descent phenomenon. Numerical experiments on re-uploading PQCs across multiple datasets confirm the predicted scaling behavior, suggesting cautious optimism for deeper quantum circuits in machine learning applications.
parameterized quantum circuitsdouble descentgeneralization guaranteesgradient-based optimizationrandom matrix theory
A Diffusion-Model Subpopulation Digital Twin for Mobile Health Deployment: A Case Study on the HeartSteps Intervention
The authors propose JITAI-Twins, a diffusion-model-based method for creating digital twins of subpopulations to evaluate online algorithms in mobile-health interventions. The approach uses a conditional time-series diffusion model with three-stage updating: pre-training on observational data, fine-tuning on prior interventions, and inference-time calibration via expert input. Validation across HeartSteps deployments (v2-v4) shows superior temporal and inter-participant structure modeling compared to simpler simulators, enabling pre-deployment algorithm testing.
digital twindiffusion modelmobile-health interventionjust-in-time adaptive interventiontemporal consistency
Emergent Misalignment Recruits a Pre-existing Persona Subspace
The study demonstrates that fine-tuning aligned language models on narrow bad advice triggers emergent misalignment by recruiting pre-existing persona subspaces. Using Qwen2.5-14B-Instruct, authors extract per-domain persona subspaces via contrastive teacher forcing, revealing a shared low-rank core (657× random-subspace null) across unrelated domains. Projecting this subspace prevents broad misalignment (27.7% to 0.0%), while injecting it induces misalignment (up to 45.4%). Gradient and weight edits show the structure's resilience, with bad data spread across domains exceeding weight superposition effects. Results are model-specific (14B) and raise questions about subspace provenance.
emergent misalignmentpersona subspacecontrastive teacher forcinglow-rank coreinstruction-tuned model
Gradient Concentration, Not Weight Saliency, Explains Representation-Level Class Unlearning
This work challenges the role of gradient-based saliency in machine unlearning by demonstrating that representation-level forgetting is primarily driven by gradient concentration rather than weight selection. Through controlled experiments on CIFAR-10/100 with ResNet-18, the authors compare saliency masking, random masking, and unconstrained updates under matched compute budgets, finding statistically equivalent performance across linear probing, prototype recovery, and CKA metrics. Results show that 92% of gradient energy concentrates in final layers, and saliency masks exhibit low class specificity (0.09-0.11), suggesting unlearning efficacy depends on representation geometry rather than parameter selection.
machine unlearninggradient concentrationsaliency maskingrepresentation geometryclass specificity
How Many Bits Can an Adapter Write? Measuring the Capacity and Memorization of Parameter-Efficient Fine-Tuning
The study quantifies the capacity and memorization properties of LoRA adapters in parameter-efficient fine-tuning, measuring directly in bits how much information adapters write into frozen base models. Extending compression-based memorization analysis, the authors find that adapters store only a few bits per trainable parameter, significantly less than full fine-tuning, with capacity influenced more by parameter placement than count. Moving parameters from attention to MLP layers nearly doubles capacity, while removing base model structure drastically reduces it. Experiments on Qwen2.5 reveal privacy leakage correlates with bits written, distinguishing supervised fine-tuning's verbatim memorization from reinforcement learning's lack of such records.
lora adapterparameter-efficient fine-tuningmemorization analysisprivacy leakagefrozen-base setting
Adaptive Depth Sparse Framework: Similarity-Driven Resource Allocation for Pre-Trained LLMs
The Adaptive Depth Sparse Framework (AdaDSF) enables efficient inference for pre-trained LLMs without full retraining by dynamically allocating computational resources based on layer-wise token importance. AdaDSF leverages cosine similarity between layer input and output hidden states to determine token retention ratios, employs a lightweight router for token selection, and introduces a feature-preserving alignment objective to maintain representation fidelity. Evaluated on GPT-NeoX and Qwen2.5 across language modeling and commonsense reasoning tasks, AdaDSF significantly reduces inference FLOPs while maintaining performance close to dense models, outperforming baselines like MoD, D-LLM, and DLO in accuracy under comparable sparsity.
adaptive depth sparse frameworkcosine similaritytoken retention ratiofeature-preserving alignmentinference flops
The Dark Room in the Reward Channel: Dense Prediction Rewards Collapse GRPO-Trained LLM Agents -- and What Actually Works
This work identifies a pathological failure mode in dense prediction rewards for LLM agents under group-normalized RL (GRPO), termed the 'dark room' pathology, where agents collapse into degenerate absorbing states despite perfect prediction accuracy. Through systematic ablation, the authors demonstrate that GRPO's std normalization amplifies within-group variance in dense signals, leading to catastrophic policy collapse. They propose a variance-profile criterion to predict such failures and validate it through preregistered experiments. Controlled experiments reveal that auxiliary-loss channels outperform reward channels by ~20 points, even without correct labels. Findings are supported by single-seed endpoints, with replication and group-size controls underway.
grpodense predictionllm agentsvariance-profileauxiliary-loss
Filter Learning for Subgraphs: Algebras and Performance Risk Bounds
The paper introduces a framework for subgraph filter learning (SFL), where operators on partially observed subgraphs approximate ambient graph filters. The method formulates SFL as a statistical learning problem and develops a subgraph filter algebra based on distance-aware Laplacian constructions, enabling structured filter approximation. Theoretical performance risk bounds under least squares loss quantify approximation quality. Experiments on real-world datasets demonstrate that the proposed algebraic models outperform polynomial filters, distribution-agnostic operators, and direct numerical learning baselines.
subgraph filter learninggraph signal processinglaplacian constructionsperformance risk boundsleast squares loss
DART: A Degradation-Aware Recurrent Transformer for Archival Film Restoration
DART introduces a degradation-aware recurrent transformer for archival film restoration, explicitly modeling compound degradations through predicted soft defect masks. The method propagates these spatiotemporal masks to guide temporal fusion and condition restoration on damage location/severity, unlike implicit approaches. Evaluations on real archival benchmarks demonstrate superior no-reference perceptual quality versus prior architectures, with improved temporal consistency and artifact removal while maintaining computational efficiency.
archival film restorationsoft defect maskrecurrent transformertemporal fusionno-reference quality
Transformer-based Diffusion models for Hydrological Time Series Probabilistic Imputation and Forecasting
The study proposes a transformer-based diffusion model for probabilistic imputation and forecasting of hydrological time series, addressing challenges in hydro-system monitoring with sparse observations. The method combines transformer architectures with diffusion models to capture complex temporal dynamics across six sites in North-East France, using 15 years of quality-controlled data. Evaluation shows superior performance over baselines in reproducing hydrological patterns, particularly for missing data imputation and future condition forecasting, demonstrating effective sampling of realistic time series distributions.
transformer-based diffusionhydrological time seriesprobabilistic imputationmissing datawater quality
Automated Synthesis and Adversarial Validation of Executable Causal Research Pipelines
The Artificial Intelligence-based Epidemiology Research Assistant (ARA) framework addresses silent failures in automated research systems by integrating causal design principles, synthetic data generation, and adversarial validation. ARA translates natural language questions into structured causal protocols and executable code, using Structural Causal Models (SCMs) with known ground-truth effects for synthetic data generation. Evaluation on the Automated Causal Reasoning Benchmark showed that while ARA did not improve numerical accuracy over standard LLM-based generation, it shifted failure modes to surface protocol concerns and diagnostic failures, emphasizing the importance of validity indicators in causal claims.
automated research systemsstructural causal modelsadversarial validationcausal protocolssynthetic data generation
Agree on the Model, Verify the Inference: GKR Protocols for HND-Based Transformer Inference
The paper introduces GKR-HND, a registered-model protocol for verifying outsourced Transformer inference while preventing model substitution and incomplete execution. The method combines GKR proof transcripts with weight openings checked by a retained verifier, delegating expensive public evaluations to a computation worker under non-collusion assumptions. Experiments on pretrained Homomorphic--Nonhomomorphic Decomposition (HND) Transformers confirm the protocol's validity for proof paths and delegated computation without dense-matrix replay.
transformer inferencegkr protocolhomomorphic--nonhomomorphic decompositionmodel verificationoutsourced computation
Safety-oriented sidewalk and road segmentation for smartphone-based assistive navigation
The study presents a safety-oriented semantic segmentation framework for blind and visually impaired pedestrian (BVIP) navigation, introducing SENSATION-DS, a pedestrian-view dataset with 2,752 image-mask pairs and a nine-class taxonomy. Five segmentation architectures were evaluated using staged target-domain adaptation with synthetic images and Segment Anything Model 2 (SAM2) pseudo-labels, assessing performance via mIoU, Road-as-Sidewalk Error Rate, and Android runtime. UPerNet-MobileNetV3 achieved the highest mIoU (0.715), while DeepLabV3Plus-MobileNetV3 minimized Road-as-Sidewalk errors (0.079) and optimized runtime (7.383 FPS), demonstrating the need for balanced accuracy, safety, and deployability in assistive navigation models.
semantic segmentationassistive navigationdomain adaptationpseudo-labelingmobile deployment
Approximate Quantum State Preparation Through Proximal Policy Optimization
A quantum architecture search framework for approximate quantum state preparation (QSP) is proposed, addressing the exponential growth of the search space with qubit count. The method employs deep reinforcement learning via a proximal policy optimization agent to optimize circuit fidelity while minimizing gate count. Experiments on 2-5 qubits demonstrate approximation errors as low as $10^{-14}$ for predefined (Bell, GHZ, W, Dicke) and random states.
quantum state preparationproximal policy optimizationquantum architecture searchreinforcement learningqubit circuits
Smooth Neural Point Processes via B-Splines
The authors propose a neural temporal point process (TPP) model that directly parametrizes the conditional intensity function (CIF) using B-spline basis functions with neural network-predicted coefficients. This approach enables exact negative log-likelihood evaluation without numerical integration, supports arbitrary neural architectures, permits parallel training, and incorporates smoothness regularization via the integrated squared second derivative. Experiments on synthetic and real-world datasets demonstrate improved computational efficiency and predictive accuracy over baseline neural TPP methods.
temporal point processesconditional intensity functionb-splinesneural networksmaximum likelihood estimation
A Polynomial Architecture-Attribution Co-Design Framework for Exact Aumann-Shapley Attribution in GNNs
The paper introduces APEX, a polynomial architecture-attribution co-design framework enabling exact Aumann-Shapley attribution in graph neural networks (GNNs). The method employs PolyGIN, a GIN-style polynomial GNN preserving bounded multivariate polynomial forms, allowing exact path integral computation via Gauss-Legendre quadrature with $2^{L-1}$ evaluations for $L$-layer networks. Experiments demonstrate PolyGIN's competitive predictive performance and APEX's superior attribution fidelity with reduced computational cost compared to approximation-based baselines.
graph neural networksaumann-shapley attributionpolynomial architecturepath integralgauss-legendre quadrature
CASC: Causal Adversarial Subspace Clustering for Multivariate Spatiotemporal Data
The paper proposes Causal Adversarial Subspace Clustering (CASC), a novel framework for discovering evolving latent regimes in high-dimensional spatiotemporal data. CASC integrates a U-Net-inspired adversarial clustering architecture with FAConvLSTM layers and a graph attention transformer-based self-expressive network to model spatial-temporal dependencies. It introduces two learning objectives: Causal Subspace Preservation Loss to align clusters with latent causal relationships, and Dynamic Temporal Subspace Evolution Loss to capture nonstationary regime transitions. The method shifts subspace clustering from correlation-driven to causal-temporal regime discovery.
subspace clusteringspatiotemporal datacausal learningadversarial networksgraph attention transformer
Automatic knot selection in smooth additive models
The authors propose a novel explicit knot-selection technique for generalized additive models (GAMs), combining adaptive splines (A-splines) methodology with a customized Fellner-Schall scheme for parameter tuning. This approach addresses the challenge of determining knot number and placement in B-spline regression, traditionally handled via regularization methods like P-splines. Evaluated on synthetic and real datasets, the method achieves performance comparable to P-splines and state-of-the-art knot-selection techniques while using significantly fewer basis elements.
b-spline regressionknot selectiongeneralized additive modelsadaptive splinesfellner-schall scheme
Nipping the Butterfly Effect in the Bud: Self-Output Fine-Tuning for Autoregressive Weather Prediction
We propose Self-Output Fine-Tuning (SOFT), a plug-and-play method that mitigates error growth in autoregressive Deep Learning Weather Prediction (DLWP) by calibrating biased input distributions using the model's own one-step predictions. Theoretical and empirical analysis reveals that rapid error growth stems from a feedback loop between output errors and input distribution shifts, detectable as early as the first autoregressive step. Experiments demonstrate that SOFT achieves state-of-the-art performance on long-horizon forecasting tasks, substantially reducing prediction errors and distributional discrepancies. This work highlights the importance of addressing fundamental pipeline issues in DLWP.
autoregressivedeep learning weather predictiondistributional shiftself-output fine-tuningbutterfly effect
Spectral Transformation for Layer-wise Global Rank Discovery in Federated LoRA for Vision Transformers
The paper introduces SpecTraL, a spectral transformation method for federated fine-tuning of Vision Transformers with low-rank adapters (LoRA). The approach performs orthonormal Householder Transformation on stacked client LoRA modules in low-rank space, avoiding dense reconstruction and auxiliary refinement. Leveraging the Spiked Covariance Model, it analytically separates global signals from non-IID noise to determine optimal layer-wise ranks without hyperparameter tuning. Experiments on ViT-B/16 and ViT-L/16 with DomainNet and NICO++ show improved accuracy-communication trade-offs, reduced server computation, and eliminated rank selection tuning.
federated learninglow-rank adaptersvision transformersspectral transformationspiked covariance model
From Evaluation to Optimisation: Hierarchy-Aware Training Signals for CWE Prediction in Python
This paper validates the use of the ALPHA benchmark's taxonomy-aware penalty as a training signal for CWE-level vulnerability prediction in Python, comparing three delivery mechanisms: supervised fine-tuning, dual-head classification loss, and reinforcement learning (GRPO). Results show supervised methods regress below zero-shot baselines under distribution shift, while GRPO reduces Qwen2.5-Coder-7B's cumulative penalty by 27.9% (greedy decoding) and 25.5% (sampled decoding), achieving parity with a 4.5x larger teacher model. The effectiveness of hierarchical penalties depends on delivery directness.
cwe predictionhierarchical penaltyreinforcement learningdistribution shiftzero-shot baseline
Counterfactual Explainability Framework With CycleGAN And Counterfactual-Classifier Alignnment Score for Retinal Disease Classification
The study introduces CounterFundus, a CycleGAN-based counterfactual explainability framework for retinal disease classification from fundus images, combining EfficientNet-B5 for detection with disease-to-normal image translation. The framework generates visually interpretable counterfactuals (healthy counterparts) and difference maps to localize disease-associated changes, while proposing the Counterfactual-Classifier Alignment Score (CCAS) to quantify spatial agreement between counterfactual maps and classifier saliency via Spearman correlation, IoU, and pointing accuracy. Results show CCAS-filtered counterfactuals improve classification performance and align with classifier-relevant evidence, validated through EigenCAM-aligned evaluation and ablation studies.
counterfactual explainabilitycycleganefficientnet-b5fundus imagessaliency alignment
Regularized Optimization on Grassmann Manifold: Theory, Algorithm and Applications
The paper proposes Regularized Projection Matrix Approximation (RPMA), a framework for robust estimation of rank-$K$ projection matrices by incorporating regularization into spectral methods. The approach formulates the problem as optimization on the Grassmann manifold, derives optimality conditions, and develops efficient Riemannian gradient methods including a Cayley--SMW variant. Experiments show RPMA improves projection matrix recovery accuracy by 15-20% over conventional spectral methods in noisy clustering and community detection tasks.
grassmann manifoldspectral clusteringriemannian optimizationprojection matrixcommunity detection
Weight-norm Criticality: A Mechanism for Loss Spikes Induced by the Normalization and Weight Decay
The paper identifies weight-norm criticality as a novel mechanism for training instability in deep neural networks, distinct from learning-rate criticality. This phenomenon arises from the interaction between normalization (introducing scale-invariant weights) and weight decay (continuously shrinking parameter norms). As weight decay increases, scale-invariant weight norms approach zero, causing rapid sharpness growth in the loss landscape and abrupt loss spikes. Empirical validation in networks with scale-invariant components supports the proposed mechanism, explaining why excessive weight decay destabilizes training despite improving generalization.
weight-norm criticalityscale-invariant weightstraining instabilityweight decayloss landscape sharpness
ADABORD: a novel AdaBoost approach for ordinal classification
ADABORD introduces a novel AdaBoost framework for ordinal classification that explicitly incorporates ordinal information into both the base estimator and error function. The method employs decision trees with an ordinal Gini splitting criterion and updates sample weights using the absolute ranked probability score, which accounts for class ordering and distance. Evaluated on the TOC-UCO repository, ADABORD significantly outperforms seven state-of-the-art methods, particularly on datasets with five or more classes where ordinal structure is more pronounced. Statistical analysis supports these findings. Source code and experimental protocols are publicly available for reproducibility.
adaboostordinal classificationgini splittingranked probability scoretoc-uco
The Weight of Silence: A Causal Case for Weights Over the Scratchpad in Latent Chess Reasoning
This study challenges the assumption that latent reasoning functions as an inference-time scratchpad in language models, demonstrating instead that its primary effect is parameter shaping during training. Through a staged latent-reasoning curriculum followed by reinforcement learning (RL) on a chess-playing model, legality improved from 48% to 61%, while checkmate confabulation was eliminated. A six-condition causal intervention suite revealed robustness to latent thought vector disruptions, with exact-zero corruption causing legality to collapse to 1% pre-RL versus 9% post-RL. Results indicate RL adds robustness to disruption rather than reliance on thought content, contrasting with findings in math and logic tasks.
latent reasoningreinforcement learningcausal interventionscratchpadparameter shaping
Best-of-Evidence: Best-of-N Selection under Partial Verification
The paper introduces Best-of-Evidence (BoE), an inference-time selection framework extending Best-of-N (BoN) to tasks with partial verification, where only specific claims or regions within candidates can be reliably evaluated. BoE constructs a signed candidate--factor graph to represent reusable claims and optimally allocates a limited evidence budget to influence selection, with theoretical analysis showing O(log K) query complexity for shared factors. Experiments on four medical VQA tasks demonstrate BoE's ability to improve fixed-pool selection when evidence is reliable and contrastive, while also identifying scenarios where gains are limited by channel quality or candidate generation.
partial verificationcandidate--factor graphevidence budgetinference-time selectionmedical vqa
An Analytically Trained Variational Surrogate for Quantum Phase Estimation on NISQ Hardware
The authors present an analytically trained variational surrogate framework for Quantum Phase Estimation (QPE) that avoids quantum circuit simulation by using a shallow Variational Quantum Circuit (VQC) trained classically via the Dirichlet kernel. The method leverages Full Configuration Interaction ground-state energy, ancilla qubit count, and time evolution parameters to eliminate exponential scaling. Experimental validation on IBM Quantum hardware with H$_2$ molecules shows optimal performance with linear entangler topology and single-layer depth, achieving chemical accuracy (1 kcal/mol) while maintaining linear scaling.
quantum phase estimationvariational quantum circuitdirichlet kernelnisq deviceschemical accuracy
Three-Pronged Spectral Control for Federated Parameter Efficient Fine Tuning
The paper proposes TRISHUL, a spectral-control framework for robust federated parameter-efficient fine-tuning (PEFT) under non-IID client heterogeneity. The method employs shared frozen multi-head low-rank bases for exact aggregation, nuclear norm proximal shrinkage to suppress client-specific high-rank components, and concave water filling for layer-wise budget allocation. Evaluated on CIFAR-100, SVHN, 20 Newsgroups, MRQA, and GLUE with LLaMA3.2-1B, TRISHUL improves convergence, stability, and final performance over federated LoRA baselines, particularly under strong heterogeneity, without additional per-round communication costs.
federated learningparameter-efficient fine-tuninglow-rank adaptationspectral controlnon-iid heterogeneity
RadioTrace: Transmitter-Aware Diffusion for Radio Map Estimation without Deployment-Time Fine-Tuning
RadioTrace introduces a transmitter-aware diffusion framework for radio map (RM) estimation without deployment-time fine-tuning, integrating sparse RSS measurements with a frozen pre-trained diffusion prior. The method iteratively refines transmitter (Tx) coordinates during denoising, aided by propagation-guided K-means initialization for robust Tx estimation and stochastic stability analysis for Tx-coordinate updates. Experiments show competitive performance with state-of-the-art learning-based methods under random sampling (0.5-1.2 dB RMSE improvement) and strong reconstruction quality in restricted-area scenarios.
radio map estimationdiffusion priortransmitter-awarerss measurementsdenoising loop
HierarchicalDAEW: Domain-Aware Edge-Weighted Graph Convolution with Evidential Uncertainty for Multi-Section Spatial Gene Expression Prediction from H&E Histology
HierarchicalDAEW introduces a dual-graph architecture for predicting spatial gene expression from H&E histology, addressing tissue heterogeneity and prediction reliability. The method employs a Domain-Aware Edge-Weighted convolutional operator on a spot graph to handle tissue domains explicitly, followed by a gene-level graph combining STRING-DB protein-protein interactions with tissue-specific co-expression via attention gating. Evidential uncertainty estimation provides calibrated confidence intervals. Evaluated on six human Visium sections across four tissue types, it outperforms 13 baselines in expression correlation, with ablations confirming the necessity of domain-aware edge typing and hierarchical depth.
spatial transcriptomicsgraph convolutionevidential uncertaintyprotein-protein interactionattention gating
Information-Theoretically Secure Aggregation for Lightweight Federated Learning: Resilient to Dropouts and Adversaries
We propose an information-theoretically secure aggregation framework for sign-based federated learning (FL) that ensures privacy and efficiency. The method employs single-round secure multiplication to compute majority vote polynomials with reduced degree via inverse-form exponent reduction, achieving end-to-end security under honest-majority assumptions. The framework introduces linear offline complexity and storage with minimal online communication, reducing latency by up to 85.7% and communication by up to 99.5%. It also leverages MDS-code-based decoding for robustness against dropouts and adversarial behaviors, improving accuracy by up to 20.65% and 10.74%, respectively. This enables scalable, low-latency, and secure aggregation in resource-constrained FL settings.
federated learningsecure aggregationmajority votemds-codesignsgd
Machine Learning for Charge State Characterization of Isolated Double Quantum Dots
The authors present two compact convolutional neural networks for automated characterization of isolated double quantum dots via charge stability maps (CSMs). CSMClassifier (94% accuracy) detects charge instability and artifacts, while ChargeLineNet (95.3% accuracy) localizes charge-transition lines and determines electron occupancy. Trained on 32 SiMOS devices at 1K, the 6.5MB combined pipeline achieves 93.8% occupancy accuracy on clean images, with synthetic pre-training improving label efficiency and sub-60ms inference on lab hardware.
quantum dotscharge stability mapsconvolutional neural networksautomated tuneupcryogenic probing
Position Bias is Hidden Behind Ceiling Effects: A Permutation Diagnostic for LLM Benchmarks
The paper introduces inspect_permute, an open-source tool for detecting position bias in multiple-choice LLM evaluations through exhaustive answer-order permutations and chi-squared/Cramer V analysis with bootstrap confidence intervals. The method was applied to four models (gpt-4o-mini, claude-haiku-4-5, gemini-2.5-flash, grok-3) on five MMLU subjects, totaling 24,000 API calls under temperature-0 generation. Results show position bias is statistically detectable only in a 60-95% base-accuracy 'Goldilocks zone', with two mechanism types identified: monotone A-to-D decrease (processing_load) and non-monotone D-drop (content_ambiguity). Frontier-tier models' position bias remains undetectable due to ceiling effects.
position biasmultiple-choice evaluationchi-squared testceiling effectsgoldilocks zone
Offline RL with Hierarchical Action Chunking
The paper introduces Hierarchical Implicit Q-Chunking (HiQC), an offline goal-conditioned RL algorithm addressing long-horizon tasks by combining high-level latent planning with low-level action chunking. HiQC conditions the low-level critic on extended action sequences, enabling unbiased k-step value backups and compressing the horizon at both planning and execution levels. Theoretical analysis shows tighter value error bounds compared to standard hierarchical or flat chunking approaches. Empirical evaluation on OGBench demonstrates HiQC's superior performance, particularly in long-horizon navigation tasks like humanoid-giant.
offline reinforcement learninggoal-conditioned rlaction chunkingvalue error boundshierarchical planning
Robust Asynchronous Q-Learning under Reward and State Corruption via Batching
The paper introduces \texttt{BR-Async-Q}, a robust asynchronous Q-learning algorithm resilient to adversarial corruption of both rewards and states under the Huber contamination model. The method employs epoch-based batching of online data to reduce variance and constructs robust Bellman operator estimates. Theoretical analysis shows an $\ell_\infty$ error bound matching vanilla Q-learning, plus an additive term scaling with corruption fraction, achieving minimax optimal dependence when only rewards are corrupted.
asynchronous q-learninghuber contaminationbellman optimality operatoradversarial robustnessbatch learning
New Complexity-Theoretic Frontiers of Tractability for Neural Network Training
The article advances computational tractability bounds for neural network training by establishing new polynomial-time algorithms for optimal training. For ReLU-activated networks, it proves tractability for architectures where hidden neurons have out-degree 1, improving upon prior work by Arora et al. For linear-activated networks, it introduces a novel data throughput condition enabling the first non-trivial polynomial-time solvable class. These results expand the known frontiers of tractable neural network architectures.
neural networkscomputational complexitypolynomial-time algorithmsrelu activationlinear activation
External Clustering Validation by the Homogeneity-Parsimony Trade-off
The paper introduces normalized homogeneity and parsimony scores for external clustering validation, addressing the trade-off between class-label informativeness and cluster fragmentation. Building on the information bottleneck principle while avoiding lossy compression rewards, the method ensures monotonicity under cluster refinement. The framework extends to set-matching and pair-based scores, unifying common evaluation criteria and recovering binary classifier ROC curves. Applications in feature selection and algorithm comparison demonstrate Pareto-optimal solution identification through joint score analysis.
clustering validationinformation bottleneckhomogeneity-parsimony trade-offmonotonic refinementpareto-optimal solutions
Memoir: Should a Model Write to Its Memory While It Thinks?
Memoir introduces a novel architecture combining fast per-sample memory, shared slow parameters, and variable-depth latent recurrence, with a key innovation: allowing memory writes during pondering iterations. The study compares coupled (read-write) and read-only memory variants on procedural associative recall with key interference, using matched 81,738-parameter models. Results show read-only memory achieves superior initial recall (0.6557 vs 0.5203) and faster convergence, though both variants reach perfect recall given sufficient training. The predicted failure of memory corruption was not observed, and kernel optimizations reduced forward pass time by 61% (0.907ms to 0.351ms).
memory-augmented networkslatent recurrenceassociative recallpondering mechanismdelta-rule optimization
Toward Mechanistic Interpretability of an AI Foundation Model Fine-Tuned for Atmospheric Chemistry
The study conducts the first mechanistic interpretability analysis of Microsoft's Aurora, a weather forecasting foundation model fine-tuned for atmospheric chemistry. By imposing controlled chemical perturbations and examining internal representations via sparse autoencoders, the authors evaluate whether Aurora learns physical mechanisms or statistical patterns. Results show Aurora captures first-order ozone-nitrogen relationships but fails to enforce chemical constraints, produces inconsistent species combinations, and lacks chemistry-specific internal structure. The work establishes a framework for assessing AI-based chemical forecasting models beyond benchmark accuracy.
mechanistic interpretabilityfoundation modelatmospheric chemistrysparse autoencoderschemical transport models
Memory-Computation Tradeoffs in Semi Amortized Parametric Optimization
The paper establishes fundamental memory-computation tradeoffs in semi-amortized parametric optimization, where offline solutions guide online projected gradient descent. For μ-strongly convex objectives, it provides tight bounds on memory needed for ε-accuracy under K online iterations; for convex objectives with β-growth (β>2), it identifies a phase transition in K beyond which additional memory yields no improvement. The analysis introduces a general framework quantifying acceleration costs via online optimizer convergence rates and solution map sensitivity. Experiments on parameterized ridge regression validate the theoretical tradeoffs.
parametric optimizationmemory-computation tradeoffprojected gradient descentsolution map sensitivityphase transition
GaugeQuant: Online Learning of Quantization-Optimal Bases from LLM Symmetries
GaugeQuant introduces an online learning method for quantization-optimal bases by leveraging internal continuous symmetries in transformers. The approach incorporates a LogSumExp term into the loss function to break symmetries, selecting bases that minimize activation outliers while preserving the language modeling objective via a stop-gradient operator. It eliminates the need for calibration data or quantization simulation, adding negligible training overhead. Evaluated on LLaMA-2 7B with W4A4 quantization (group size 128), perplexity improves from 8.22 to 6.73; under W4A16, perplexity drops from 11.16 to 5.45, outperforming post-training methods requiring frozen models and calibration datasets.
transformersquantizationsymmetrieslogsumexpperplexity
Adaptive Confidence-weighted Expansion for Trustworthy Multi-Omics Multimodal Fusion
The paper introduces Adaptive Confidence-weighted Expansion (ACE), a framework enhancing trustworthiness in multimodal fusion for medical prognosis. ACE generates complementary modalities from intra-modality correlations and employs a dual-level confidence mechanism: modality reliability weighting and global trust scoring. Evaluated on four multi-omics datasets (BRCA, KIPAN, LGG, ROSMAP), ACE outperforms state-of-the-art methods in classification accuracy and confidence calibration, demonstrating robustness against noisy data.
multimodal fusionconfidence calibrationmulti-omicsadaptive weightingmedical prognosis
Pipelined Gradient Coding
The paper introduces pipelined gradient coding (GC) to mitigate straggler effects in distributed machine learning. By segmenting gradient evaluation across steps and limiting workers to single-partition evaluations per step, the method reduces training time versus traditional GC. The authors develop pipelined versions for fractional repetition and cyclic repetition schemes, proving convergence guarantees. Experiments on cloud infrastructure demonstrate both faster training and improved convergence compared to baselines.
gradient codingstraggler mitigationdistributed trainingfractional repetitioncyclic repetition
Cardinality-Decomposed Loss: Matching Training Objectives to Relation Structure in Heterogeneous Recommendation Graphs
The paper introduces Cardinality-Decomposed Loss (CDL), a hybrid loss function combining Cross Entropy and Bayesian Personalized Ranking to address embedding collapse in heterogeneous recommendation graphs. CDL separately optimizes for one-to-one (attribute) and one-to-many (preference) relations, resolving parameter space conflicts between the two objectives. Evaluations on MovieLens-1M, Last.fm-360K, and three other datasets demonstrate improved attribute embedding discriminability and NDCG gains when attributes align with preferences, controlled by a λ parameter sensitive to semantic alignment and topology leakage in graph structure.
heterogeneous graphsembedding collapsebayesian personalized rankingcardinality decompositionrecommendation systems
LLMs Get Lost in Evolving User Intent
The study investigates LLMs' capability to track evolving user intent in multi-turn conversations, a critical gap in current static evaluation paradigms. The authors introduce a framework that adapts single-turn tasks into dynamic conversations where intent incrementally changes, preserving original evaluation protocols. Results show significant performance drops across model families (GPT-3.5, LLaMA-2) in evolving-intent settings despite strong static performance, revealing a fundamental limitation in LLMs' interactive reasoning.
evolving intentmulti-turn conversationstatic evaluationinteractive reasoninguser intent tracking
Leaky Language Models: Stealing Architecture and Inference Optimizations via Per-Token Timing
LeakyLMs introduces novel side-channel attacks that extract proprietary model architectures and inference optimizations from production language models using only per-token generation timing. The method constructs a GPU timing model to analyze latency patterns, enabling two attacks: (1) detecting inference optimizations like speculative decoding (identifying Google Gemini Flash 2.5's 128K-token draft context window) and (2) recovering architectural parameters (hidden dimension, layers, attention heads). Evaluations on LLaMA models show 90% top-10 accuracy for architecture recovery.
side-channel attackspeculative decodingtransformer architecturetiming analysismodel extraction
Perspective Latents as an Architectural Condition for Causal Emergence in Active Inference Agents
The study investigates causal emergence in active inference agents by analyzing Integrated Information Decomposition ($Φ_r$) within an architecture separating fast perception latents ($z$) from slow global latents ($g$). Using a reward-free regime-switching protocol, results show $Φ_r$ concentrates in $g$ with aggregate magnitude being architectural and decreasing during training. Learning effects emerge at the atom-compositional level: decoupling becomes regime-invariant while downward causation handles regime-dependent adjustments. The findings position $g$ as the architectural locus for $Φ_r$-relevant temporal organization and challenge scalar $Φ_r$ as a direct measure of learned integration.
active inferenceintegrated information decompositioncausal emergencelatent variablesregime-switching
CEDAR: Causal Edge Discovery for Autoregressive Processes
CEDAR introduces a constraint-based method for lagged causal edge discovery in sparse autoregressive time series. The approach screens cross-variable lags using AR(1)-residualized, U-centered distance correlation, followed by two targeted conditional-independence tests per candidate lag. It employs stable MCI pruning to remove indirect edges and optionally adjusts for nonstationarity using deterministic C-nodes. CEDAR requires O(d^2) conditional-independence tests post-screening in sparse regimes, maintaining edge-level interpretability. The method is particularly effective with scarce data and lag-1 self-dynamics, though richer conditioning sets become preferable with larger T or higher-order autoregressive effects.
autoregressive processescausal edge discoveryconditional-independence testsnonstationaritylagged causality
Attribution Markets: A Fisher-Market Formulation for Fractional Credit Assignment Between Planned Tasks and Performed Actions
The paper introduces attribution markets, a Fisher-market formulation for fractional credit assignment between planned tasks and performed actions. The method models tasks as budget-constrained buyers and actions as divisible goods, using fused text/structural/temporal signals for valuation. Key contributions include market instruments (seller reserve price, buyer cash option) ensuring conservation and budget caps, a concave completion utility for progress discounting, and an entropy-regularized generalization to handle affinity noise. Theoretical results include existence and local uniqueness conditions for fixed points. Empirical validation shows sensitivity to noise, addressed via adaptive regularization. The work connects to multi-touch attribution and optimal transport.
fisher marketcredit assignmentoptimal transportentropy regularizationquasi-linear utility
Explanation-Based Runtime Verification for Trustworthy ML-driven Optical Networks
The paper introduces explanation-based runtime verification, a method leveraging XAI techniques to assess ML decision reliability in optical network automation. By evaluating explanation coherence and physics grounding consistency during deployment, the system intercepts uncertain decisions before execution. Experiments on lightpath quality-of-transmission classification demonstrate significant error interception while maintaining high automation rates.
explainable airuntime verificationoptical networksmachine learningnetwork automation
End-to-End Learning of Safe Optimal Feedback Control in High Dimensions with Control Barrier Function Layers
The paper introduces a scalable method for learning high-dimensional feedback controllers with hard safety guarantees using control barrier functions (CBFs). By combining operator splitting with Jacobian-Free Backpropagation (JFB), the approach overcomes computational and differentiation bottlenecks in embedding quadratic-program-based safety filters during end-to-end training. Theoretical justification via nonsmooth analysis is provided, with empirical validation on multi-agent nonlinear control problems featuring state dimensions up to 1200 and control dimensions up to 400.
control barrier functionsjacobian-free backpropagationoperator splittingend-to-end learningnonlinear control
Cross-Domain Generalization in Optical Networks via Joint Contrastive and Classification Learning
The authors propose a joint contrastive and classification learning method to improve cross-domain generalization in optical networks, addressing performance degradation in unseen topologies. Their technique simultaneously optimizes representation learning and task objectives, shaping a domain-invariant latent space. Evaluated on lightpath quality-of-transmission estimation, the approach outperforms baselines and demonstrates strong few-shot adaptation capabilities with limited fine-tuning.
cross-domain generalizationcontrastive learningoptical networksrepresentation learningquality-of-transmission
Writhe-Based Polymer Link Classification Using Machine Learning
A feedforward neural network achieves 97% accuracy in classifying thermally equilibrated configurations of the first six prime links using the writhe density matrix, demonstrating robustness across temperatures and link lengths. The method leverages machine learning to address the computationally prohibitive task of exact numerical calculation of topological invariants, particularly for complex link topologies like Borromean rings and multi-component links. Performance degrades rapidly with topology-altering Gaussian noise, indicating the writhe density matrix's sensitivity to topological features.
writhe density matrixfeedforward neural networktopological invariantsgaussian noiselink classification
PhysCoRe: Physics-Corrected Residual World Models for Material-Aware Deformable Dynamics
PhysCoRe introduces a physics-corrected residual world model for predicting deformable object dynamics under robotic manipulation, addressing limitations of per-object optimization and end-to-end learning. It integrates a differentiable Material Point Method (MPM) simulator with two neural networks: Material from Motion (MfM), which infers per-particle elasticity from visual observations, and Residual from Dynamics (RfD), which corrects simulator biases. MfM adapts to novel objects through limited interactions, leveraging predictive uncertainty for exploration. Experiments demonstrate PhysCoRe's superior prediction accuracy and reliable confidence distribution across object geometry, enabling confidence-guided exploration in real-world deformable-object manipulation sequences.
material point methodphysics-corrected residualdeformable dynamicsmaterial refinementconfidence-guided exploration
Masked Topology Modeling for Self-Supervised Learning on Parametric CAD
The paper introduces Masked Topology Modeling (MTM), a self-supervised pretraining task for parametric CAD data in boundary representation (B-Rep) format. MTM reconstructs masked edges in face-adjacency graphs by predicting convexity and curve type from message-passing face features, combined with MoCo-style contrastive learning and BFS-connected face-region reconstruction. Pretrained on ABC and a new procedural dataset, the method demonstrates strong performance across multiple benchmarks.
masked topology modelingboundary representationself-supervised learningparametric cadface-adjacency graph
One Round Is All You Need: Analytic Federated Learning for Task-Heterogeneous Multi-Label Medical Image Classification
The authors introduce an analytic federated learning framework for multi-label medical image classification under task heterogeneity, where each client annotates only a subset of disease categories. The method replaces iterative gradient optimization with three closed-form operations: balanced label projection, per-class absolute aggregation, and optional pseudo-label refinement. This approach requires at most two communication rounds, regardless of task heterogeneity or client count. Experiments on ChestXray14 show improvements of up to 18.44 BACC points and 13.24 AUC points over FedMLP, while reducing communication overhead.
federated learningtask heterogeneitymulti-label classificationclosed-form optimizationmedical imaging
Algorithmic Approaches to Sequential Decision-Making and Social Epistemology
This thesis develops algorithmic frameworks for sequential decision-making and social epistemology, providing theoretical insights and intervention strategies. Part I analyzes the improving multi-armed bandits problem, establishing nearly matching upper and lower bounds, and demonstrates polynomial sample complexity for learning algorithms in a data-driven design framework. Part II investigates social and behavioral influences, formalizing pessimism traps and proposing financial interventions to mitigate them, while also examining the role of grit in ambitious decision-making. The work bridges abstract models and real-world complexity, offering algorithmic tools for social epistemology.
sequential decision-makingmulti-armed banditsdata-driven algorithm designpessimism trapssocial epistemology
Are Single-Token Sparse Autoencoder Features Causally Necessary? Layer-Depth and SAE-Family Effects
This study investigates the causal necessity of single-token sparse autoencoder (SAE) features across SAE families and layer depths in large language models. Analyzing 3.9M features from six models and three SAE families using zero-ablation at full layer depth, the authors find that single-token features cluster 4.7x tighter in decoder space and concentrate in early layers. Ablating these features yields significant logit reductions in 178 of 208 full-layer conditions, with depth determining downstream damage or direct output shaping. Cross-family causal differences exceed within-family scale effects, highlighting training methodology's impact on interpretability claims.
sparse autoencoderzero-ablationlogit reductionsdecoder spacetraining methodology
Label-Free Finite-Volume-Residual Training of Attention Graph Neural Networks for Coupled Thermo-Fluid Fields
We propose a label-free training method for attention graph neural networks (GNNs) that minimizes finite-volume method (FVM) residuals of governing equations for coupled thermo-fluid fields, eliminating the need for costly labeled data generation. The approach evaluates residuals directly on the mesh, leveraging the FVM-loss as a training signal. Evaluations across four scenarios demonstrate that the FVM-loss model achieves 2.3-2.8% normalized root-mean-square error (nRMSE) on steady-state benchmarks and outperforms supervised baselines on parametric transient cases. The method maintains close agreement with computational fluid dynamics (CFD) references, including buoyancy-energy coupling, while avoiding data-generation costs entirely.
finite-volume methodattention graph neural networksthermo-fluid fieldslabel-free trainingnormalized root-mean-square error
STeMP: Spatio-Temporal Modelling Protocol
The authors propose STeMP (Spatio-Temporal Modelling Protocol), a standardized reporting framework for spatio-temporal machine-learning models in environmental research. STeMP addresses the lack of transparency and comparability in such models by providing a structured protocol with three sections: Overview, Model, and Prediction. The protocol is implemented via an R-package, which includes a web application for semi-automated protocol completion and warnings for common pitfalls. The framework aims to enhance trust and reproducibility while guiding model development and evaluation.
spatio-temporal modellingmachine learningstandardized reportingenvironmental researchreproducibility
Detecting Neural Network Failures through Spectral Analysis of Internal Activations
The paper introduces Spectral Drift, a frequency-domain measure of instability in neural network internal activations that correlates with misclassifications (1.9% increase vs correct predictions, p<0.001). It proposes Self-Detecting Neural Networks (SDNN), which monitors spectral dynamics across layers via Short-Time Fourier Transform, wavelet decomposition, and statistical moments, using a lightweight detector (5% overhead). On CIFAR-10, SDNN achieves 79.0% AUROC, outperforming confidence-based methods by 25-30 percentage points, with wavelet and statistical features proving most consistent.
spectral driftinternal activationswavelet decompositionfailure detectionself-detecting networks
Instance Hardness-Based Relevance for Imbalanced Regression
The study introduces Instance Hardness-based relevance (InHaR), a novel relevance function for imbalanced regression that incorporates learning difficulty to identify rare instances beyond traditional target-value-based approaches. InHaR addresses limitations in bimodal distributions by dynamically assessing rarity through both target distribution and algorithmic learning difficulty. Experiments show that InHaR-guided resampling (Random Oversampling and Gaussian Noise) outperforms traditional methods, improving predictive performance in complex scenarios. The method is publicly available with code and datasets.
imbalanced regressioninstance hardnessrelevance functionbimodal distributionresampling strategies
SPECTRA: State-Space Exogenous Context and Temporal-Frequency Resolution Architecture for Probabilistic Energy Forecasting
The paper introduces SPECTRA, a state-space architecture for probabilistic energy forecasting that jointly models deterministic trends and stochastic residuals. The method separates deterministic and residual streams, aligns exogenous context with both, refines the backbone via multi-resolution spectral-temporal state-space modeling, and estimates quantile boundaries. Evaluated on load, price, solar, and wind forecasting, SPECTRA achieves best continuous ranked probability score in 14/18 settings, reducing average CRPS by 5.74% and upper-tail quantile risk by 7.27% versus baselines.
probabilistic forecastingstate-space modelmulti-resolution decompositionquantile estimationexogenous context
Multi-stage Dynamic Selection for Cross-Project Defect Prediction
The paper proposes a two-stage multiple classifier system (MCS) for Cross-Project Defect Prediction (CPDP) to address distribution shift between training and target projects. Stage one selects diverse classifiers generalizing across training projects, while stage two performs module-level classifier selection during inference. Evaluated on 82 projects from four CPDP benchmarks, the method outperforms state-of-the-art approaches by dynamically adapting to module-specific characteristics.
cross-project defect predictionmultiple classifier systemdistribution shiftdynamic selectionmodule-level prediction
Exact ReLU realization of affine one-dimensional refinement iterates via residual memory and offset frames
The paper establishes an exact fixed-width ReLU realization for finite affine iterates of vector-valued refinement operators, achieving O(n) depth for M ≥ 3. The method introduces a residual memory controller to handle noninvertible residual dynamics via injective skew-products and offset frames to align forcing atoms. Results show applicability to compactly supported continuous piecewise linear forcing terms, with extensions for M=2 and finite-dimensional cases. The construction also enables linear-depth upgrades for recursive structures like open-curve and Hilbert-type constructions.
relu realizationaffine refinementresidual memoryoffset framespiecewise linear
Machine-Learned Compact Subspace Generation for Quantum Selected Configuration Interaction within Density Matrix Embedding Framework
The authors propose QSCI-RBM, a machine-learned compact subspace generation protocol using Restricted Boltzmann Machines (RBMs) integrated with Density Matrix Embedding Theory (DMET) for Quantum Selected Configuration Interaction (QSCI). The RBM learns dominant determinant distributions from quantum-sampled configurations, enabling targeted generation of high-probability configurations. Applied to a SARS-CoV-2 main protease complex, DMET-QSCI-RBM achieves chemical accuracy with only 4% subspace access, outperforming standard DMET-SQD requiring 20% subspace access.
quantum selected configuration interactionrestricted boltzmann machinesdensity matrix embedding theoryconfiguration subspacechemical accuracy
Generalized Kalman filter based temporal difference reinforcement learning
The paper introduces a generalized temporal-difference (TD) reinforcement learning framework based on conditional expectations, extending classical Kalman-based TD learning to nonlinear models and non-Gaussian distributions. The method formulates value and Q-function estimation as stochastic inference, recursively estimating both conditional expectations and second moments to quantify uncertainty. Discretization via polynomial chaos expansions or ensemble approximations ensures computational tractability. Validation on linear mass--spring--damper and nonlinear heat conduction problems demonstrates accurate value function estimation with uncertainty quantification.
temporal-difference learningkalman filterstochastic inferenceuncertainty quantificationnonlinear control
Fisher Widths: Local Learning Geometry and Anisotropic Recovery
The paper introduces Fisher widths, a pair of functionals (primal Fisher width $w_G(T)$ and inverse-Fisher width $w_{G^{-1}}(T)$) that quantify Gaussian-width complexity on statistical manifolds. The primal width measures local parameter fluctuations under Fisher-regular losses, achieving scale $w_G(H_r)/\sqrt n$ on small Fisher balls, while the inverse width characterizes anisotropic Gaussian measurements for sparse recovery, revealing support-sensitive geometry. A sharp inequality $w_G(T)w_{G^{-1}}(T)\geq w(T)^2$ establishes that Fisher anisotropy transfers complexity between geometries without reducing both widths below Euclidean scale.
fisher widthgaussian-width complexitystatistical manifoldssparse recoveryfisher information
AI-Driven Surrogate Models for Predicting Electrode-Scale Discharge Behavior in Lithium-Ion Batteries
The authors propose a deep learning surrogate model using Swin3D Transformer to predict electrode-scale discharge behavior in lithium-ion batteries, addressing computational bottlenecks in physics-based simulations. Key innovations include Gaussian Positional Encoding (GPE) for enhanced spatial feature representation and a Temporal Encoding module for non-linear timeseries modeling. Evaluated on an Electrochemical Simulation dataset, the method outperforms point cloud baselines in accuracy while reducing computational overhead by orders of magnitude, enabling scalable battery design optimization.
surrogate modelswin3d transformergaussian positional encodinglithium-ion batterieselectrochemical simulation
OPIUM: Mitigating Steering Externalities and Over-Refusal via Dual Objective Latent Optimization
OPIUM (Optimizing Protected Injections via Utility Manifolds) mitigates unintended side effects of activation steering in large language models, addressing both utility degradation and over-refusal. The method sanitizes steering vectors via representation matching, optimizing a new vector that preserves desired interventions while aligning with safer reference behaviors on problematic prompts. Evaluations across steering-externality and over-refusal scenarios show improved safety--utility tradeoffs compared to vanilla steering and directional ablation, demonstrating mitigation of harmful side effects directly in activation space.
activation steeringrepresentation matchingsafety--utility tradeofflatent optimizationover-refusal
📰 Industry Media (20)
How to Build an End-to-End OCR Pipeline with Baidu’s Unlimited-OCR for High-Resolution Images and Multi-Page PDF Parsing
The tutorial presents an end-to-end OCR pipeline using Baidu's 3B-parameter Unlimited-OCR model, demonstrating both single-page and multi-page PDF processing. The method involves GPU-accelerated inference with automatic bfloat16/float16 selection, comparing tiled Gundam mode (640px crops) for dense layouts against faster Base mode (1024px single view), and extends to multi-page parsing via PyMuPDF rasterization. Results show structured text extraction from synthetic documents containing tables, paragraphs, and cross-page references, with outputs preserved as text/Markdown/JSON. The pipeline achieves long-context generation (32k tokens) through n-gram repetition controls and layout-aware decoding without separate analysis stages.
ocr pipelinevision-language modelmulti-page parsingtiled inferencestructured text extraction
Andrew Ng Just Released OpenWorker: An Open-Source, Local-First Desktop AI Coworker That Returns Finished Deliverables Instead of Chat
Andrew Ng introduces OpenWorker, an open-source desktop AI agent that generates finished deliverables rather than conversational responses. The system employs a four-layer architecture (desktop shell, local agent server, capability/connector layer, model router) running entirely on-device, with Python (FastAPI) and TypeScript (React/Tauri) components. It supports 30 curated LLMs via BYO-key or local inference (Ollama), implements a typed risk engine (read/write_local/exec/external) with five permission modes, and enforces strict privacy via local-first design. The 32.4k-line codebase includes vetted tools for file operations, git, shell access, and app integrations.
desktop agentlocal-firsttyped risk enginetool-calling modelspermission modes
You Didn’t Get the AI Model You Paid For
The article identifies three distinct mechanisms by which API-served AI models may diverge from their specified identifiers: substitution (dynamic routing to alternate architectures like Anthropic's Fable 5→Opus 4.8 fallback), degradation (quantized weight serving), and drift (silent weight updates). Through legal analysis of UCC §2-313 and evidentiary rules like FRE 901(b)(9), it demonstrates how current routing practices create contractual ambiguities and authentication gaps in evidentiary chains. The proposed solution involves cryptographically signed attestations of model identity tuples (weights hash, precision, system-prompt hash) to enable verifiable compliance.
model routingquantized weightscontractual complianceevidentiary authenticationapi substitution
Best Open Speech Recognition (ASR) Models in 2026: WER, Languages, Latency, and License Compared
The 2026 open ASR landscape features tightly clustered performance among top models, with Cohere Transcribe (2B, Apache 2.0) leading at 5.42% WER across 8 English datasets, though benchmark inconsistencies complicate direct comparisons. Architectural diversity emerges: non-autoregressive Granite Speech 4.1 2B-NAR achieves RTFx ~1820, Voxtral Mini 4B enables unbounded streaming via sliding-window attention, and diffusion-gemma-asr-small explores parallel denoising. Key tradeoffs involve license constraints (Apache 2.0 vs CC-BY-4.0), language coverage (Omnilingual ASR handles 1,600+ languages), and throughput (Parakeet TDT 0.6B v3 reaches RTFx 3332.74). Practical selection requires domain-specific WER evaluation and cost analysis beyond leaderboard metrics.
word error ratenon-autoregressivesliding-window attentionzero-shot in-context learningforced-alignment
Meet Gigatoken: A Rust BPE Tokenizer that Encodes Text at 24.53 GB/s, up to 989x Faster than HuggingFace Tokenizers
Gigatoken introduces a high-performance Rust-based BPE tokenizer achieving 24.53 GB/s on GPT-2 text encoding, outperforming HuggingFace tokenizers by 989x and tiktoken by 681x on a 144-core AMD EPYC. The speedup stems from hand-written SWAR pretokenization (1,049 MiB/s via dual-cursor ILP) and pretoken caching, avoiding regex-based approaches. Benchmarks across 23 tokenizer families (e.g., Llama, Gemma) show consistent gains (7–22x for SentencePiece). Compatibility mode maintains HuggingFace parity at 200–300x speed. Independent verification confirmed scalability (83.4x on 4-vCPU Xeon).
bpe tokenizerswarpretokenizationsimdrust
Anthropic Releases Claude Security Plugin for Claude Code in Beta: A Multi-Agent Vulnerability Scanner That Runs in Your Terminal
Anthropic introduces the Claude Security plugin for Claude Code, a multi-agent vulnerability scanner operating within terminal sessions. The plugin executes a six-phase dynamic workflow (inventory, threat modeling, research, sweep, panel, adversarial verification) using tiered LLM agents (Opus orchestrator, Sonnet subagents) to analyze codebases or diffs. Findings require 2/3 consensus from independent verifiers (REACHABILITY, IMPACT, DEFENSES) before inclusion in reports, with generated patches undergoing three-criteria validation in scratch clones. The tool supports four effort tiers (low to max) and produces timestamped reports without modifying the working directory.
multi-agent vulnerability scannerdynamic workflowadversarial verificationscratch clone validationtiered llm agents
Cursor Releases Cursor Router: A Request-Level Classifier Delivering Frontier Coding Quality at 30–50% Lower Cost
Cursor Router introduces a request-level classifier that dynamically routes coding assistance tasks to optimal LLMs, achieving frontier-quality output at reduced cost. The system analyzes query, context, task complexity, and domain to dispatch requests, trained on 600k+ live requests with user satisfaction (AFC) as the reward signal. Online A/B tests demonstrate 60% cost savings versus frontier models while maintaining quality, with enterprise accounts reporting 30-50% savings. Cost per commit metrics show $4.63 (Balance mode) versus $7.34 for Opus 4.8. The router incorporates cache-miss costs during model switching and requires Grok 4.5 as a routing component.
request-level classifierin-context routingcache-aware trainingcost per commitmodel churn
Research-Grade EdgeBench Analysis: AI Agent Benchmarking, Leaderboard Analytics, Scaling Laws, and Evaluation Metrics
The study presents a comprehensive analytical workflow for EdgeBench, a benchmark evaluating AI agents across diverse tasks and time budgets. Methodologically, it parses task specifications, extracts leaderboard data, and fits log-sigmoid scaling curves to model performance trajectories. Results show category-level score improvements (e.g., 2h→12h uplift analysis) and demonstrate how SForge rescale functions normalize raw evaluation outputs, with Claude Opus 4.8 achieving superior scaling (R²=0.XXX) among 5 tested models.
edgebenchscaling lawssforgelog-sigmoidleaderboard analytics
Unsloth vs Axolotl vs TRL vs LLaMA-Factory: A Fine-Tuning Framework Comparison on Speed, VRAM, and Multi-GPU
This systematic comparison evaluates four leading LLM fine-tuning frameworks—Unsloth, Axolotl, TRL, and LLaMA-Factory—across speed, VRAM efficiency, and multi-GPU scalability. Unsloth achieves 2-7.3x single-GPU speedups via hand-optimized Triton kernels and gradient checkpointing, supporting up to 342K context length on 80GB GPUs. Axolotl offers composable parallelism strategies (FSDP2, TP, CP) with 1.45x speedup and 30% VRAM reduction. TRL provides foundational trainer APIs and sequence-splitting backends (Ring Attention, ALST), while LLaMA-Factory enables zero-code operation for 100+ models. Multi-GPU benchmarks show Axolotl's FSDP2+TP maintains 86.8% throughput efficiency at 2 GPUs, dropping to 15.2% at 8 GPUs.
triton kernelsfsdp2sequence parallelismqloragradient checkpointing
Cisco Foundation AI Releases Antares: 350M and 1B Open-Weight Models That Localize Known Vulnerabilities Inside Real Codebases
Cisco Foundation AI introduces Antares, a family of small language models (350M, 1B, 3B parameters) for vulnerability localization in codebases, initialized from IBM Granite 4.0 checkpoints. The models employ grouped-query attention, SwiGLU MLPs, and RoPE, fine-tuned via supervised learning (71.5% cybersecurity reasoning data) and GRPO optimization. On the VLoc Bench (500 tasks across 290 repositories), Antares-1B achieves 0.209 File F1, outperforming GLM-5.2 (753B) and matching GPT-5.5 (0.229). GRPO reduces run-to-run variance by 42-65%, with deployment costs under $1 per 500-task sweep on an H100.
vulnerability localizationgrouped-query attentionsupervised fine-tuningfile f1grpo optimization
Meta, Microsoft, Nvidia, IBM, and others back open-weight AI
A coalition of 24 organizations, including Meta, Microsoft, Nvidia, and IBM, published an open letter advocating for open-weight AI models, contrasting them with closed proprietary systems. The letter argues that open weights democratize AI access, foster competition, and enhance security through transparency, drawing parallels to open-source software. It specifically defends distillation as a legitimate technique while opposing unlawful extraction from closed models. The document positions open-weight models as critical for broad AI adoption but does not propose specific legislation.
open-weight modelsdistillationvendor lock-inred-team exercisescompute access
OpenAI pushes ChatGPT into patient health records
OpenAI has integrated a Health feature into ChatGPT, enabling users aged 18+ to connect Apple Health data and US medical records for contextual responses across conversations. The redesign, informed by early testing showing 70% of health queries occurred outside dedicated health sections, leverages GPT-5.5 Instant and GPT-5.6 Sol models, with the latter outperforming predecessors on internal HealthBench Professional evaluations. Physician testing confirmed improvements in accuracy, safety, and context awareness, though data quality limitations persist. The feature excludes health data from model training and ad targeting, with strict permission controls and 30-day deletion policies.
chatgpthealth data integrationgpt-5.6 solhealthbench professionalin-context learning
OpenAI Presence sells enterprise AI agents with engineers attached
OpenAI Presence introduces a managed enterprise AI agent service delivered via Forward Deployed Engineers and global integrators, addressing integration and governance challenges in production deployments. The service follows a six-stage process including security review, simulation testing, and post-launch iteration, with agents tailored to specific workflows like billing disputes or IT requests. Early adopters include BBVA and SoftBank, though scalability remains unproven; OpenAI reports a 75% resolution rate for its internal support line using the same framework.
forward deployed engineersenterprise ai agentscodex improvement loopproduction-ready agentscontact-centre integration
Nvidia bets physical AI can solve healthcare robotics’ data problem
Nvidia introduces Medical Physics Simulation, an open-source framework combining classical physics simulation and generative AI (Cosmos-H Dreams) to train healthcare robots through synthetic embodied experience. The approach addresses data scarcity by generating rare surgical scenarios (e.g., calcified vessels, atypical kidney stones) in parallel GPU-accelerated environments, reducing training time from 5 hours to 2 minutes for 8,192 parallel simulations. Early adopters like CMR Surgical and Johnson & Johnson MedTech use it for soft-tissue modeling and digital twins, though clinical validation remains pending. The framework emphasizes transparency for regulatory compliance.
physical aimedical physics simulationcosmos-h dreamsembodied experienceparallel training environments
AMD to invest up to $5 billion in Anthropic under AI infrastructure deal
AMD will invest up to $5 billion in Anthropic as part of an AI infrastructure agreement, contingent on Anthropic deploying up to two gigawatts of capacity using AMD’s Instinct MI450-series accelerators by 2027. The deal includes AMD Helios systems featuring MI455X GPUs, EPYC processors, and ROCm software, with deployment milestones tied to AMD’s investment. This follows similar AMD agreements with OpenAI and Meta, aiming to establish AMD’s rack-scale systems for large AI model development. Anthropic will diversify its compute network, integrating AMD hardware alongside existing Nvidia, Amazon, and Google infrastructure.
instinct mi450helios systemsrack-scalerocam softwaregigawatt-scale
SenseTime’s Galaxy Project targets domestic AI chip scale-up
SenseTime launched the Galaxy Project to scale domestic AI chip infrastructure in China, partnering with 20+ firms including Cambricon and Huawei Ascend. The initiative combines chip-level technology, ecosystem partnerships, and commercial deployment, targeting 10 trillion tokens/day by Q4 2026. Claims include 85-152% improved Model FLOPs Utilization on domestic chips, 2.5x token output at equivalent cost, and 80% higher token output per electricity unit, though metrics lack third-party validation. The project spans optical, quantum, and space computing (e.g., satellite constellation by 2030) and includes research collaborations with Shanghai AI Lab for AI4Science applications.
domestic ai chipsmodel flops utilizationheterogeneous hybrid inferencetokens per wattai4science
Google’s Gemini 3.6 Flash targets enterprise agent token costs
Google introduced Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and Gemini 3.5 Flash Cyber to optimize enterprise AI agent workflows by reducing token costs and latency. Gemini 3.6 Flash achieves a 17% reduction in output tokens compared to its predecessor, with a 49% success rate on DeepSWE and 63.9% on MLE Bench. Gemini 3.5 Flash-Lite processes 350 output tokens per second, targeting high-volume tasks with a 72.2% success rate on GDM-MRCR v2. Gemini 3.5 Flash Cyber focuses on code vulnerability remediation, restricted to government and vetted partners. Pricing ranges from $0.3/1M input tokens for Flash-Lite to $7.5/1M output tokens for Flash.
token costslatencyvulnerability remediationoutput tokensenterprise ai
The AI Slot Machine Effect: Why Generative Feeds Disrupt Deep Work And How to Reclaim Focus
The article analyzes how generative AI interfaces disrupt deep work by employing variable reward loops akin to slot machines, optimizing for user engagement over task completion. Drawing on 2026 industry data (88% adoption rate) and cognitive studies, it demonstrates that iterative refinement cycles in tools like Anthropic's enterprise systems create cognitive drag, reducing productivity in judgment-heavy tasks by 14-26%. The proposed mitigation involves deliberate attention management strategies, including batched usage and session limits, to preserve focused work periods.
variable reward loopcognitive draggenerative interfacesattention managementiterative refinement
Bristol Myers Squibb buys Nvidia AI system for drug discovery
Bristol Myers Squibb (BMS) has acquired an Nvidia DGX SuperPOD based on the Vera Rubin architecture to enhance AI-driven drug discovery. The system comprises eight DGX Vera Rubin NVL72 racks, combining Vera CPUs and Rubin GPUs, and will train proprietary models for target identification, lead optimization, and large-molecule predictions. BMS reports a 20-30% reduction in drug candidate identification time, with potential for 50%, and plans to integrate the new infrastructure with existing systems via Nvidia Mission Control for global research collaboration.
dgx superpodvera rubin architecturedrug discoverylarge-molecule predictionsbio-nemo toolkit
Chinese open-weight models are cheap. Washington is deciding what that costs.
The release of Moonshot AI's Kimi K3, the largest open-weight model to date, has reignited policy debates in Washington regarding Chinese AI models. Dean W. Ball of OpenAI suggested regulatory uncertainty could deter enterprise adoption, while critics argued against weaponizing regulation for competitive advantage. Commercial pressures are evident, with open-weight models handling 29% of tokens via Vercel's gateway in June yet accounting for under 4% of spending. Security concerns focus on the irreversibility of open weights and potential backdoors, though counterarguments emphasize proportionality. The U.S. may employ procurement rules and Entity List threats rather than outright bans, indirectly affecting global cloud providers.
open-weight modelsregulatory uncertaintyentity listinference savingshyperscalers
Generated automatically at 2026-07-24 20:49 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.