Daily Digest — 2026-08-21
249 items · 3 research labs, 236 arxiv papers, 10 industry media
🏛️ Research Labs (3)
How ChatGPT Work helps Stampli move ideas to market
Stampli leveraged OpenAI's Codex and ChatGPT Work to accelerate product launches and content creation by automating knowledge integration and asset generation. The team used these tools to consolidate product context, meeting notes, and messaging guidelines into a shared system, reducing production time by 3.16x (from 243 to 77 hours) for the Deep Finance™ launch. Codex handled 90% of animation work and enabled rapid data retrieval during executive meetings. Weekly content output increased 10x, with hundreds of pieces generated weekly. The tools also facilitated brainstorming, stakeholder persona creation, and strategic advising, compressing prototype-to-launch timelines from months to six weeks.
codexchatgpt workdeep financeprototype-to-launchknowledge integration
Offering Zero Data Retention for frontier models
OpenAI introduces Zero Data Retention (ZDR) for frontier models, ensuring that customer prompts and responses are not retained post-processing, with enterprise data excluded from model training unless explicitly opted-in. To address safety risks across multiple interactions, Private Safety Processing is previewed, enabling pattern identification without exposing content to OpenAI personnel. This system operates on both customer-controlled and OpenAI-provided encrypted infrastructure, returning limited safety signals for enforcement. Early testing with customers is underway, with a technical white paper and broader rollout planned for September, emphasizing customer control and regulatory compliance.
zero data retentionprivate safety processingfrontier modelsencrypted infrastructuresafety signals
Up to 3.2x Faster Inference with LFM2.5-DSpark
LFM2.5-DSpark introduces a speculative decoding framework that accelerates inference for large language models by up to 3.2x on GPU and 2.87x on-device. The method combines a DFlash-style parallel backbone, a lightweight sequential head modeled as a Markov chain, and a confidence-scheduled verifier to optimize token acceptance rates. Evaluated on LFM2.5-2.6B, the framework achieves a 57% reduction in function-calling latency and maintains quality parity under greedy decoding. Implementation supports llama.cpp and SGLang, with draft models (~300M parameters) trained on diverse datasets. Results demonstrate significant throughput improvements on H100 GPUs and M4 Max MacBook Pro devices.
speculative decodingdflashmarkov chainllama.cppsglang
📜 arXiv Papers (236)
SPADE: Self-Play in Adaptive Synthetic Executable Environments
SPADE introduces a self-play RL framework where a single LLM acts as both Environment Designer and Reasoning Agent, generating adaptive training environments as executable code with Gym-style interfaces. The method estimates agent regret via reward gaps with/without hints, optimizing for capability-edge environments while maintaining feasibility. Key innovations include document-grounded environment design and accumulated memory. Scaling to 30B parameters, SPADE outperforms fixed-environment baselines by +5.3 average on eight benchmarks, with +5.7 on BFCL-v4 and +13.9 on ACEBench-Agent for tool use, demonstrating scale-dependent gains in game settings.
self-playadaptive environmentsexecutable coderegret estimationopenai gym
ADEPT: Accelerating Dexterity via Pre-Training and Post-Training using Reinforcement Learning
The paper presents ADEPT, a reinforcement learning framework for sim-to-real transfer of dexterous manipulation skills across high-DoF robots. The method combines pre-training on a generic object reposing task with a post-training recipe involving behavior-cloning distillation, critic warm-up, and conservative on-policy updates to maintain transferability. Experiments demonstrate zero-shot sim-to-real transfer on 23-DoF Kuka-Allegro and 29-DoF Flexiv-Sharpa robots, achieving human-level speed in long-horizon tasks from raw visuo-tactile perception.
sim-to-real transferdexterous manipulationreinforcement learningvisuo-tactile perceptionhigh-dof robots
Beyond Teacher Likelihood: Group-Calibrated On-Policy Distillation for Long-Context Reasoning
Group-Calibrated On-Policy Distillation (GC-OPD) enhances long-context reasoning by addressing teacher-verifier disagreement in on-policy distillation. GC-OPD normalizes verifier rewards and trajectory-level OPD scores within rollout groups, using their difference as a signed residual. Relative-advantage-based credit assignment (RACA) distributes this residual across tokens based on their relative OPD advantages. Evaluated on five long-context benchmarks, GC-OPD improves Qwen3-4B and Qwen3-8B averages from 29.08 to 40.47 and 35.12 to 44.65, respectively, outperforming vanilla OPD (39.31 and 43.56). Ablations confirm the efficacy of signed residuals and RACA over uniform token allocation.
on-policy distillationlong-context reasoningverifier rewardsrelative-advantage-based credit assignmenttrajectory-level normalization
Finetuning Strategies for Querying Sounds by Vocal Imitation
The authors present the winning submission to the AES AIMLA 2025 Challenge on sound-effect retrieval by vocal imitation, detailing two fine-tuning approaches. First, contrastive learning employs a frozen, pretrained CED encoder; second, joint contrastive-triplet learning with semi-hard negatives utilizes a MobileNetV3 encoder. The report, updated post-challenge, provides implementation specifics previously undisclosed.
contrastive learningvocal imitationsound retrievalfine-tuningsemi-hard negatives
Interpretable AI predicts a 2026 summer dry anomaly in central China
A deep learning framework translates dynamical atmospheric circulation predictions into interpretable precipitation forecasts, identifying a summer 2026 dry anomaly in central China. The model employs layer-wise relevance propagation (LRP) to attribute predictions to specific circulation features, notably northerly winds driven by central equatorial Pacific warming. Retrospective validation shows higher skill in analogue years with persistent warming, where perturbation tests confirm the mechanistic link (removing LRP-identified features eliminates 100% of the predicted anomaly).
deep learninglayer-wise relevance propagationprecipitation anomalyatmospheric circulationperturbation test
Beyond the Transcript: Detecting Covert Co ordination in Latent Multi-Agent Communication
We introduce Verifiable Latent Alignments (VLA), a framework for detecting and mitigating covert coordination in multi-agent systems through latent communication channels. VLA combines a three-layer monitor for representation anomaly detection, counterfactual action-distribution influence, and sparse-autoencoder interpretation, alongside black-box behavioral instructions and white-box matched-neutral counterfactuals for steerability. Evaluated on a multi-agent auction benchmark, the sequential monitor achieves mean AUROC scores of 0.993 for homogeneous agents and 0.854 for heterogeneous pairs. In Qwen3-0.6B auctions with 25-100 bidders, white-box steering achieves 100% bid-distribution recovery and reduces collusive low-bid behavior by 47.3 percentage points, demonstrating effective monitoring and mitigation without attack-specific training.
latent communicationcounterfactual analysissparse-autoencodermulti-agent systemsverifiable latent alignments
Pre-Compiled Pipeline Shards for Distributed LLM Inference on Intel AI PC Fleets
The paper introduces a distributed LLM inference system leveraging idle Intel AI PCs with integrated GPUs/NPUs (16+GB unified memory) to serve models exceeding single-node capacity. Using pipeline parallelism, models are split into pre-compiled OpenVINO graph shards per stage, optimized via IndirectKVCache fusion and speculative decoding. A two-node Llama 3.1 8B INT4 pipeline achieves 1.79x single-user throughput with two concurrent users, while a four-node Lunar Lake deployment runs 70B models interactively with token-identical outputs to non-speculative decoding.
pipeline parallelismopenvinoindirectkvcachespeculative decodingmicro-batching
Grouping the Stochastic Machine: Precision, Not Capability, as the Frontier Metric for AI Systems
The article proposes precision, not capability, as the key metric for comparing frontier AI systems, arguing that current benchmarks focus erroneously on central tendency rather than output consistency. It introduces a method for measuring precision by repeatedly executing fixed tasks at constant temperature and computing outcome consistency, requiring no human grading. Results demonstrate that precision distinguishes correctable systematic errors from inherent model limitations, with empirical validation showing rule-based improvements can eliminate certain inconsistencies while constructed tasks fail to reveal practical discipline gaps.
precision metricfrontier aioutput consistencybenchmark designstochastic systems
Leaf Values as Coordinates: Exact Contrastive Explanation for Gradient-Boosted Ensembles
The paper introduces an exact contrastive explanation method for gradient-boosted ensembles by interpreting leaf values as coordinates in R^M, enabling linear model operation on instances. This approach leverages inherent additivity in the ensemble's structure, eliminating the need for fitting, sampling, or feature additivity assumptions. A recourse method based on this representation was evaluated on five tabular datasets under repeated cross-validation, achieving near-perfect reconstruction accuracy (6.2 x 10^-15). The method demonstrated Pareto-non-dominance on effort vs. realism in credit datasets and retained 58% validity under practical constraints, outperforming baselines by 17%.
gradient-boosted ensemblescontrastive explanationleaf valuesrecourse methodtabular datasets
Tuning the Stochastic Machine: A Systems Engineer's Operating Model for Human-AI Engineering
The article proposes a systems engineering framework for persistent error correction in LLM assistants, addressing the operational gap where corrections are typically session-bound. Drawing parallels between traditional machine operations (firmware, configuration) and stochastic LLM behavior, the author identifies seven principles for governing corrections, including versioning, recurrence monitoring, and retirement of stale rules. Three practical cases demonstrate the framework's application, revealing instances where controls inadvertently caused the harm they aimed to prevent. The work concludes with a proposed measurement framework and lab study for validation.
llm operationserror correctionsystems engineeringstochastic generationversioning
PGFS++: Molecular Property Improvement under Synthesis and Diversity Constraints
The paper introduces PGFS++, a synthesis-aware reinforcement learning framework for molecular property improvement under synthesis and diversity constraints. Building on PGFS+, which uses trainable embedding lookup tables for reaction templates and reactants, PGFS++ addresses reward-hacking by maintaining input-specific trajectories that preserve structural similarity. The method combines learned reaction templates with compatible building blocks to produce molecules with improved target properties while ensuring synthetic feasibility and output diversity. Experiments demonstrate PGFS++'s effectiveness in property improvement tasks without sacrificing molecular diversity.
molecular property improvementsynthesis-aware rlreward-hackingforward-synthesisstructural similarity
Discretizing Continuous Time Series for Imputation with Masked Diffusion Training
The paper introduces MDTIM, a Masked Diffusion Time-series Imputation Model that structurally separates missing and observed values using MASK tokens and directly predicts original signals. The method employs Stochastic Discretization to map continuous time series to ordinal-aware tokens while preserving dynamics. Experiments on diverse benchmarks show MDTIM outperforms state-of-the-art deterministic and generative baselines in robustness and scalability across missing-data scenarios.
time series imputationmasked diffusion modelstochastic discretizationordinal-aware tokensgenerative baselines
Open-MOPD: Diagnosing and Fixing Capability Imbalance in Multi-Teacher On-Policy Distillation
The paper introduces Open-MOPD, a framework addressing capability imbalance in multi-teacher on-policy distillation (M-OPD) for reinforcement learning. Through controlled experiments on SmolLM3-3B-Base with oracle routing, the authors identify three key factors causing optimization budget misallocation: sequence-length disparities, dynamic convergence drift, and reward staleness. Open-MOPD combines token-share balancing, gap-aware budget allocation, and student reward refresh, improving headroom recovery from 35.6% to 83.4% compared to standard M-OPD. The work provides fully open-sourced training recipes and evaluation suites for reproducible research.
multi-teacher distillationon-policy learningcapability imbalancereward stalenessoptimization budget
Detecting Backdoors in Object Detection via Pre-NMS Prediction Distribution Shift
DistScan introduces a novel backdoor detection framework for object detection models by leveraging pre-NMS prediction distribution shifts, addressing limitations of existing methods that fail in scene-level attacks. The method aggregates intermediate class predictions over a clean validation set and identifies backdoors by detecting significant deviations from training class frequencies, requiring no model weight access, trigger knowledge, or additional training. Evaluated on MS-COCO and PASCAL VOC across two architectures and three scene-level attack scenarios, DistScan improves average detection accuracy by 27.32 percentage points over the best-performing baseline.
backdoor detectionobject detectionpre-nms predictionscene-level attackstraining class frequencies
DA-WAM: Decision-Aligned Future Latents for Driving World Models
DA-WAM introduces a unified framework for decision-aligned future modeling in autonomous driving, addressing the gap between predictive representation learning and planning optimization. The method integrates an online encoder, a stable momentum target, and an action-conditioned predictor to generate distinct future latent states per trajectory candidate, evaluated by a future-latent-conditioned factorized scorer. Predictive supervision is maintained throughout planner optimization, with additional supervision from safety-critical hard negatives near planning boundaries. Experiments on NAVSIM-v1 and NAVSIM-v2 demonstrate state-of-the-art performance, validated through ablations and diagnostic analyses.
decision-aligned modelingfuture latent statesaction-conditioned predictortrajectory scoringautonomous driving
ReWEIGH the Evidence: Calibrating Token-Level Ordinal Visual Evidence to Mitigate Hallucinations in Large Vision-Language Models
ReWEIGH introduces a training-free decoding intervention to mitigate hallucinations in large vision-language models (LVLMs) by calibrating token-level ordinal visual evidence. The method aggregates vocabulary ranks across visual positions and compares each candidate with a token-specific reference estimated from unlabeled images, applying a bounded penalty during inference. Evaluated on four 7B backbones, ReWEIGH reduces hallucinated object mentions by up to 21.3% while maintaining or improving descriptive and general performance, with an average added latency of 1.33% per token. The approach scales across six architecture families up to 32B parameters.
hallucinationstoken-levelvocabulary ranksvisual-token statesdecoding intervention
Robust Risk Under Evolving Uncertainty: A Wasserstein Counterpart of the Entropic Value-at-Risk
The paper introduces Wasserstein entropic value-at-risk (WEVaR), a coherent risk measure using optimal-transport balls instead of relative-entropy balls to account for reachable catastrophes ignored by entropic value-at-risk. The method derives a variational dual mirroring the entropic formula, with transport price replacing inverse temperature, and positions WEVaR in the risk hierarchy. Numerical verification confirms both dualities. A closed-form robust dynamic-programming operator, driven by belief entropy, demonstrates contracting caution as belief sharpens, with certified safety guarantees and a sharp safety switch.
wasserstein entropic value-at-riskcoherent risk measureoptimal-transport balldynamic-programming operatorsafety sandwich
What is Missing from AI Post-Training AI: An Empirical Analysis
The study identifies a critical gap in AI post-training agents: while they demonstrate execution-level capability for local optimization within fixed training strategies, they lack strategy-level capability to adapt high-level approaches based on accumulating evidence. Through analysis of public post-training trajectories and three intervention experiments (experience scaffolding, human guidance, and increased inference compute), the authors show execution improvements (+12.6 GSM8K, +40.8 HumanEval) but persistent strategy rigidity. Results indicate current agents require explicit mechanisms for dynamic strategy reevaluation during execution.
post-trainingexecution-level capabilitystrategy-level capabilitylocal adjustmenttraining trajectory
GS-VLA: Plug-and-Play Viewpoint Canonicalization for Frozen VLA Policies via Gaussian Splatting
GS-VLA introduces a lightweight, plug-and-play framework for viewpoint canonicalization in Vision-Language-Action (VLA) policies without retraining, addressing robustness to camera shifts. The method reformulates viewpoint shifts as a localized novel-view synthesis problem, leveraging a 4M-parameter 3D-Gaussian canonicalizer prepended to frozen VLA policies. This approach improves performance across diverse policy architectures, unseen task suites, and perturbation scales, recovering significant performance lost under viewpoint shifts. Experiments on the LIBERO benchmark demonstrate success rate improvements from 10% to near 90% in worst-case scenarios, showcasing effectiveness without policy weight modifications.
viewpoint canonicalizationnovel-view synthesisgaussian splattingvla policieslocality assumption
Eureka: Task-Conditioned Meta-Agent Orchestration for Scientific Discovery
Eureka introduces a task-conditioned Meta-Agent architecture for scientific discovery, compiling long-horizon tasks into dynamic obligation graphs with explicit acceptance semantics. It employs receding-horizon planning, architecture promotion, and minimal-sufficient compilation to form specialized Macro-Agents, updating local architectures via cost-benefit-gated evolution when bottlenecks recur. Eureka successfully completes 170/170 recursive tasks, generates 3,948 certificates with no false acceptances, compresses median input from 9,490 to 4,005 tokens, and avoids 65.38% recomputation across 12,000 tasks. It instantiates both a Theory-Discovery Agent, yielding structural results in quantum-process and spacetime theory, and a Math/Conjecture Agent, advancing a positivity certificate for Suzuki's localized Weil quadratic form to 0 < a <= 69/200.
meta-agentobligation graphsreceding-horizon planningmacro-agentscost-benefit-gated evolution
Bernstein-Vazirani Networks: Quantum Machine Learning by Interference
The authors propose Bernstein-Vazirani Networks (BVNs), a non-variational quantum machine learning framework utilizing quantum interference for supervised learning. BVNs employ quantum Fourier sampling to extract globally informative features from superpositioned labeled data, with generalized variants enabling interference in problem-adapted bases for enhanced expressivity. The gradient-free framework achieves universal function approximation via (over)complete interference bases, demonstrating competitive performance on synthetic/real-world classification and implicit image representation tasks compared to classical and quantum baselines.
quantum machine learningquantum interferencefourier samplinguniversal function approximationgradient-free training
Counterfactual Contrastive Analysis
We propose a classifier-free method for Visual Counterfactual Explanations (VCEs) using Contrastive Analysis (CA) to generate model-agnostic counterfactual images. Our approach disentangles common and salient generative factors between two datasets, leveraging StyleGAN2's feature space F for improved detail preservation. Unlike conventional CA, we introduce a framework supporting multiple salient factors per dataset. Evaluated on three medical imaging datasets, our method demonstrates superior counterfactual generation quality compared to existing approaches, reducing sensitivity to classifier biases.
visual counterfactual explanationscontrastive analysisstylegan2generative factorsmodel-agnostic
Adaptive Memory and Reflection Multi-Agent System for Medical Question Answering
This work introduces the Adaptive Memory and Reflection (AMR) multi-agent system for medical question answering, addressing limitations of single-agent architectures and static retrieval. The system employs specialized agents with dedicated memory and reflection-based feedback to retrieve prior cases and enhance reasoning. Questions are routed through solo, collaborative, or escalated workflows based on complexity, supported by consensus and ethical overseer modules for reasoning consolidation and output review. Evaluations on MedQA and MedMCQA demonstrate superior performance compared to baselines, with ablation studies confirming the effectiveness of combining agent-specific memory, reflection, and external retrieval. The source code is publicly available.
multi-agent systemadaptive memoryreflection-based feedbackcomplexity assessmentethical overseer
Self-prompting and cross-model consensus enable reproducible data extraction from scientific literature with large language models
This work introduces a reproducible framework for extracting contextualized scientific data using large language models (LLMs), combining self-prompting, cross-model consensus, and human oversight. Four escalating workflows were evaluated: 1) expert-curated prompts with LLMs, 2) LLM-authored prompts, 3) autonomous literature discovery, and 4) dataset creation from published guidelines. Results show that frontier LLMs perform well with expert prompts but struggle with scientific nuance; self-authored prompts are nearly as effective; autonomous discovery suffers from hallucinations; and LLM-generated datasets align closely with human judgments but require human-in-the-loop validation. The framework establishes an auditable division of labor for scalable scientific data curation.
large language modelsself-promptingcross-model consensusdata extractionhuman-in-the-loop
One-Stage Object Detectors in Autonomous Driving
This survey systematically reviews one-stage object detectors for autonomous driving, analyzing architectures including YOLOv1, SSD, RetinaNet, EfficientDet, FCOS, CenterNet, and YOLOv10. It examines design choices, feature-fusion strategies, loss functions, deployment trade-offs, and benchmark performance across these models. The paper also evaluates autonomous-driving datasets, metrics, and identifies open challenges in achieving reliable real-world performance. Results indicate that while one-stage detectors effectively balance speed, accuracy, and efficiency, a significant gap persists between benchmark metrics and dependable autonomous-driving applications.
one-stage detectorsfeature-fusionloss functionsbenchmark performanceautonomous-driving
Harness Continual Learning: Continual Adaptation Beyond Model Parameters
We introduce Harness Continual Learning (HCL), a paradigm where agents evolve through harness components (Task Interface, Experience Memory, Capability Map, Adaptive Router) while keeping foundation models frozen, addressing harness-level forgetting. The method employs guarded harness evolution, separating update generation (Continual Optimizer) from state commitment (Continual Evaluator) based on improvement, retention, and validity criteria. Experiments across textual reasoning, multimodal perception, and open-world interaction demonstrate >10% relative gains over baselines, with component ablations and retention sweeps quantifying harness-level forgetting and stability--plasticity trade-offs.
harness continual learningfoundation modelharness-level forgettingguarded harness evolutionstability-plasticity trade-off
From Threat Intelligence to Detection: Knowledge-driven Enrichment and Template-based Rule Grounding for Automated Sigma Rule Generation
AUTOSIGMA introduces an automated system for converting unstructured cyber threat intelligence (CTI) into Sigma rules, addressing scalability and adaptability challenges in threat detection. The method combines knowledge-driven enrichment, template-based rule grounding, and a multi-stage validation process using an LLM-as-a-Judge mechanism. Evaluations demonstrate superior performance in rule validity, relevancy, MITRE ATT&CK technique coverage, and robustness to input quality compared to alternative solutions and standalone LLMs.
sigma rulescyber threat intelligencemitre att&ckllm-as-a-judgetemplate-based grounding
A Theory of Post-hoc Debate Judgement
The paper introduces a theoretical framework for post-hoc debate judgement in AI systems, focusing on reproducibility, robustness, groundedness, and explainability. It evaluates two debate judgement methods: LLMs as judges and formal semantics from computational argumentation. Experimental results show comparable accuracy between the methods, but formal semantics provide stronger theoretical guarantees. The study concludes that argumentation semantics are preferable for principled debate-driven AI systems.
debate judgementllmsformal semanticscomputational argumentationclaim verification
GrabVG: Graph-Attentive Binding for Visual Grounding in UAV Imagery
The authors propose GrabVG, a novel visual grounding framework for UAV imagery that addresses challenges of visual redundancy and topological ambiguity. GrabVG decomposes grounding into preattentive hypothesis search and graph-attentive feature binding, leveraging distillation-guided proposal induction, text-aware hypothesis filtering, and graph attention for language-guided intra-instance visual cues and inter-instance topological relationships. Experiments on AerialVG and AerialSense demonstrate GrabVG's superior performance, achieving 67.31% and 80.34% Acc@0.5, outperforming baselines by 10.55 and 8.76 percentage points, respectively.
visual groundinguav imagerygraph attentionhypothesis searchtopological ambiguity
DeepWeaver: Bridging the Evidence Synthesis Gap in Open-Ended Question Answering
DeepWeaver introduces a framework for evidence synthesis in open-ended question answering, addressing limitations of retrieve-then-generate pipelines where direct generation often underuses evidence and misaligns citations. The method employs Thought Block Chains (TBCs), a structured representation that groups claims, salient information, keywords, and supporting evidence, enabling inspection of residual evidence, TBC revisions, and discovery of new claims prior to final generation. Evaluated on the LoQA benchmark and DeepResearch Bench, DeepWeaver enhances content sufficiency, citation grounding, and detail preservation across multiple LLMs, demonstrating improved citation quality and deeper insights.
evidence synthesisthought block chainsretrieve-then-generateopen-ended qacitation grounding
rEDMRec: Distilling Large Language Model Reasoning into an Editable Experience Memory for Recommendation
We propose rEDMRec, a framework that distills large language model (LLM) reasoning into a compact, editable experience memory for recommendation systems, decoupling online inference cost from reasoning depth. The method employs four typed experience channels—long-term preference, short-term context, item-perception, and counterfactual hard-negative comparisons—managed by an LLM memory controller using Add/Delete/Modify/Keep operations and refined via K-agent debate. A lightweight student LLM retrieves from this memory for ranking, avoiding repeated teacher LLM invocations. Experiments on ML-1M, Amazon Beauty, and Steam datasets show rEDMRec improves HR@1 by up to 13.3% over baselines, with short-term context consistently beneficial across capacity tiers.
experience memoryllm reasoningk-agent debaterecommendation systemsmemory controller
AlphaClifford: Efficient Clifford Synthesis and Transpilation with Model-based RL
AlphaClifford introduces a model-based Reinforcement Learning framework using Monte Carlo Tree Search to optimize Clifford circuit synthesis with H, S, and CNOT gates. By leveraging symplectic group algebra for state-space modeling, it minimizes gate counts more effectively than existing heuristics. Results show reductions in total and CNOT gates for unconstrained optimization, superior hardware-constrained transpilation, and utility as a post-synthesis optimizer in Clifford+T pipelines, demonstrating RL's efficacy for quantum compilation challenges.
clifford circuitsmodel-based rlmonte carlo tree searchquantum compilationsymplectic group
Training Chemical Plausibility-Aware Large Language Models for Single-Step Retrosynthesis
We introduce Top-K prompting for training chemical plausibility-aware large language models (LLMs) in single-step retrosynthesis, addressing the one-to-many nature of reaction prediction. Our method integrates fine-tuning with ChemCensor-based and novelty-oriented rewards on CREED-CCV-2+USPTO-XL, a dataset of ~45.6 million verified reactions. The resulting Chemistry Constraint-Consistent Language Model (C3LM) achieves state-of-the-art performance on the OOD URSA-expert-2026 benchmark. Analysis reveals that LLMs and conventional models explore complementary reaction spaces, suggesting ensemble-based systems for retrosynthesis. This establishes Top-K, plausibility-aware training as a robust direction for LLM-based synthesis planning.
top-k promptingretrosynthesischemcensorc3lmursa-expert-2026
Breaking the weakest link to evade vision language models
This work investigates the vulnerability of Vision Language Models (VLMs) to adversarial attacks targeting multimodal alignment, proposing a gradient-based method that optimizes exclusively on the vision encoder to reduce computational cost. The study examines untargeted attacks, which disrupt the model's interpretation of the original image, and targeted attacks, which force specific semantic descriptions unrelated to the image. Evaluations on Qwen2.5-VL, Granite-Vision, FastVLM, and Phi-3.5-Vision demonstrate that small, imperceptible perturbations significantly alter textual interpretations. The findings underscore the susceptibility of VLMs to adversarial manipulation and the need for enhanced robustness in multimodal AI systems.
vision language modelsadversarial attacksmultimodal alignmentgradient-based optimizationvision encoder
MedUAG: Unified Understanding and Generation for Medical Multimodal Models
The authors introduce MedUAG, a unified framework for medical multimodal understanding and generation, addressing the lack of benchmarks and validated models in the domain. They construct MedUAGCorpus, the largest medical multimodal dataset with over 6 million instances across 14 imaging modalities, and MedUAGBench, a benchmark evaluating 12 diverse generation tasks. MedUAG, an end-to-end trained unified model, demonstrates strong performance across multiple understanding and generation tasks, establishing a competitive baseline for future medical multimodal systems.
multimodal large language modelsmedical generationend-to-end trainingimaging modalitiessystematic benchmark
Graphical Design of Interpretable Architectures
The authors introduce a graphical notation for designing interpretable AI architectures, adapted from Penrose tensor notation, to address limitations in existing representations. This notation provides a global view of architectures while maintaining a one-to-one mapping with PyTorch einsum code, enabling reproducibility and insight into tensor manipulations. The method is demonstrated by describing interpretable architectures such as concept bottlenecks, sparse probes, and neural additive models, and by diagramming Steerling-8B, a frontier interpretable language model. The resulting diagram reveals global architectural insights, geometric interpretations of operations, and translates directly into 33 lines of PyTorch code.
graphical notationinterpretable architecturespenrose tensor notationpytorch einsumconcept bottlenecks
SkillForge: Self-Distilling Agents for Project-Specific Issue Resolution
SkillForge introduces a self-distillation framework for enhancing LLM-based agents' project-specific issue resolution by proactively acquiring repository knowledge. Instead of relying on historical issue-resolution signals or costly test-time exploration, SkillForge synthesizes project-specific issues by re-implementing test-covered core functionalities. It distills reusable, entity-grounded skills from resolving these synthetic issues and associates them with relevant repository entities. Experiments across open-source and closed-source models demonstrate consistent performance improvements in downstream issue resolution, validating the efficacy of proactive knowledge acquisition.
self-distillationproject-specific knowledgeentity-grounded skillsissue resolutionsynthetic issues
Test-Time Scaling in the Wild: Why Exploitation, Not Exploration, Is the Bottleneck
This work introduces the first compute-normalized comparison of five test-time scaling (TTS) families across five open-ended generation benchmarks in medicine, law, finance, general chat, and creative writing. The study decomposes TTS effectiveness into exploration and exploitation, revealing that while exploration improves steadily with compute, exploitation fails due to weak correlation (ρ_v ≈ 0.12) between reward models and true quality. Tree search exacerbates this through diversity collapse, and refinement shows limited gains. Only synthesis across candidates (Fusion) consistently outperforms single-sample baselines, recovering ~40% of available quality, demonstrating that candidate selection, not generation, is the bottleneck.
test-time scalingexplorationexploitationtree searchfusion
SMTrap: Cost-Effective DoS Attacks Against Large Reasoning Models via SMT Conflict Guidance
We introduce SMTrap, a cost-effective Denial-of-Service (DoS) attack framework targeting Large Reasoning Models (LRMs) via SMT conflict guidance, eliminating reliance on model feedback or GPU computation. The method leverages search amplification, synthesizing inference-heavy Constraint Satisfaction Problem (CSP) instances using conflict counts from Satisfiability Modulo Theories (SMT) solvers as an external signal. Empirical evaluations across seven frontier LRMs demonstrate SMTrap's state-of-the-art DoS capabilities, producing effects multiple times stronger than existing baselines. A token-usage reduction mitigation strategy is also presented to counter the threat.
denial-of-servicelarge reasoning modelsconstraint satisfaction problemsatisfiability modulo theoriessearch amplification
Learning-State-Aware Dynamic Generative Data Augmentation on Small-Scale Datasets
The paper introduces Learning-State-Aware Dynamic Generative Data Augmentation (LSADA), a method for small-scale image classification that dynamically adjusts augmentation strength per sample based on learning state (loss and loss-decrease rate) and employs a decoupled diffusion fusion strategy to balance diversity and class semantics. LSADA applies strength-controlled transformations to class-relevant regions while diversifying class-irrelevant regions. Evaluated on nine datasets, it outperforms state-of-the-art dynamic GDA by 4.5% (natural images) and 2.5% (medical images) on average.
generative data augmentationdiffusion fusionsmall-scale datasetslearning-state-awareclass semantics
\textsc{TestifAI}: Tomography-Based Testing for Deep Learning Systems
TestifAI introduces a deep learning testing framework for efficient robustness estimation against combinatorial input perturbations. The method enables users to specify operational conditions as structured spaces of semantic perturbations and severity levels, querying robustness for arbitrary combinations. TestifAI employs partial model tomography, reconstructing model behavior in multi-perturbation spaces from lower-order projections, avoiding exponential test execution. Experiments on five image and language classification tasks demonstrate TestifAI predicts higher-order (3-4 perturbations) outcomes from low-order (1-2 perturbations) observations with <7% robustness estimation error, reducing inferences by 60-80%.
partial model tomographyrobustness estimationcombinatorial perturbationssemantic input perturbationslower-order projections
Syntactic Simplification of OWL Class Expressions
We introduce Class Expression Simplifier (CES), a novel algorithm for syntactic simplification of OWL class expressions in Description Logics. CES preserves formal semantics while reducing complexity through systematic application of rewriting rules that eliminate redundancies and identify equivalent but simpler expressions. Evaluated on class expressions from two medium-sized ontologies, CES demonstrates measurable improvements in reasoning efficiency and reductions in verbosity. Implemented within the OWLAPY Python framework, CES enhances ontology-driven applications by producing more compact and human-readable representations without altering logical entailments.
owl class expressionsdescription logicssyntactic simplificationreasoning efficiencyowlapy
Training-Free Inference-Time Self-Reflection and Cost-Bounded Early Stopping for Large Language Models
EvoResearcher introduces a training-free, inference-time protocol for cost-bounded self-reflection in frozen large language models (LLMs). The method iteratively generates, self-critiques, and revises responses until a maximum depth is reached or a CONFIRMED sentinel triggers early stopping, enabling self-verification under strict compute constraints. Four meta-reward components—correctness, efficiency, reflection depth, and tool-call diversity—are implemented as prompt-level mechanisms without gradient updates. Evaluated on Big-Bench Hard (100 questions), GSM8K (500), and MATH (500), the protocol achieves early stopping in 82-88% of cases with equal accuracy, averaging 2.1 generations per question.
self-reflectionearly stoppingmeta-rewardinference-timefrozen llm
DentAgent: Evidence-Centric Multi-Agent Coordination for Multimodal Dental Reasoning
We introduce DentAgent, an evidence-centric multi-agent framework for multimodal dental reasoning that integrates domain knowledge, radiographs, intraoral photographs, and 3D dental data. The system employs an Orchestrator to coordinate five specialized agents, each converting observations into structured evidence records managed by an Evidence Blackboard. This approach standardizes evidence representation, tracks coverage, gaps, and conflicts, and integrates isolated dental capabilities into a unified workflow. DentAgent achieves leading performance across four benchmarks, surpassing senior specialists by 17.3 percentage points on multi-label diagnosis, demonstrating its potential for broadly applicable and traceable multimodal dental reasoning in population oral health assessment.
evidence-centricmulti-agentmultimodalorchestratorevidence blackboard
SkillGate: Training In-Policy Skill Selection in Long-Horizon Agents
SkillGate introduces a novel training framework for in-policy skill selection in long-horizon agents, addressing selector credit starvation—a structural issue where skill-naming tokens receive vanishing and increasingly wrong-signed credit under outcome-rewarded RL. The method partitions token support into two disjoint credit channels: outcome credit for execution tokens and action-local advantage for skill-naming tokens, ensuring positive credit only for correct skill selections. Evaluated on five agentic benchmarks with a 16-candidate slate, SkillGate improves trial success rates from 40.8% to 53.2% for a 9B policy, reduces exposure to misleading candidates by two-thirds, and decreases the number of skills read.
skill selectioncredit starvationtoken supportaction-local advantagelong-horizon agents
ORBITER: Conflict-Aware Decision-Making for Agentic Last-Mile Delivery
ORBITER introduces an agentic Order Arbiter for conflict-aware next-order decision-making in last-mile delivery, addressing LLMs' unreliability as direct predictors. It models courier service through decision points containing spatiotemporal states and visible orders, exposing local trade-offs. Fixed proposers rank candidates, while a structured report identifies ranking disagreements. The LLM employs task-specific tools to gather evidence on alternatives, with an independent critic verifying decisions against evidence. Evaluations across four cities demonstrate ORBITER's effectiveness, outperforming state-of-the-art baselines by up to 9.2% on average.
last-mile deliveryspatiotemporal dependenciesdecision pointstask-specific toolsindependent critic
Verifiable abstention makes AI leak diagnosis accountable in water distribution networks
We propose a verifiable abstention framework for accountable AI-assisted leak diagnosis in water distribution networks, addressing the gap between localization accuracy and operational trust. The method employs a physics-grounded executor agent that falsifies hypotheses against a digital twin, coupled with an independent supervisor agent using a large-language-model auditor to verify evidence against a code-verifiable contract. Evaluated under field-grade noise, the system achieves 96% decision precision on acted events, acting correctly on 4 of 33 leaks in an independent benchmark. A real-world register of 194 events yielded five excavation dispatches with three correct identifications and 44% survey recovery at full district precision, demonstrating operational viability.
verifiable abstentionleak localizationdigital twinlarge-language-modeldecision precision
MLREF: Efficient Module Reuse for Reward Design in Reinforcement Learning via Large Language Models
MLREF introduces a module-level reward evolution framework for reinforcement learning, addressing the instability of monolithic reward functions by maintaining a reusable module pool. The method optimizes the pool via reflection-based refinement, hybrid credit assignment, and merge strategies with rollback, enabling component reuse and iterative improvement. Experiments on 17 tasks demonstrate improvements of 25.2% (locomotion) and 6.6% (manipulation) over baselines, with more stable optimization dynamics.
reward function designmodule poolreinforcement learninglarge language modelscredit assignment
Understanding Multilingual Medical ASR Adaptation Through Layer-Wise Analysis
The paper analyzes how multilingual medical adaptation reshapes Whisper's encoder representations through layer-wise analysis, comparing zero-shot decoding, English-only, German-only, and combined EN+DE fine-tuning. Experiments show Whisper-Medium achieves the lowest English WER (7.72%) and combined EN+DE WER (26.30%), while German-only Whisper-Large-v3 yields the lowest German WER (44.96%) on limited training data. Layer-wise analysis reveals English medical fine-tuning dominates representation shifts, with multilingual continuation preserving adapted spaces; domain and language features remain recoverable across layers.
medical asrmultilingual adaptationlayer-wise analysiswhisper modelsword error rate
Identifying Implicit Premises for Logical Reconstruction of Argument Graphs
The paper proposes a neuro-symbolic pipeline for generating implicit premises to enable logical reconstruction of argument graphs from natural language text. The method combines large language models (LLMs) for generating intermediate implicit premises with symbolic translation into logical formulae, facilitating entailment, contradiction, or neutrality analysis between explicit premises and claims. Evaluation on the Microtext Argumentative Corpus demonstrates the approach's effectiveness in handling enthymemes.
enthymemesneuro-symboliclogical reconstructionargument graphsimplicit premises
Pairwise Logical Selection of Enthymeme Completions under Semantic-Link Uncertainty
The paper introduces Possible-World Atom-Link Formalization (PWAL), a neuro-symbolic method for pairwise logical selection of enthymeme completions under semantic-link uncertainty. PWAL extends prior work by handling both missing-premise and missing-claim selection, replacing binary entailment with logical-resistance scores and marginalizing over alternative semantic-link configurations. Evaluated on five tasks (ARCT, CDED-derived, iDebate-derived, AAE2-derived, alphaNLI), PWAL improves strict accuracy by 2.95-30.86 percentage points and reduces tie rates by 4.57-58.00 points versus Top-Link, while providing transparent traces of each comparison.
enthymeme completionlogical resistancesemantic-link uncertaintypartial maxsatneuro-symbolic
Do Large Language Models Hallucinate Electric Fata Morganas?
The paper argues that hallucinations in large language models (LLMs) have philosophical implications for machine consciousness, beyond being mere engineering flaws. It analyzes hallucination causes (source-target divergence, training-inference discrepancies, overfitting) and conducts two empirical studies: (1) GPT generations on ambiguous questions show temperature sampling affects hallucination rates (higher temperatures increase plausible-but-wrong answers), and (2) an encoder-only model trained on encyclopedic data produces factual answers without embellishment, suggesting hallucinations stem from diverse training data rather than cognitive ability. The work references Turing, Searle, and cybernetics to posit that machine consciousness may be epistemically indistinguishable from advanced hallucinations.
large language modelshallucinationsmachine consciousnesstemperature samplingencoder-only model
A strengthening of the MCFL-ness of $O_2$
The article strengthens existing proofs that the monoid $O_2$ is a multiple context-free grammar (MCFG) by refining the characterization of string tuple factorizations. Building on recent work that framed the proof in terms of factorization properties, the authors derive a more precise structural description of these factorizations. This result advances theoretical foundations applicable to computational linguistics and algebra, though specific quantitative improvements or algorithmic implications are not detailed.
multiple context-free grammarstring tuplefactorizationcomputational linguisticscomputational algebra
Forgetting, plasticity, and co-observation: a third facet of continual learning
The paper identifies data co-observation as a distinct factor in continual learning beyond catastrophic forgetting and loss of plasticity. Through controlled experiments decoupling data access constraints from stability-plasticity trade-offs, the authors demonstrate consistent performance gaps between joint and separate training in supervised and self-supervised chunking scenarios. Results show memory replay's benefits exceed mere forgetting mitigation by reintroducing co-observation advantages, while distillation-based methods primarily aid knowledge retention.
continual learningco-observationcatastrophic forgettingmemory replayknowledge retention
Decomposing Wrong-Consensus Agreement in LLM Self-Consistency: A GPT-4.1 Case Study
The paper quantitatively analyzes failure cases in LLM self-consistency through majority voting by decomposing wrong-consensus agreement into mechanical and residual components. It introduces a pluralistic agreement index Γ, normalized by reference scale d=(1-p)/(C-1), and uses difficulty-matched, leak-free resimulation to isolate per-case answer preferences. Results show GPT-4.1's mechanical preference explains 81-93% of agreement on GPQA-Diamond but only 59-78% on AIME, with residual agreement (1.56-2.80 Γ units) suggesting preference heterogeneity. Self-consistency backfire occurs on hard questions (voting gap -0.09, CI [-0.12,-0.07]), with highest-agreement bin accuracy limited to 0.42-0.83.
self-consistencypluralistic agreement indexmajority votingpreference heterogeneitymechanical null
SIDScope: A Diagnostic Resource for Semantic-ID Interfaces in Generative Recommendation
SIDScope introduces a diagnostic resource for evaluating Semantic-ID interfaces in generative recommendation systems, addressing gaps in coherence, structure, and validation. The method normalizes item-to-code artifacts, verifies provenance, profiles mapping structure, and analyzes path-to-item outcomes across nine tokenizer exports from Amazon and Yelp data. Key findings include prefix alignment's conditional impact on retrieval (1.2-3.0 percentage points gap in target path survival) and the necessity of separate handoff checks for model reuse. The package provides frozen evidence summaries, conformance reports, and CPU-only verifiers.
semantic-idgenerative recommendationtokenizerprefix alignmenttrace accounting
Beyond Predictive Fairness: Quantifying Attribution Consistency Across Demographic Groups in Diabetic Retinopathy Screening
The study introduces the Explanation Consistency Score (ECS), a fairness metric based on Jensen-Shannon divergence to assess attribution map similarity across demographic subgroups in medical imaging. Using diabetic retinopathy screening as a case study, ECS is evaluated globally and within disease severity levels. Results show high explanation consistency across ethnic groups despite predictive performance disparities, indicating that ECS and predictive fairness capture complementary aspects of model behavior.
explanation consistency scorejensen-shannon divergenceattribution mapspredictive fairnessdiabetic retinopathy
Epistemic Subordination: Generative AI and the Infrastructure of Knowledge
The article introduces 'epistemic subordination' as a structural harm in generative AI, where dominant cultural frameworks become the default knowledge infrastructure through probabilistic modeling. The authors argue that minority epistemologies are absorbed but subordinated in outputs due to training data compression, creating a unified harm across legal domains. They critique current anti-discrimination, cultural rights, and democratic pluralism laws for addressing only downstream applications, proposing instead governance at the model-training level.
epistemic subordinationgenerative aiprobabilistic modeltraining data compressionminority epistemologies
Metrics That Write Themselves: Evolving an Evaluator from Its Own Blind Spots
The paper introduces EvalCEGAR, a method for automatically evolving evaluation metrics by refining a pool of defect-detecting Python operators through counterexample-guided abstraction refinement. The approach identifies collisions (pairs of correct/incorrect answers scored identically) to iteratively improve operator specificity, avoiding prompt-based generation limitations. On MBPP+ and HumanEval+, a 55-line operator closed 15.4% of the performance gap versus a perfect filter (p=0.0010), matching hand-written operator effectiveness on 33% of flags while reducing false positives. Six of eight runs produced useful operators, outperforming a combined hand-written filter and matching LLM judge accuracy without per-candidate inference costs.
counterexample-guided abstraction refinementautomatic metric evaluationdefect-detecting operatorsin-context learningprogram verification
A Multi-Agent Platform for Automated Enterprise Analytics and Insight Generation
The paper introduces a multi-agent framework for conversational business intelligence, leveraging CrewAI to deploy five specialized agents in a sequential pipeline for query processing, data analysis, visualization generation, and insight delivery. The system incorporates a defense-in-depth security architecture and query parameterization for reusable dashboard components. Evaluated on 300 test cases, it achieves 95.3% functional accuracy, 24-second mean latency, and 4.52/5.0 response quality, with a 93.0% hallucination-free rate, outperforming single-agent baselines by 22.6 percentage points in accuracy and 20.2% in quality. Ablation studies identify Data Analysis and Report Aggregation agents as key to performance.
multi-agent frameworkconversational business intelligencedefense-in-depth securityquery parameterizationllm-as-a-judge
Flama: a Python framework for development and deployment of production-ready APIs, machine learning, and LLM services
Flama introduces an open-source Python framework for unified development of production-ready web APIs, ML services, and LLM applications. Built on ASGI, it combines seven subsystems: type-driven dependency injection, pluggable schema layers, automatic CRUD generation, portable model packaging (.flm format), multi-backend LLM serving (vLLM/MLX), Rust-accelerated core, and Model Context Protocol. The framework supports REST, WebSocket, SSE, NDJSON streaming, OpenAPI 3.2.0 generation, and offers JWT authentication, pagination, and background tasks. Comparative analysis demonstrates its capabilities against existing frameworks and inference engines.
asgidependency injectioncrud generationllm servingopenapi
A Few Cases Are All You Need: An Empirical Study of Annotation-Efficient LoRA Fine-Tuning of MedSAM3
This work demonstrates that parameter-efficient LoRA fine-tuning of MedSAM3 achieves clinically useful medical image segmentation with minimal annotation. The authors adapt MedSAM3 for five abdominal organs in CT and MRI using only 1, 2, 5, and 10 annotated cases, evaluating on AMOS22 dataset. With just 10 cases, models achieve performance competitive with specialist systems trained on orders of magnitude more data, including reliable gallbladder segmentation (Dice 0.68 CT, 0.59 MRI) where existing tools fail (Dice 0.0004). External validation on Whole Heart Segmentation shows generalization to cardiac segmentation. Training requires 3--5 hours per organ on a single GPU, 2--3 times faster than nnU-Net.
loramedsam3amos22dicennunet
Budget-First Tariff Recommendation (BFTR): A Complete Algorithmic Framework for Telecom Plan Recommendation without Overcharging
The paper introduces Budget-First Tariff Recommendation (BFTR), a complete algorithmic framework guaranteeing zero overcharging in telecom plan selection by aligning final prices with catalog references. BFTR integrates eight strategies, including novel Recursive Hybrid and Knapsack-First Hybrid approaches, with formal proofs of offer existence and zero price deviation for non-interpolated strategies. Evaluation on 974 Nigerian MTN customers shows Recursive Hybrid achieves optimal utility (0.946) with 29.9GB volume, Piecewise maximizes volume (39.7GB), and Power Law offers balanced performance (38.1GB, 99.9% budget utilization), all maintaining zero surcharge and sub-10ms execution times.
tariff recommendationovercharging guaranteehybrid strategiesutility optimizationtelecom pricing
Competence, Not Accuracy: A Diagnostic for Reference-Free Judge Gates in Skill Optimization
The paper introduces a diagnostic for assessing LLM-judge gates in text-space skill optimization, formalizing reference-free judges as latent solvers whose evaluation capacity is bounded by their problem-solving competence. It derives a closed-form bound on discriminability (ROC-AUC) based on competence $c$ and answer-space size $k$, showing a necessary condition $c > 1/k$ and demonstrating that benchmark accuracy overstates relevant competence. Empirical results reveal discriminability at chance when competence is low and usability when competence is sufficient, validated through a non-intervening probe and closed-loop study.
llm-judgeskill optimizationroc-aucdiscriminabilityreference-free evaluation
The Impact of CutMix on Reliability and Robustness in Semantic Segmentation
This work systematically evaluates CutMix's impact on semantic segmentation reliability and robustness, isolating its effects from other augmentation techniques. Using DeepLabV3+ (CNN) and SegFormer (transformer) architectures, the study measures accuracy, calibration, and uncertainty quality across in-domain and out-of-domain scenarios. Results indicate CutMix minimally affects segmentation accuracy but consistently improves model reliability, particularly under distribution shifts, by enhancing calibration and uncertainty estimation rather than raw prediction performance.
cutmixsemantic segmentationmodel calibrationdistribution shiftuncertainty estimation
A Critical Synthesis of Uncertainty Quantification and Foundation Models for Semantic Segmentation
This paper presents the first systematic evaluation of uncertainty quantification (UQ) methods for semantic segmentation foundation models, addressing interpretability and overconfidence issues. The authors fine-tune a lightweight DPT decoder on SAM2's pretrained encoder, benchmarking four UQ approaches (Monte Carlo Dropout, Deep Sub-Ensemble, Test-Time Augmentation, Evidential Deep Learning) across Cityscapes, NYUv2, and out-of-domain settings. Results reveal trade-offs between segmentation accuracy, calibration, uncertainty quality, and computational cost, highlighting gaps in joint optimization of reliability and efficiency for real-world deployment.
uncertainty quantificationsemantic segmentationfoundation modelsdomain shiftcalibration
MemFuse: Multi-Source Memory Fusion from Fragmented Observations
The paper introduces MemFuseBench, a benchmark for multi-source memory fusion, addressing the limitation of existing systems that focus on single-source textual histories. It proposes MemFuse, a structured memory system that organizes event-layer atomic memory and cluster-layer fused memory within a causal fusion graph, preserving source provenance. Experiments on MemFuseBench demonstrate MemFuse's superior performance in temporal reasoning, cross-source evidence fusion, and robustness to noise across three LLM settings.
memory fusionmulti-source benchmarkstructured memorycausal fusion graphsource provenance
Impact of Iterative Fine-Tuning on Transcription Accuracy in Complex Historical Sanskrit Manuscripts
The paper introduces a traditional OCR pipeline for digitizing complex historical Sanskrit manuscripts, addressing challenges from heterogeneous layouts and non-standard appearance through iterative fine-tuning at layout and appearance levels. The method reduces human annotation effort by adapting to target manuscript distributions, demonstrated on three manuscripts with newly released granular layout annotations in PAGE-XML format. Quantitative improvements from iterative fine-tuning are shown, alongside benchmarking of Multi-Modal Large Language Models on the introduced dataset.
ocr pipelineiterative fine-tuninghistorical manuscriptslayout-level annotationspage-xml
Composed Historical Image Retrieval by Modeling Temporal Representations
The paper proposes Temporally Decomposable Image Representations (TDIR), a representation learning algorithm that decomposes historical photographs into orthogonal date and content subspaces. TDIR enables transitive temporal operations on embeddings without label supervision, while maintaining performance on downstream tasks. Theoretical analysis proves decomposition conditions and characterizes approximation errors. Evaluated on Composed Image Retrieval with historical photographs, TDIR preserves temporal structure while achieving competitive accuracy in both date estimation and object retrieval.
representation learningorthogonal subspacescomposed image retrievaltemporal decompositionembedding spaces
Europe's Climate Ambition Under Scrutiny: Evidence from Deep Learning Emission Projections
This study evaluates the EU27's progress toward its 2030 emissions reduction target using deep learning to project sectoral CO$_2$ trajectories from high-resolution socioeconomic data. The analysis extrapolates current trends without assuming new policy interventions, revealing a projected 35% shortfall (620 Mt CO$_2$) relative to the target. While the Power sector shows alignment due to renewable adoption, Mobility remains stagnant, contributing disproportionately to emissions. The findings highlight a systemic ambition-implementation gap requiring urgent policy action.
deep learningemissions projectionsectoral analysisclimate policyco$_2$ trajectories
Aslema at NADI 2026: Augmentation through Fewshot for SLU
The paper introduces Aslema, a system for NADI 2026 Shared Task 5 on intent recognition and slot filling in Tunisian Derja. It compares zero-shot performance of four omni LLMs with fine-tuned models, demonstrating fine-tuning's superiority. Synthetic data augmentation via LLM-generated culturally grounded utterances and voice cloning further improves performance. The final Qwen3-Omni-30B-based system achieves 86.8% intent accuracy and 34.7 WER on devtest, ranking 1st in slot filling (59.5 CoER) and 4th in intent recognition (66.1% accuracy) on the test set.
intent recognitionslot fillingsynthetic data augmentationvoice cloningtunisian derja
RTPO: Reverse-Turn Policy Optimization for Stabilizing Agentic RL Training
The paper introduces Reverse-Turn Policy Optimization (RTPO), a method to stabilize multi-turn reinforcement learning (RL) training for agentic workflows. RTPO addresses three instability sources—rollout-training context mismatch, weak turn-level credit assignment, and asynchronous policy drift—by organizing rollouts as sparse reverse trees and performing turn-level updates in reverse temporal order. Theoretical analysis shows RTPO eliminates context mismatch and drift while improving credit assignment. Experiments on multi-turn RL benchmarks demonstrate performance gains of 21.50% over trajectory-level and 10.76% over turn-level baselines.
multi-turn rlpolicy optimizationcredit assignmentasynchronous driftagentic workflows
Sanyu Studio: A Multi-Agent System for Art-Historical Narrative Construction
Sanyu Studio introduces a multi-agent dialogue system for constructing plural art-historical narratives, addressing concerns about generative AI's standardization effects. The system models 321 Sanyu oil paintings as agents equipped with fact retrieval, interpretation, organization, and memory-filtering mechanisms. A seven-day workshop with eight art-university participants demonstrated that user prompts, evidence organization, and cognitive biases yield divergent yet coherent narrative versions, suggesting AI can enhance human agency in art interpretation under limited historical evidence.
multi-agent systemart-historical narrativememory-filteringgenerative aiinterpretation mechanisms
Orienteering Problem with Uncertain Time-Varying Rewards: Framework and Benchmark for Everyday Service Robotics
The authors introduce the Orienteering Problem with Uncertain Time-Varying Rewards (OP-UTVR), extending classical orienteering by modeling stochastic, dynamic rewards observed during execution. They propose three planners with varying planning horizons and online adaptivity, providing theoretical performance bounds under reward uncertainty. A mobile service robot benchmark demonstrates trade-offs between planning horizon length and adaptation capability, with long-horizon adaptive planning showing superior performance in pedestrian-rich indoor environments.
orienteering problemtime-varying rewardsonline adaptationservice roboticsstochastic planning
Candidate-Fate Accounting for Transparent Sensor Diagnostic Pipeline Search
The paper introduces candidate-fate accounting, an audit framework for transparent sensor diagnostic pipeline search in AutoML/AutoDL systems. The method records all pipeline candidates (valid/invalid, pruned, or unfitted) via hashing, legality checks, allocation rationales, and a closed fate ledger. Experiments on three bearing-diagnostic datasets reveal 30–41 omitted candidates in fitted-trial-only reports while maintaining diagnostic performance, with the framework verifying complete candidate accounting.
automlsensor diagnosticsaudit frameworkpipeline searchcandidate-fate accounting
Change Point--Aware Evaluation and Re-Calibration of PPG-Based Blood Pressure Estimation
The study introduces a fluctuation-aware evaluation framework for photoplethysmography (PPG)-based blood pressure (BP) estimation, addressing limitations of aggregated metrics by focusing on performance during BP change points detected via time-series analysis. The method identifies abrupt distributional shifts in BP trajectories, revealing significant performance degradation in state-of-the-art models during fluctuations. A targeted re-calibration framework triggered by change points improves robustness without architectural modifications, demonstrating the importance of dynamic evaluation for real-world BP monitoring.
photoplethysmographyblood pressure estimationchange point detectiontime-series analysisre-calibration
Preference Reasoning under Indeterminacy in Large Language Models
The paper identifies indeterminacy in preference reasoning as a critical challenge for large language models (LLMs), distinguishing between epistemic (incomplete/partial preferences) and structural (nonexistent solutions) forms. Through a hierarchical task framework, the authors demonstrate that state-of-the-art LLMs fail to differentiate determined from undetermined instances, exhibiting miscalibrated reasoning even in verification tasks. This highlights a fundamental limitation in current models' alignment and decision-making capabilities under real-world conditions.
preference reasoningindeterminacylarge language modelssocial choicealignment
CTIFoundry: An Agent-Native Corpus Scaffold for Cyber Threat Intelligence
CTIFoundry introduces an agent-native corpus scaffold for cyber threat intelligence (CTI) to address the substrate bottleneck in LLM-agent investigations. The system materializes latent CTI structure through an ontology graph spanning four knowledge bases (CVE, CWE, CAPEC, ATT&CK), span-grounded report layers, and hybrid retrieval surfaces, exposed via seven typed tools and three procedural skills. Evaluated on CTIConnect, it improves F1 scores by +0.19 to +0.28 across four models, with smaller models outperforming flagship models on flat substrates while reducing tool calls by ~50%.
cyber threat intelligenceontology graphspan-groundedhybrid retrievalprocedural skills
Denoising-Aware Inversion: Revealing Privacy Risks in Noise-Protected Text Embeddings
The paper introduces DAEI, a denoising-aware embedding inversion pipeline that exposes privacy risks in noise-protected text embeddings. Addressing the 'Double Noise Trap' that hinders existing generative inversion methods, DAEI combines a residual denoising autoencoder with unsupervised generative text inversion using Stein's unbiased risk estimate. Experiments demonstrate a 154% relative BLEU improvement over baselines, with 32–60% gains in token-level F1 and ROUGE-L, challenging the sufficiency of Gaussian noise for privacy protection.
embedding inversiondenoising autoencoderprivacy leakagegaussian noisestein's unbiased risk estimate
Can a Lightweight Multimodal Model Estimate LLM Reasoning Performance? A Study for Compute-Optimal Document Inference
The paper introduces BudgetDoc, the first multimodal benchmark for model-budget-performance trade-offs in document tasks, and DRB (Document-Reasoning Balancer), a 1B-parameter estimator (SigLIP-2 + Qwen3-0.6B) that predicts ordinal LLM performance across budget levels (0.753 weighted F1). DRB dynamically allocates reasoning budgets for five frontier models on three datasets, matching or improving F1 scores in 9/15 configurations versus maximum-budget baselines while reducing cost. Preliminary results suggest cross-model generalization potential.
multimodal benchmarkreasoning budgetpre-flight estimatordocument taskscross-model selection
OmniHandwritingOCR: A Diagnostic Benchmark for Evaluating Multimodal LLMs in Handwritten OCR Scenarios
OmniHandwritingOCR introduces a diagnostic benchmark for evaluating multimodal LLMs in handwritten OCR, addressing gaps in existing benchmarks by covering multilingual handwriting, writer errors, and complex mathematical expressions. The benchmark comprises six subtasks and twelve subsets, totaling 77.57K labeled images from public datasets and newly collected student writings, including a difficulty-stratified multi-line formula corpus. Evaluation of thirteen open- and closed-source systems reveals significant performance drops on complex formulas, inconsistent rankings across languages, and hallucinated corrections, highlighting challenges in visual grounding and structural understanding.
multimodal llmshandwritten ocrdiagnostic benchmarkmathematical expression recognitionvisual grounding
From Storage to Access: Verifiable Activation of Parametric Knowledge in LLMs via Explicit Priming and Implicit Reasoning
The paper introduces VAKE (Verifiable Activation of Parametric KnowledgE), a two-stage reinforcement-learning framework for improving factual question answering in LLMs by decoupling knowledge elicitation from reasoning. VAKE first uses an explicit Priming policy to insert verifiable bridging triples into insufficient retrieved subgraphs, then transfers this capability to implicit Reasoning via a frozen model. Evaluated across seven benchmarks on 3B-14B models, VAKE outperforms baselines, with LLM-based analysis showing 80% of inserted triples provide novel factual knowledge and over 50% elicit latent parametric knowledge inaccessible through direct prompting.
parametric knowledgeknowledge elicitationreinforcement learningfactual question answeringverifiable evidence
FACET: Preserving Source Intent and Executable State in Terminal Task Synthesis
FACET introduces a framework for synthesizing high-quality terminal tasks by preserving source intent and ensuring cross-artifact consistency through environment-grounded construction. The method reconstructs agent skills into coherent scenarios, realizes and repairs execution environments, and uses container state as shared grounding for instructions, solutions, and verifiers. Execution-based validation and targeted repair correct failures without regenerating valid components. Results show improved performance on Terminal-Bench 2.1, with fine-tuned models achieving consistent gains, demonstrating the importance of source-intent preservation and shared executable-state grounding for scalable task synthesis.
terminal task synthesisexecutable-state groundingenvironment-grounded constructionsource-intent preservationexecution-based validation
MR-IQA-2: Faithful Image Quality Reflection via Fine-Grained Credit Assignment
MR-IQA-2 introduces a framework for faithful image quality assessment (IQA) by decoupling credit assignment for reasoning and rating. The method employs an actor-editor-judge pipeline: an actor generates quality reasoning, an editor modifies the image based on identified factors, and a frozen judge compares original/edited images to provide reflective supervision. Fine-grained credit assignment separates reasoning (supervised by judge feedback) from rating (supervised by human ratings), using masked token-specific updates. On IQA benchmarks, MR-IQA-2 achieves competitive human-aligned ratings while enabling richer visual understanding for quality optimization. Code is available.
multimodal large language modelsimage quality assessmentcredit assignmentreasoning faithfulnessreflective supervision
The Role of Grid Cells in Reducing Spatial Aliasing in Hippocampal Place Representations
The study demonstrates that grid cells significantly reduce spatial aliasing in hippocampal place representations by providing periodic, internally generated spatial signals that complement boundary vector cell (BVC) inputs. The authors integrate analytically constructed grid cell modules with BVC-driven place cells, testing the approach in three environments: an open space, a cross-shaped obstacle layout, and a maze. Results show a 94--99% reduction in spatial aliasing compared to BVC-only baselines, with the greatest improvement in the most visually symmetric environment. This highlights grid cells' role in disambiguating perceptually identical locations.
spatial aliasinggrid cellshippocampal place representationsboundary vector cellsenvironmental symmetry
MorphoGP: A Nonparametric Framework for Predicting Equilibrium Beach Profiles Under Tidal Influence
MorphoGP proposes a nonparametric Gaussian process framework for predicting equilibrium beach profiles under tidal influence, addressing limitations of traditional models in diverse coastal environments. The method combines a ContourCluster model for morphology classification via contrastive learning with category-specific Gaussian process experts, integrated through a Gating Net. Evaluated on 180 tide-influenced Chinese beach profiles, MorphoGP reduces test RMSE by 59.3% versus baselines, achieving 0.297 m RMSE.
gaussian processcontrastive learningequilibrium beach profilestidal influencenonparametric framework
Performance Drift Detection in Machine Learning as a Service (MLaaS) for IoT Environments
The paper proposes a novel MLaaS Performance Drift Detection (MPDD) framework for IoT environments, addressing challenges in black-box MLaaS systems where data distribution shifts and model updates cause performance drift. The framework combines an MLaaS extraction model that learns service behavior from input-output pairs with an Adaptive-Temporal Performance Drift Detection Mechanism (APDDM) that dynamically adjusts monitoring frequency. Experiments on real-world datasets show MPDD achieves 22-25% higher accuracy than baselines, while APDDM yields 4% accuracy gain and 9% lower miss detection rate versus fixed-interval monitoring.
mlaasperformance driftiotadaptive monitoringblack-box learning
CentaurBench: Benchmarking LLM Capabilities on Augmenting vs. Automating Real-World Work Tasks
The paper introduces CentaurBench, a framework evaluating LLMs on both automation and assistance capabilities for real-world tasks. Seven tasks assess how assistant models (providing guidance text) augment lower-capacity worker models versus directly automating outputs, scored via blind pairwise LLM-judged comparisons. Results show modest correlation (r<1) between automation and augmentation rankings, with the automation-optimal model losing in augmentation for 5/7 tasks. Assistance quality varies: unaided workers outperform all assisted conditions in 3/7 tasks, and only one model's guidance consistently improves outcomes. Findings highlight automation benchmarks as insufficient proxies for assistance efficacy in multi-agent systems.
llm benchmarkinghuman-ai collaborationmulti-agent systemsassistance qualityautomation-augmentation tradeoff
Bridging Search and CRM: Productionizing AI Product Research Agents for Customer Re-Engagement
The paper introduces a production-deployed AI framework that bridges search and CRM systems for proactive customer re-engagement in e-commerce. The system employs AI-powered Product Research Agents to identify users with exploratory purchase intent, conducts multi-agent product research using behavioral signals, external knowledge, and catalog data, and delivers personalized recommendations via WhatsApp. In a 23-day deployment with 15K notifications, the framework achieved higher CTR than traditional campaigns, with measurable downstream purchases and GMV impact.
e-commercecustomer re-engagementmulti-agent systemsbehavioral signalswhatsapp recommendations
Evaluating and Explaining Prompt Sensitivity of LLMs Using Interactions
The paper introduces interactions as a fine-grained tool to analyze prompt sensitivity in large language models (LLMs), proposing an Interaction-based Prompt Sensitivity (IPS) metric to quantify changes in interactions due to subtle prompt modifications. By decomposing LLM output scores into nonlinear interactions among input variables, the authors identify four factors reducing prompt sensitivity: supervised fine-tuning, increased model scale, dense architectures, and few-shot learning, all of which stabilize low-order interactions. Evaluation across 50 open-source LLMs reveals that these factors mitigate instability in interactions even when final outputs remain unchanged.
prompt sensitivityinteraction decompositionllm instabilitynonlinear relationshipslow-order interactions
FinRCA-Bench: Benchmarking Evidence Retrieval and Reasoning for Financial AI Systems
Introduces FinRCA-Bench, a synthetic benchmark of 2,250 financial reconciliation cases with 1,500 injected failures across 15 causal categories, designed to isolate evidence retrieval from reasoning in financial AI systems. Evaluates five retrieval methods (Rules/SQL, classical ML, dense retrieval, relational expansion, Typed Provenance Graph Retrieval) with hidden evidence contracts, showing retrieval architecture critically impacts performance: exact accuracy ranges from 2.05% to 72.44% across methods, with structural retrieval failures dominating reasoning errors (95 vs 15). Demonstrates correct root-cause labels poorly indicate auditable diagnoses (5.72% evidence contract accuracy).
financial reconciliationevidence retrievaltyped provenance graph retrievalsynthetic benchmarkroot-cause analysis
Pairwise Ranking Outperforms Single-Action RL for Offline Explanation Selection: A Practical Lesson
Pairwise learning-to-rank outperforms single-action reinforcement learning for offline explanation selection in LLM-based recommendation systems, achieving F1 = 0.500 on a Google Local subset and F1 = 0.329 on MovieLens-1M. The method separates explanation generation from selection, using a frozen candidate pool (six prompt styles, two LLMs) and a CPU-resident selector to reduce latency (<100 ms) and cost ($15 end-to-end). Evaluations on 2,958-pair XRec Google Local and 300-pair MovieLens-1M splits show LambdaRank surpassing PPO, GRPO, DPO, and teacher-student distillation, with seed variance <0.003 F1. KG-path methods achieve USR = 1.000 on Google Local and 0.997-1.000 on MovieLens-1M, avoiding template-collapse failures.
pairwise learning-to-rankoffline explanation selectionlambda rankkg-path methodstemplate-collapse failures
DART-SD: Diamond-topology Aware Retrieval and Tuning for Self-Distillation of Multi-Turn Tool-Calling Agents
DART-SD introduces a diamond-topology aware framework for self-distillation of multi-turn tool-calling agents, addressing the topological collapse caused by monolithic trajectory imitation. The method models execution as an Interaction-State Transition Graph (ISTG), identifies Critical Topological Breakpoints (CTB) during rollouts, and applies localized supervision via CTB-guided recovery references. This preserves valid reasoning prefixes while correcting errors. Experiments on multi-turn tool-calling benchmarks show significant improvements over full-trajectory baselines.
self-distillationinteraction-state transition graphcritical topological breakpointmulti-turn tool-callingdiamond topology
Prior-Conditioned Gaussian Discriminants for Generalizable AI-generated Image Detection
The paper introduces prior-conditioned Gaussian discriminants as a method for improving generalizability in AI-generated image detection across domain shifts. The approach constructs closed-form classifier heads using first- and second-order feature statistics under nested covariance assumptions, evaluated on Percept-Lens (39 datasets, 7.1M images). Results show competitive or superior performance compared to existing detector heads when matched on prior and encoder, while highlighting sensitivity to training priors and data efficiency of moment-based heads. The work advocates for (prior, encoder, head)-level reporting and stronger analytical baselines in AIGI transfer studies.
gaussian discriminantsai-generated image detectionfeature statisticsdomain shiftstransfer learning
GCNO: Gramian Chebyshev Neural Operator for Physics-Based Compression of Wireless Channels
The Gramian Chebyshev Neural Operator (GCNO) is a physics-based, variable-rate compressor for wireless channel feedback that identifies dominant propagation paths instead of compressing the full channel matrix. GCNO leverages receive-transmit channel structure to locate paths, employs a first-order Taylor correction for direction refinement, and uses least squares to recover complex strengths, operating without path labels or a learned decoder. Evaluated across three ray-traced environments, GCNO outperforms neural baselines in reconstruction accuracy at fixed payloads (or reduces payload at equal accuracy) and generalizes to unseen antenna counts without retraining.
neural operatorwireless channel compressionphysics-based compressionvariable-rate feedbackray-traced environments
Which Negatives Matter? Ask Your Text Encoder: Adaptive Similarity Margins for Dense-Caption Retrieval
HN-CLIP introduces adaptive similarity margins for dense-caption retrieval by leveraging the text encoder's text-text geometry to assign per-negative margins, addressing premature optimization saturation in InfoNCE. The method adds a detached caption-similarity matrix to negative logits, requiring no auxiliary data or inference overhead. Evaluations on four benchmarks show improvements of +2.4--+4.3 R@1 over competitors, with 2.4x--5.4x faster training than GOAL and StructXLIP, and strong performance with only 20% training data.
dense-caption retrievaladaptive similarity marginsinfoncetext-text geometrycontrastive fine-tuning
OptiModNet: A UNet-Transformer Hybrid with Grouped-Query and Channel Attention for Optic Disc and Cup Segmentation
OptiModNet proposes a lightweight UNet-Transformer hybrid architecture for optic disc and cup segmentation, combining grouped-query attention and channel attention mechanisms to enhance local and global feature representation. The model employs an Aggregated Pyramid Loss to supervise multi-scale decoder predictions, improving gradient flow and structural consistency. Evaluated on the REFUGE2 dataset, OptiModNet achieves state-of-the-art performance (+2.5% over prior methods) with 3.73 GFLOPs and 1.93M parameters, balancing accuracy and computational efficiency.
unet-transformer hybridgrouped-query attentionaggregated pyramid lossoptic disc segmentationcomputational efficiency
Science Done on a Machine by a Machine: AI Agents in Computational Chemistry
This Perspective surveys the rapid proliferation of AI agentic systems in computational chemistry, documenting growth from half a dozen in 2024 to nearly fifty by August 2026. These systems are evolving from task assistance to autonomous experiment design, execution, analysis, and manuscript writing, trending toward fully autonomous AI scientists. While current implementations retain human oversight, the authors highlight commoditization via generalist agents and express uncertainty about the field's future roles for specialists and educators.
agentic systemscomputational chemistryautonomous experimentationin silico experimentsai scientists
UMER: Unifying Embedding and Ranking via Pair-Aware Discriminative Reasoning for Universal Multimodal Retrieval
UMER introduces a Unified Multimodal Embedding and Ranking framework for universal multimodal retrieval, addressing limitations in existing Chain-of-Thought (CoT) and contrastive embedding methods. The key innovation is Pair-Aware Discriminative Reasoning, which compares query-candidate pairs to identify instruction-relevant matching and discrepancy evidence, replacing item-wise reflection. UMER jointly learns contrastive embeddings for global matching and discriminative ranking for pairwise relevance judgment within a single MLLM, enhanced by mutual distillation. On the MMEB-V2 benchmark, UMER achieves state-of-the-art performance under comparable settings while supporting budget-adjustable inference.
universal multimodal retrievalpair-aware discriminative reasoningcontrastive embeddingsmutual distillationmllm
Physics-Unrolled Neural Operator for Wireless Field Modeling
The paper introduces Physics-Unrolled Hybrid Neural Operator (PU-HNO), a three-stage cascade model for high-fidelity indoor radio-map prediction from low-fidelity ray-tracing outputs and scene priors. PU-HNO progressively captures reflection, diffraction, and scattering effects, leveraging conditionally unbiased label noise to learn stable propagation structure. Experiments demonstrate superior performance over image-to-image baselines, wireless learning models, and monolithic neural operators in both image-quality and wireless deployment metrics across diverse floorplans.
radio-map predictionneural operatorray-tracingpropagation effectsmonte carlo noise
Partition the Support, Reconstruct the Residual: Training-Free Sparse Attention for Video Generation and World Models
The paper introduces SparsePR, a training-free block-sparse attention method for accelerating video transformers by combining Response-Coupled Partitioning with Probe-Fitted Residual Reconstruction. Sampled-query key responses form K/V groups for shared routing, while probe queries calibrate an affine correction to reconstruct residuals. Evaluated on four video generation and world models, SparsePR reduces attention-reconstruction error, maintains generation quality at 22.0-26.0% executed-pair density, and achieves 1.48x-2.61x end-to-end speedups. Ablations highlight probe fitting's dominant role in error reduction and partitioning's benefits for hard-drop error.
sparse attentionvideo generationworld modelsresidual reconstructionquery-response partitioning
Coverage-Driven RTL Assertion Generation with Formal Exploration and Neuro-Symbolic Refinement
NeuroAssertion introduces a coverage-driven framework for generating high-quality RTL assertions by combining formal trace generation, syntax-guided synthesis (SyGuS), and neuro-symbolic refinement. The method first identifies hard-to-reach control-flow conditions, uses model checking to produce diverse traces, and mines initial assertions via SyGuS. A two-stage LLM process then refines assertions: one proposes candidates for uncovered regions, while another repairs failures via constrained symbolic synthesis. Experiments demonstrate ~2× more assertions and ~2× higher mutation coverage versus traditional methods.
register transfer levelsyntax-guided synthesisneuro-symbolic refinementmodel checkingassertion mining
ERASE: EaRly bAckpropagation SchEdule for Faster Training of Modern Recommendation Systems
ERASE proposes an early backpropagation schedule to accelerate training of recommendation systems by overlapping forward and backward passes. The method leverages Forward-Forward's detachment mechanism to decouple gradient dependencies, enabling concurrent execution of backward passes for detached subgraphs on separate CUDA streams. Experiments on a lightweight transformer and large-scale click-through-rate model show throughput improvements up to 9.51% while maintaining comparable normalized entropy to the baseline.
backpropagationforward-forwardcuda streamsrecommendation systemsthroughput optimization
Formal Verification of Romanov's Triplet Logic: A Verified Filter for Sliding-window 3-CNF with Application to Structured Formulas
The authors present the first mechanized formalization of Romanov's Triplet Logic (TLS) in the Rocq proof assistant, establishing its core properties and verifying key procedures. They formalize Compact Triplets Structures (CTS), hyperstructures, clearing, and Simple Vertex Intersection (SVI), proving polynomial-time bounds for filter stages and identifying a precise correctness boundary for joint satisfying sets. The work includes VFR, an extracted OCaml prototype for verified decision procedures, validated through benchmarks on random and structured instances. The Rocq development comprises 23,000+ lines of code, 427 lemmas/theorems, and zero admitted goals.
triplet logicformal verificationcompact triplets structuressimple vertex intersectionrocq proof assistant
Pedagogical AI in Mental Health: A Tri-Stream Fine-Tuned LLM Framework for Automated Clinical Supervision and Risk Triage
The paper introduces a tri-stream fine-tuned Mistral-7B-instruct framework for automated clinical supervision in mental health, addressing the supervision gap. The model analyzes therapeutic alliance, latent risk, and supervisory triage via a Dynamic Clinical Urgency Index (D-CUI), leveraging 106 sessions from DAIC-WOZ with multi-modal VAL (Visual-Acoustic-Linguistic) fusion. Results show 95% technique identification accuracy, alliance assessment MAE of 0.105, and mean D-CUI of 0.370, reducing triage latency from 72 hours to real time (~10 seconds). Training achieved 85.2% loss reduction on a Tesla T4 GPU.
tri-stream analysisdynamic clinical urgency indexmulti-modal fusiontherapeutic alliancelatent risk prediction
FM-Bench: A Benchmark for Long-Horizon Management with Competing Agents
FM-Bench introduces a novel benchmark for evaluating long-horizon decision-making in LLM agents through a football management simulation spanning 20 in-game years. Agents interact with 26 tools across 340-400 decision points, managing squad dynamics, finances, and board relations within a deterministic scoring framework. Testing 15 frontier models revealed that Claude-Fable-5 achieved top performance, with behavioral strategies (e.g., proactive renewals, cash investment) outperforming scale or token spend, while no model learned hidden market prices and memory management exhibited critical failures.
long-horizon decision-makingllm agentsdeterministic enginebehavioral capabilitiesself-managed memory
Mechanistic Interpretability of Structure-Aware Numerical Reasoning in LLaMA 3.1 8B
The study provides mechanistic interpretability evidence that LLaMA 3.1-8B performs structure-aware numerical reasoning by tracking first differences in sequences. Through probing and activation patching on a custom sequence modeling task requiring structural understanding, the authors demonstrate that the model internally computes and stores first differences without supervision, using an induction-circuit-like mechanism to retrieve and apply them. Results show LLaMA 3.1-8B successfully solves the task by adding retrieved differences to current values, representing early evidence of concept induction in LLMs.
mechanistic interpretabilitynumerical reasoningactivation patchinginduction circuitfirst differences
Improving Natural-Language Combinatorial-Optimization Accuracy in Resource-Constrained Language Models via Formal Abstractions
The paper introduces SDDL, a neuro-symbolic framework that improves combinatorial scheduling accuracy in resource-constrained language models by translating natural-language problems into solver-aligned representations. SDDL delegates low-level modeling and search to a deterministic compiler and external solver, preserving feasibility constraints. Evaluated on a 300-instance scheduling benchmark, SDDL boosts feasibility rates from 23.7%/1.3% (direct-generation) and 21.7%/7.0% (solver-code) to 55.3%/28.3%, with 0.0% median optimality gap among feasible solutions, enabling smaller models to match larger frontier models.
combinatorial optimizationneuro-symbolicfeasibility preservationsolver-aligned representationsresource-constrained models
Vector Symbolic Policy Gradient
Vector-Symbolic Policy Gradient (VSPG) introduces a discrete-action actor representing actions as unit-norm hypervectors, scored by similarity to encoded states. The method's update is proven to be advantage-weighted hypervector bundling with normalization, compatible with standard advantage estimators. Trained action hypervectors function as fixed-size compressed kernel memories, storing advantage-weighted kernel expansions over visited states for sample-efficient learning. For bipolar action memories, greedy selection is stable under random bit flips, with failure probability decaying exponentially in hypervector dimension. VSPG bridges Vector Symbolic Architectures (VSA), log-linear policy gradients, and kernel policy search while offering robustness guarantees.
vector symbolic policy gradienthypervector bundlingkernel memoryadvantage-weightedbipolar action memories
LEDGER: Claim-to-Evidence Trace Graphs for Auditing LLM Agents
LEDGER introduces layered trace graphs for auditing LLM agent workflows by structuring execution events into Trace Records, Evidence Nodes, and Workflow Nodes with typed semantic edges. The system connects claims to supporting actions, artifacts, and validation steps, enabling granular analysis of workflow decisions, artifact lineage, and validation coverage. Demonstrations in data analysis and coding show its utility for evidence-centered audits of complex agent outputs.
llm agentstrace graphsevidence nodesartifact lineagevalidation coverage
When Clean Signals Are Not Enough: Detecting Structural Ambiguity for Safe Wearable Stress Classification
The paper introduces the Individual Conformal Coupling Monitor (ICCM), a lightweight pre-inference monitor addressing structural ambiguity in wearable stress classification, where individually plausible physiological signals form patterns unsupported by non-stress references. ICCM quantifies subject-specific coupling divergence and routes each window to classify, defer, or abstain without retraining the downstream classifier. Evaluated on WESAD (N=15) and Stress-Predict (N=35) datasets, ICCM shows negative Pearson associations between ambiguity and accuracy (r=-0.607, p=0.016; r=-0.412, p=0.014), though robustness analyses temper these findings. ICCM reduces false positives marginally (29 to 27, 94 to 92) but does not resolve missed-stress failures, positioning it as an interpretable signal rather than a standalone safety guarantee.
structural ambiguitywearable stress classificationindividual conformal coupling monitorcross-signal couplingfalse-positive reduction
A Jagged Frontier: Evaluating Robustness of Code Agents to Semantics-Preserving Transformations
The study evaluates the robustness of AI code agents to semantics-preserving transformations (SPTs) in repository-level code repair tasks. Using a random variant sampler to apply SPTs (control-flow rewrites, dead-code injection, identifier renaming), the authors test two agentic scaffolds (mini-SWE agent, OpenCode) backed by four frontier models (Claude Opus 4.5, Kimi K2.5, MiniMax M2.5, Qwen 3.6-27B) on SWE-bench Verified and SWE-bench Pro. Results show up to 6.7 percentage points mean resolve-rate drop, with significant degradations in 6 of 16 configurations, revealing scaffold-dependent robustness and no consistent model ranking.
semantics-preserving transformationscode agentsresolve-rateagentic scaffoldsrepository-level repair
TTSD-FAR: Test-Time Self-Distillation with Fisher-Anchored Restoration for Missing-Modality Emotion Recognition in LVLMs
The paper introduces TTSD-FAR, a parameter-efficient framework for missing-modality emotion recognition in large video-language models (LVLMs). The method combines test-time self-distillation (TTSD), where a frozen teacher guides a low-rank student, with Fisher-Anchored Restoration (FAR) to monitor Fisher information stability and prevent drift during adaptation. Evaluated on MELD, DFEW, and BAH datasets with 0%-50% missing modalities, TTSD-FAR outperforms entropy-based adaptation, retrieval-augmented generation (RAG), and perplexity-based methods, maintaining consistency over long adaptation horizons where baselines degrade.
test-time adaptationself-distillationfisher informationlow-rank adaptationmissing-modality
Selection, Recombination, or a Fresh Solve? A Candidate-Free Control for Single-Pass Test-Time Aggregation
The study introduces a candidate-free control for single-pass test-time aggregation to evaluate whether candidate conditioning improves reasoning accuracy when correct candidates are unavailable. Using Qwen3-4B on AIME-2025 and HMMT-2025, results show candidate conditioning improves accuracy with multiple correct candidates (Δ_cand(c2+) = +0.290) but reduces it when all candidates are wrong (Δ_cand(c0) = -0.123). The findings hold after conservative correction, revealing that conditioning on all-wrong candidates harms performance versus fresh solving. Limitations include model and benchmark specificity, truncated candidate fragments, and single-pass aggregation.
test-time aggregationcandidate conditioningqwen3-4bmathematics benchmarkssingle-pass reasoning
One Gate Is Not Enough: Composing Stateful Pre-Action Controls for Agentic AI
The paper formalizes remediation-induced control coupling in agentic AI systems, where one control's remediation affects another's evaluation, and proposes a remediate-and-regate protocol to restore per-action soundness. Using finite-model checking, it demonstrates non-commutativity of remediation operators (evidence substitution and resource-budget downroute) and identifies vulnerabilities in governed evidence buffers. Theoretical results include conditions for linear aggregation to compensate vetoes and unified evidence sets. Empirical validation on three unmodified engines confirms protocol adherence in 5/6 cases, with one workflow-dependent exception.
remediation-induced couplingpre-action controlsfinite-model checkingevidence substitutionresource-budget downroute
Task-Conditioned Least-Privilege Learning for Executable Terminal and MCP Agents
The paper introduces task-conditioned least-privilege learning for executable terminal and Model Context Protocol (MCP) agents to mitigate excess-authority errors in tool-using LLMs. The method audits actions via deterministic verifiers across six risk dimensions, optimizing trajectories using task-specific sufficient-authority envelopes during post-training. Evaluated on Qwen3.5-4B over 1,500 tasks, the approach achieves 98.48% safe success (vs. 64.36% baseline) on 500 held-out tasks, reducing excess-authority errors from 4.56% to 0.79%, with evidence of generalization in a 400-task continuation study.
least-privilege learningexcess-authority errorsdeterministic verifierssufficient-authority envelopesmodel context protocol
Coupled-cluster molecular properties across the main group that extrapolate beyond training size
MEHnet-MG, an equivariant neural network, predicts coupled-cluster-quality molecular properties from a single B3LYP/def2-SVP calculation, achieving 3.8x to 230x error reduction over DFT methods across nine main-group elements. The model derives properties (energy, optical gap, dipole, etc.) from a predicted Hamiltonian, ensuring correct size-scaling, and generalizes to 58-atom chains where pooling-based architectures fail. Trained on a CCSD(T)-labeled dataset, it matches CCSD polarizability and EOM-CCSD optical gaps to ~2% at 44/37 atoms and adds only ~25 ms inference time per molecule.
coupled-cluster theoryequivariant networkmolecular hamiltoniansize-scalingdft acceleration
Low-Power, Neuromorphic, Acoustic Anomaly Detection for Persistent Machine Monitoring
The study presents a neuromorphic approach for low-power acoustic anomaly detection using an Intel Loihi 2 processor. The system employs an autoencoder with log-mel features computed off-chip, while normalization, inference, L1 reconstruction scoring, and thresholding run on-chip. Evaluated on the ToyADMOS ToyCar benchmark, it achieves 0.9959 AUC (clean) and 0.7990/0.6466 source/target AUC (noisy DCASE 2026 Task 2), outperforming baselines. Power measurements show 0.0406–0.0426 mJ per sample, two orders lower than CPU/GPU implementations, enabling persistent monitoring.
neuromorphic computingacoustic anomaly detectionautoencoderlog-mel featureslow-power inference
From Inference to Adaptation: A Unified Optimal Transport View of Vision Language Model
The paper proposes \algname, a unified optimal transport (OT) framework for vision-language model (VLM) test-time adaptation (TTA) that bridges the gap between inference and adaptation objectives. The method formulates zero-shot image classification as a Wasserstein OT problem for robust pseudo-labeling, then adapts VLMs via a soft-label InfoNCE loss derived from the same OT formulation, enabling fine-grained contrastive learning. Theoretical analysis shows the InfoNCE loss equivalently expresses a Wasserstein OT, unifying inference and adaptation. Experiments demonstrate state-of-the-art performance, outperforming prior methods by up to 7% with efficient computation.
optimal transportvision-language modelstest-time adaptationwasserstein distancecontrastive learning
Measuring the Partial-Credit Gap: A Strict Benchmark on Vietnam's 2025 Convex Marking Scheme
The paper introduces THPT-Ladder, a benchmark for evaluating language models on Vietnam's 2025 National High School Graduation Examination, which uses a convex grading scheme that penalizes partial knowledge. The benchmark comprises 632 items from 21 official exams across 11 subjects, graded identically to human candidates. Results show that standard accuracy metrics inflate model scores by 0.020 to 0.159 points per question compared to the official rubric, altering percentile rankings (e.g., Qwen3.5-27B drops from the 90th to 77th percentile). Error distribution affects scores significantly, with identical accuracy yielding 0.869 to 0.932 points per question.
convex gradingpartial creditlanguage model evaluationbenchmarkhigh-stakes testing
Governance Records as Supervision: Verifier-Selected Self-Training for Structured Workflow Repair
The paper introduces verifier-selected self-training for structured workflow repair, demonstrating how machine-verifiable governance records can supervise bounded models without oracle targets or stronger teachers. Using PlanBench replanning cases, Qwen3-14B generated 24 VAL-verified plans that trained the same model for non-thinking execution, increasing VAL-accepted plans from 1 to 57 on 80 test cases. VAL-selected supervision outperformed schema-selected and model-self-selected alternatives by +33 net gains (p<0.000002), while maintaining schema validity and reducing latency by 56×. Complementary experiments with Phi-4 showed similar improvements (2→51 accepted plans).
governance recordsverifier-selected supervisionstructured workflow repairplanbenchschema-valid
FedCoRe: Target-Adaptive Completion for Missing Modalities in Healthcare Federated Learning
FedCoRe introduces a federated learning framework for missing modality completion in healthcare, focusing on representation- or logit-space corrections rather than synthetic data generation. The method uses paired supervision from clients observing both complete and missing modalities, with updates restricted to clients possessing such pairs and validation potentially retaining baseline predictions. Evaluated on a MIMIC-derived respiratory deterioration task, FedCoRecovered 49.0% of lost AUROC for missing ECG and 52.8% for missing CXR, demonstrating partial performance restoration through task-specific corrections rather than literal modality recovery.
federated learningmissing modalitiesrepresentation completionlogit-space correctionmultimodal models
ComponentBench: Diagnosing Component-Level Failures in Computer-Use Agents
ComponentBench introduces a benchmark and diagnostic pipeline for evaluating computer-use agents at the component level on modern web UIs, addressing the gap between atomic tests and long-horizon workflows. The benchmark comprises 2,910 programmatically verified tasks across 97 UI components, with cleaned human reference trajectories for assessing task success and interaction efficiency. Evaluating seven models (e.g., GPT-5.4, Gemini 3 Flash) across four observation/action spaces reveals performance variations up to 30%, with spatial manipulations remaining challenging and agents taking 3.7x longer than humans.
computer-use agentsui componentsbenchmarkobservation spaceaction space
SESSE: Sketch, Expand, Sort, Summarize, Evaluate -- LLM-as-Judge Evaluation via Structured Decomposition
The paper introduces SESSE, a training-free framework for LLM-as-judge evaluation that decomposes holistic judgments into structured sub-questions derived from the judge's error cases, eliminating the need for oracle responses or task-specific rubrics. SESSE operates via five steps: Sketch, Expand, Sort, Summarize, and Evaluate, providing interpretable per-criterion vote evidence. On RewardBench (n=1,000), SESSE matches chain-of-thought baselines and approaches the performance of fine-tuned RISE-Judge-32B (92.7%) while remaining training-free. The method offers an audit trail for diagnosing label ambiguity and judge failures.
llm-as-judgetraining-freestructured decompositionrewardbenchinterpretable evaluation
The Lifecycle of LLM-as-a-Judge for Large-Scale Recommendation Explanations
The paper proposes a lifecycle framework for LLM-as-a-Judge systems in production, focusing on Netflix's recommendation explanation pipeline. The framework comprises four phases: (I) Birth (benchmark creation with human labels), (II) Training (Reasoning-Aligned Rubric Tuning, RART), (III) Deployment (quality gating and reflective generation), and (IV) Monitoring (human-in-the-loop drift detection). A five-week A/B test demonstrated improved member engagement, with increased novel content viewing (+X%) and browse-to-play conversions (+Y%), without quality issues.
llm-as-a-judgereasoning-aligned rubric tuninghuman-in-the-looprecommendation explanationsdrift detection
FairGlucose: A CGM Fairness Benchmark Reveals Subgroup Disparities Hidden in Population-Level Validation
The study introduces FairGlucose, a 300-patient continuous glucose monitoring (CGM) benchmark with 132,480 forecasting samples and 3,945 behavioral events, balanced across 12 demographic strata. Evaluating 33 models on 2-hour glucose forecasting, it reveals that population-level validation masks significant subgroup disparities: type 1 diabetes patients exhibit 6 mg/dL higher prediction error than type 2 (p < 0.001), with performance gaps correlating with clinically hard cases. Behavioral events contribute minimally (≈0.1 mg/dL), and frontier LLMs underperform specialized neural models by 1-6 mg/dL. The findings advocate for subgroup-disaggregated reporting in digital health AI validation.
continuous glucose monitoringsubgroup disparitiesglucose forecastingpopulation-level validationdemographic strata
Debiased Inference for AI-Generated Data without Gold-Standard Labels: Identification via Multiple Imperfect Measurements
The paper proposes Debiased Inference with Multiple Imperfect Measurements (DMM), a framework for valid downstream statistical analysis using AI-generated data without gold-standard labels. DMM leverages multiple error-prone AI measurements, assuming conditional independence given latent true labels and unit-level features (e.g., text embeddings), while allowing unknown misclassification rates to vary across methods and units. Theoretical analysis proves DMM's consistency and asymptotic normality, with simulations demonstrating valid inference and efficiency gains from additional imperfect measurements. Diagnostics for assessing conditional independence are provided, focusing on large language model annotation applications.
debiased inferenceai-generated dataconditional independencesemiparametric inferencemisclassification rates
Evaluating Structured Information Extraction with Open Models in a High Risk Public Sector Application
This work benchmarks open-source structured information extraction systems on a high-risk public sector application: processing international student applications. The evaluation compares Optical Character Recognition (OCR) engines, Large Language Models (LLMs), and Vision-Language Models (VLMs) in multi-step pipelines. Results show VLMs generally outperform OCR+LLM combinations, but even state-of-the-art open models achieve limited reliability (only 4/35 configurations exceed F1=0.5). Performance scales non-linearly with model size, and OCR output structural preservation proves critical for downstream accuracy.
structured information extractionoptical character recognitionlarge language modelsvision-language modelshigh-risk ai applications
What Makes Software Issue Resolution Tasks Difficult for Agents?
The study proposes a measurement framework to quantify structural properties affecting agent success in software issue resolution tasks, using CoderForge-Preview dataset analysis. Methodologically, it extracts patch, repository, and prompt features, applying ensemble methods, SHAP attribution, and effect size analysis. Results show task difficulty is highly predictable from static features (AUC=0.863), primarily driven by patch fragmentation and repository scale, with prompt linguistic features emerging in mid-difficulty tasks.
agentic systemstask difficultypatch fragmentationshap attributioncoderforge-preview
SeisEvo: Evolution of Seismic Data Reconstruction Algorithms by Agents
SeisEvo introduces an LLM-driven multi-agent system for evolving seismic data reconstruction algorithms through constrained program evolution, producing standalone white-box operators. The method modifies classical algorithms via agentic search while enforcing physical constraints, evaluating candidates through execution. Results show Evo-POCS improves SNR by 3.49 dB over POCS for interpolation, while Evo-MSSA achieves 7 dB SNR gain over MSSA for simultaneous interpolation and denoising, with maintained performance on unseen data.
seismic reconstructionalgorithm evolutionllm-driven agentswhite-box operatorssnr improvement
How AI Prompts Can Teach Us About the Structure of Human Behavior
The study introduces an AI-based method for analyzing human behavior structure by assigning multidimensional 'type vectors' (e.g., Altruism, Risk Aversion) to a large language model and optimizing these dimensions to match 119,147 human decisions across 10 economic games. Using type vectors like (2,4) to condition model prompts, the method identifies three key dimensions (Risk Aversion, Strategic Sophistication, Trust) that parsimoniously explain behavior, clustering human types into <12 groups with cross-game predictive validity. Results support low-dimensional, portable representations of behavior across diverse settings.
type vectorsstrategic sophisticationeconomic gamesbehavioral dimensionscross-game prediction
Cacheable by Design? Training Mixture-of-Experts Routers for Locality Against the Edge Memory-Bandwidth Wall: A Pre-Registered Negative Result with a Systems Measurement Study
This work presents a pre-registered evaluation of cacheability training in Mixture-of-Experts (MoE) models, focusing on edge-serving memory-bandwidth constraints. Using llama-moe-trace, a router-telemetry tool, the authors measure expert reuse patterns in Qwen3-30B, finding 66% of requests served by a 13.4% LRU expert cache. They train 137M MoE models with auxiliary locality and domain router losses, achieving up to 60% cache-miss reduction but failing pre-registered perplexity gates (<=1%). Cache-aware rerouting combined with trained locality achieves ~80% miss reduction at <=3.4% perplexity, outperforming standalone methods. Results show cacheability tax persists across scales, with 340M models exhibiting slightly increased costs.
mixture-of-expertscacheabilityrouter-telemetryperplexityedge-serving
Redakto - The Incognito Tab for LLMs
Redakto introduces an open-source tool for anonymizing text before LLM processing, addressing privacy concerns under EU legislation. It offers redaction and pseudonymization functionalities via web app, REST APIs, and model context protocol hooks, requiring modest compute resources. Evaluations on legal and medical texts show anonymized texts maintain utility scores comparable to originals, indicating minimal task impact for the studied domains.
privacy preservationpseudonymizationrest apimodel context protocolutility evaluation
Visual-Prompt Guided Wildlife Instance-Level Recognition
The authors propose a one-stage end-to-end model for wildlife instance-level recognition, combining detection and re-identification in a unified framework. The method integrates DINOv2 for spatial geometry and MegaDescriptor for re-identification, enhanced with visual-prompt guided latent queries. A detection decoder operates directly on the latent space to localize target identities. Preliminary results achieve 30.584% mean average precision, approaching the 44.89% benchmark of two-stage methods, with qualitative analysis confirming effective bounding box generation and identity matching.
instance-level recognitionvisual promptinglatent space queryingend-to-end detectionwildlife re-identification
Bidirectional representational alignment between biological and artificial neural networks
The study demonstrates that steering representational geometry during training improves bidirectional alignment between biological and artificial neural networks. Using a computational framework combining spectral regularization and bidirectional predictivity analyses, the authors evaluated self-supervised contrastive vision models. Spectral geometry steering increased reverse predictivity by 55% with modest forward predictivity reductions, achieving symmetric alignment at intermediate spectral exponents while reducing effective dimensionality and reorganizing the shared representational subspace.
representational alignmentspectral regularizationbidirectional predictivityeffective dimensionalitycontrastive learning
GenEx: A Graph-Based Representational Paradigm for SARS-CoV-2 Variant Detection via Codon Co-occurrence Networks
The authors propose GenEx, a graph-based pipeline for SARS-CoV-2 variant detection that models codon sequences as co-occurrence networks rather than linear strings. The method employs Multi-Scale Codon Co-occurrence Graph (MSCG) and Linear-time Adjacency PMI Codon Graph (LAPCG) to construct graphs, followed by spectral feature extraction using squared singular values (σ²) via SVD to enhance inter-class separability. Evaluated on 23 benchmarked ML models, GenEx achieves strong performance in distinguishing SARS-CoV-2 variants (Beta, Gamma, Delta, Omicron), demonstrating the efficacy of graph-based genomic representations.
codon co-occurrence graphspectral feature extractionsingular value decompositionsars-cov-2 variantsgraph-based representation
GigaBrain-WBC-0.5: A Behavior World Model for Robust Whole-Body Control with Environment Interaction
GigaBrain-WBC-0.5 introduces a Behavior World Model (BWM) for robust whole-body control in humanoid robots, addressing limitations of existing motion trackers that fail in environment-dependent scenarios. The method trains a causal Transformer to jointly predict next actions, states, and latent behavior commands, leveraging an automatic terrain-annotation pipeline to recover 3D contact geometry from retargeted motion. Results show superior performance: 81.3% success on terrain interaction (4.3x baseline), 83.1% under implausible commands, and 99.3% fall recovery (16.8x baseline), with successful hardware transfer between Unitree G1 and Maker L01 robots.
behavior world modelwhole-body controlcausal transformerterrain annotationfall recovery
On the Triangle Inequality for the Jaccard Distance in Arbitrary Lattices
The paper generalizes the Jaccard distance to arbitrary lattices with strictly positive, monotone, and modular valuations, proving it satisfies the triangle inequality without requiring distributivity. For relatively complemented distributive lattices, the triangle inequality holds under positive, monotone, supermodular, and log-submodular valuations. Necessary conditions are established, showing supermodularity is essential for the generalized Jaccard distance to be a valid metric. Applications in quantum information theory, formal concept analysis, and machine learning are highlighted.
jaccard distancelattice theorytriangle inequalitysupermodularitydistributive lattices
Bound-Aware Per-Organ Recall Risk Control for Multi-Organ CT Segmentation under Clinical Domain Shift
The paper introduces a distribution-free method for per-organ recall risk control in multi-organ CT segmentation under clinical domain shift. Using conformal risk control (CRC) and risk-controlling prediction sets (RCPS), the authors calibrate organ-specific thresholds for an AMOS-trained nnU-Net, auditing transfer to RAOS and estimating recertification costs via voxel false-negative rate (FNR). Results show AMOS control passes, but 7/12 organs exceed α=0.10 post-transfer; smaller calibration sets risk conservative thresholds. Waudby-Smith-Ramdas (WSR) bounds recertify six Tier-1 organs with 25 cases versus 30–40 for Hoeffding-Bentkus (HB), while CRC requires 10–15 cases but exhibits heavier tail risk. No Tier-2 organ meets precision criteria with 25 cases.
conformal risk controlfalse-negative ratennu-netdomain shiftrisk-controlling prediction sets
MotoSafety: Edge-AI with Learned Temporal Importance for Two-Wheeler Collision Risk Assessment Under Time Pressure
The paper introduces MotoSafety, an edge-AI framework for two-wheeler collision risk assessment under time pressure (TP), using Learned Temporal Importance (LTI) to process multivariate time-series data. The method leverages a novel dataset of 129,000 labeled samples across 153 simulator rides, capturing 64 attributes including motion, control, and compliance metrics. MotoSafety achieves 94.97% accuracy (99.33% ROC AUC) for risk assessment and 0.039 MSE for forecasting, with 1.15M parameters and 0.135 ms latency, enabling edge deployment on low-cost CPUs. It also demonstrates transferability to human activity (97.66%) and clinical (99.65%) domains.
edge-ailearned temporal importancecollision risk assessmenttime-series forecastingintelligent transportation systems
D$^2$ACCI: A Dual-Loop Diagnostic Protocol for Evidence-Preserving Agent Memory
The paper introduces D$^2$ACCI, a dual-loop diagnostic protocol for LLM agent memory systems that enables evidence-preserving iteration through stage-level fault localization. The method combines an outer diagnostic gate with graded observability metrics (DCR) and reusable evaluation artifacts (D$^2$ACCI-Eval), featuring protected-slice monitoring and trace-level analysis. Evaluation on LoCoMo (93.59%), LongMemEval (90.93%), and PersonaMem-V2 (57.20%) shows statistically significant gains from key components (+1.9 to +3.7pp, p ≤ .003), while diagnostic artifacts achieve 98-100% DCR@3 versus 0% for result-only logs.
llm agentspersistent memorydiagnostic protocolgraded observabilityfault localization
The Curious Case of Exploding DecPOMDPs: Containing the Fire through Policy Counting
The paper introduces policy-counted DecPOMDPs to address the exponential complexity of decentralized partially observable Markov decision processes (DecPOMDPs) in multi-agent systems. By shifting focus from agent counting to policy counting, the method achieves tractability through a compact representation that reduces model complexity and evaluation cost to polynomial dependence. The authors present policy-counted dynamic programming as an efficient solution approach for these policy-counted DecPOMDPs.
decpomdpsmulti-agent systemspolicy countingdynamic programmingtractability
Accuracy and Robustness of Model Cascades Under Data Perturbations
This paper analyzes the robustness of confidence-based model cascades for image classification under input data perturbations. The authors construct an energy-efficient cascade (10× lower CO₂ emissions) operating at the Pareto-optimum of accuracy, routing quality, and energy consumption. Through systematic evaluation under static corruptions and sequential perturbations, they identify three failure modes: (1) broken routing with preserved large-model utility, (2) dual-model degradation nullifying deferral benefits, and (3) suppressed deferral yielding stable but unreliable predictions. Results demonstrate that cascade reliability requires explicit evaluation under distribution shift beyond clean-data accuracy metrics.
model cascadesconfidence-based routingdistribution shiftinput corruptionsenergy efficiency
A systematic review of machine learning techniques to address diagnosis and treatment of autism: challenges and opportunities
This systematic review evaluates 55 studies (2017-2023) on machine learning applications for autism spectrum disorder (ASD) diagnosis and treatment. Supervised learning dominates current approaches, while deep learning gains traction with increased data availability. Key findings highlight the potential of hybrid methods (combining unsupervised learning, deep learning, and fuzzy logic) and multimodal data integration (genetic, clinical, wearable sensors) to improve diagnostic accuracy and enable continuous monitoring. The review identifies interdisciplinary collaboration and expanded ASD-specific datasets as critical needs for future research.
autism spectrum disordersupervised learningdeep learningmultimodal integrationwearable sensors
Lévy Attention: Single-Pass Predictive Uncertainty for Continuous-Time Attention
Lévy Attention introduces a stochastic cross-attention operator for continuous-time series that provides predictive uncertainty estimates at no computational overhead. The method formulates attention as a stochastic integral against an inhomogeneous Poisson random measure, where query-key compatibilities define an intensity and the output averages interpolated values at scattered atoms. The deterministic pass yields both expected attention (reducing to mollified cosine-kernel attention) and closed-form uncertainty metrics: evidence Λ_q and disagreement trΣ_V(q). Empirical results show the operator maintains accuracy (≤5.6% drop vs. softmax) while providing calibrated uncertainty, outperforming 20-pass MC dropout and achieving nominal coverage in conformal prediction tasks.
stochastic attentionpredictive uncertaintyinhomogeneous poissoncontinuous-time seriesconformal prediction
Learned, Then Lost: A Measured Single-Example Counterfactual in Pre-training
The study measures the transient impact of a single training example on model behavior through controlled counterfactual experiments. Using 32 GPT-2 (124M parameter) models trained on OpenWebText, the authors injected one of three text conditions (fluent corpus-attested prose, matched fabricated prose, or random characters) or no injection into a batch at step 200. The injected passage was learned temporarily (50 steps post-injection: +0.039–0.044 nats cross-entropy improvement, p < 1e-4) but decayed to undetectability by training conclusion (p ≥ 0.25). Weight displacement reached 44.1% of seed-to-seed variance, indicating relocation within the optimization basin without escaping it. Geometric measures (interpolation loss barrier, CKA) showed no significant differences across conditions.
counterfactual measurementtransient learningweight displacementoptimization basincross-entropy decay
Continuous-Time Reinforcement Learning for Controlled Hawkes Jump-Diffusions
The authors propose a continuous-time reinforcement learning method for controlling multivariate Hawkes-driven stochastic differential equations in non-Markovian settings. They develop a finite-dimensional Markovianization procedure to approximate Hawkes processes with exponential kernel mixtures, proving convergence of the approximation to the original problem. A model-free algorithm, Hawkes-CT DDPG, is introduced to solve the optimization problem using only observed event times and decay filters, outperforming discrete-time RL methods across exponential, Erlang, and power-law kernels.
hawkes processstochastic controlmarkovianizationcontinuous-time rlnon-markovian
Geometric Iterative Retrieval for Neural Audio Codec Resynthesis
The paper introduces geometric iterative retrieval, a novel paradigm for high-fidelity audio resynthesis from coarse Residual Vector Quantization (RVQ) tokens. Departing from discrete token prediction or continuous regression, the method leverages RVQ's hierarchical structure to perform contrastive retrieval in continuous codebook space, iteratively refining the output. Evaluated on speech and music codec restoration tasks, it outperforms single-pass token prediction and one-step regression baselines in fidelity.
residual vector quantizationneural audio codeccontrastive retrievalcodebook spacetoken-based generation
SCORE: Subject Coordinate Recovery for Label-Free Cross-Subject EEG-to-Image Retrieval
The paper introduces SCORE (Subject Coordinate Recovery), a label-free framework for cross-subject EEG-to-image retrieval that addresses performance degradation for new users. The method aligns source subject EEG with a common image space during training, simulates unseen-subject recovery, and at deployment estimates an orthogonal transformation using hubness-corrected landmark matching without target labels. SCORE achieves 53.23%/83.55% and 12.01%/32.16% Top-1/Top-5 accuracy on THINGS-EEG2 and Alljoined-1.6M benchmarks, outperforming baselines by up to 17.45 percentage points.
eeg-to-image retrievalcross-subject decodingcoordinate alignmenthubness correctionorthogonal transformation
Beyond Trial Averaging: Anchoring Neural and Visual Representations for Few-Repetition Brain-to-Image Retrieval
The paper introduces NEAR (neural-anchor-based retrieval), a framework improving few-repetition brain-to-image retrieval by addressing both query noise and gallery misalignment. NEAR uses a high-repetition center as anchor, applying a denoiser to queries and a small network to predict pseudo-anchors for gallery images. Evaluated on EEG, MEG, and fMRI datasets, NEAR increased 200-way Top-1 accuracy on THINGS-EEG2 by 5.7pp (1-rep) and 9.3pp (4-reps) compared to trial-averaging baselines.
brain-to-image retrievalneural decodingfew-shot learningrepresentation alignmentsignal denoising
Enhancing EBSD throughput of battery electrode materials using super-resolution generative adversarial networks
A super-resolution generative adversarial network (SRGAN) framework enhances electron backscatter diffraction (EBSD) throughput for LiNixMnyCozO2 cathode materials by computationally upscaling low-resolution datasets. The method outperforms classical interpolation, preserving grain boundaries and small grains across 2x-12x upscaling factors. At 5x upscaling (25x speed-up), relative errors were +5.7% (grain diameter), +8.2% (maximum inscribed diameter), and -14.6% (grain boundary length), enabling high-throughput microstructural analysis.
super-resolutiongenerative adversarial networkelectron backscatter diffractionlithium-ion batterymicrostructural characterization
Pretraining Reusable Inference Across Views with Synthetic Task Priors
The paper introduces SIMPLE, a task-conditioned multi-view in-context learner that pretrains reusable inference procedures across diverse view configurations. The method constructs a synthetic task prior in embedding space to generate varied support-query episodes with controlled variations in class structures, cross-view dependencies, and missingness patterns. A hierarchical architecture performs within-view, cross-view, and support-query reasoning. Experiments on multi-view and multi-omics benchmarks show that frozen SIMPLE achieves competitive performance without backbone updates, while adapter calibration attains leading results, demonstrating that multi-view reasoning can be pretrained and reused.
multi-view learningin-context learningsynthetic task priorhierarchical inferenceadapter calibration
Does Mapping Non-Maximal Probabilities to GMM Components Matter for S-JEPA Encoder Representations?
The study demonstrates that the mapping of non-maximal probabilities to Gaussian mixture model (GMM) components significantly impacts learned Encoder representations in S-JEPA, beyond just the probability values. Two controls were tested: FIXED-RANDPERM, which reassigns non-maximal probabilities with a fixed mapping, and UNIFORM-TAIL, which distributes non-maximal mass uniformly. REAL SOFT outperformed both controls across three seeds, showing better GMM tail recovery and spectral dynamics accessibility in frozen Encoder readouts. Results indicate that the numerical probability structure alone is insufficient; component mapping matters for representation learning.
gaussian mixture modelencoder representationsnon-maximal probabilitiesspectral dynamicsfrozen readouts
Learning Random Geometric Graphs Drawn in Probabilistic Metric Spaces
The authors propose a data-driven method for learning Random Geometric Graphs (RGGs) in probabilistic metric spaces, applicable to generic datasets regardless of observable type, distribution, or size. They introduce a disparity variable to define a distance function via its closed-form cdf, enabling edge existence probabilities (Soft RGG) and a rejection sampling-based edge probability estimation. The expected degree distribution is local and correlation-dependent, with a closed-form posterior for learning unknown correlation matrices. The method is demonstrated on highly multivariate real datasets.
random geometric graphsprobabilistic metric spacesdisparity variablerejection samplingdegree distribution
SPK: Eliciting Structured Prior Knowledge for Interpretable Out-of-Distribution Detection in Real-Time Object Detection
The paper introduces Structured Prior Knowledge (SPK), a framework for interpretable out-of-distribution (OoD) detection in object detectors by explicitly eliciting latent semantic priors. SPK leverages in-distribution data and hallucination-inducing samples to decode part-level semantic concepts, integrating them with geometric and contextual priors into a compact 5D representation. Experiments across diverse architectures (e.g., Faster R-CNN, YOLOv3) and benchmarks (e.g., OpenImages-OoD) show SPK achieves state-of-the-art OoD detection, revealing pretrained detectors encode richer latent knowledge than previously exploited. The approach enables structured, interpretable reliability analysis without modifying the base detector.
out-of-distribution detectionlatent priorsobject detectorssemantic conceptsinterpretability
Diffusion Models for High-Dimensional Clustered Data: Intrinsic-Dimension Adaptivity via Bayesian Classification
(No summary returned.)
Multi-Agent Off-Policy Deep Reinforcement Learning for Smart Campus Coverage
The paper proposes a multi-agent off-policy deep reinforcement learning framework for optimal millimeter-wave base station deployment in non-convex smart campus environments. Addressing the NP-hard max-min fairness problem, the authors benchmark four DRL approaches: single/multi-agent DQN and DDPG variants. Results show multi-agent DDPG achieves full coverage with 0.94 Jain's index fairness and superior computational efficiency in dense scenarios (400 users), outperforming single-agent methods.
deep reinforcement learningmillimeter-wavemarkov decision processmulti-agent systemsmax-min fairness
Monroe: A Molecular Foundation Model for In-Context Probabilistic Inference
Monroe introduces a molecular foundation model (MFM) with innovations including pre-training on 81M PM6 quantum chemistry molecules, enhanced stereochemistry graph representations, conformer denoising, and TabPFN-based in-context prediction. The model employs multi-task learning and embedding decorrelation, achieving state-of-the-art performance on Polaris benchmarks and significant gains on activity cliff tasks. Ablations demonstrate that TabPFN adaptation generalizes, improving MiniMol and CheMeleon to create MiniMol_PFN and CheMeleon_PFN variants.
molecular foundation modelin-context learningstereochemistryconformer denoisingtabpfn
Fuzzy Accuracy Compensates for Label Subjectivity in Classification of Skin Tone Using Wearable Photoplethysmography Signals
The study demonstrates improved skin tone classification from photoplethysmography (PPG) signals using fuzzy accuracy to compensate for subjective Fitzpatrick scale labeling. Three approaches were evaluated: deep learning/tree-based models on raw PPG signals (55% accuracy, 96% fuzzy accuracy), deep learning on Symmetric Projection Attractor Reconstruction (SPAR) images (44%, 85%), and feature-based machine learning (42%, 87%). A fuzzy cross-entropy loss function yielded optimal results, showing PPG signals contain skin tone information discernible within ±1 class of human raters.
photoplethysmographyfitzpatrick scalefuzzy accuracysymmetric projection attractor reconstructioncross-entropy loss
Transportable Causal Effect Estimation across Networks under Interference
The paper introduces TranCE, a doubly-robust algorithm for transporting causal effect estimates across networks under interference, addressing covariate and structural network shifts. The method employs selection diagrams to derive transport formulas for direct, spillover, and total effects, combining interventional outcome models, domain density-ratio correction, and cross-fitted inference. Experiments on semi-synthetic benchmarks (real-world social networks) and a weather-insurance field experiment demonstrate effectiveness, with transported effects validated against held-out randomized estimates.
network interferencecausal transportabilityselection diagramdoubly-robust estimationspillover effects
Lost in Aggregation: How Benchmarks Overlook Irreplaceable Model Strengths
The paper proposes evaluating tabular ML benchmarks via a data-centric peak performance frontier, identifying models as irreplaceable, sufficient, redundant, or fallible based on their best statistically supported performance per dataset. Analyzing TabArena, the authors show conventional aggregation metrics (averages, ranks, pairwise wins) correlate highly but poorly reflect irreplaceability, favoring consistently decent models over those with unique dataset-specific strengths. They advocate measuring benchmark progress both by aggregate improvements and expansion of attainable peak performances.
tabular benchmarkspeak performance frontieraggregation metricsirreplaceabilitydataset-specific strengths
Score the Algebra, Not the Span: Dimension Reduction for Transfer Operator Models of Dynamical Systems
The paper identifies linear masking, a failure mode in spectral dimension reduction for dynamical systems with weakly interacting components, where components may be entirely omitted from models. The authors propose scoring the generated σ-algebra instead of rank, enabling efficient representation of component interactions through algebraic operations. Their χ²-divergence criterion guarantees embedding with twice the intrinsic dimension suffices to capture the full operator spectrum. Experiments show rank-based methods miss masked components at k<100, while 10 algebra coordinates recover them, enabling better prediction from few labels compared to high-dimensional or VAMP-feature regression.
linear maskingtransfer operatorσ-algebraχ²-divergencevamp score
Simple, Safe, and Overlooked: Reclaiming Sustainable Domain Generalization with Statistical Color Matching
The paper introduces Colorist, a training-free domain generalization method using statistical color matching in RGB space to address medical image classifier failures caused by hardware and color variations. The approach applies global mean-standard deviation matching, preserving anatomical structures while outperforming deep generative models in structural fidelity (+9% accuracy over SOTA domain generalization methods, +13% over baseline). It demonstrates efficacy across histopathology, peripheral blood, dermatology, and retinal datasets, offering interpretability and computational efficiency without neural network overhead.
domain generalizationstatistical color matchingdata augmentationmedical imagingstructural fidelity
Converting Expert Deliberation into Financial Signals Through A Context-Aware NLP Pipeline
The paper introduces CDSP, a context-aware NLP pipeline that converts investment committee meeting transcripts into structured predictive features for financial markets. The method segments transcripts into topical chunks, applies LLM-based asset-class labeling, maps financial keywords to a taxonomy, and extracts sentiment polarity and mention frequency features. Evaluated on 48 monthly meetings predicting equity vs. bond performance, CDSP features combined with sentence embeddings achieved 73% accuracy (F1=0.73), outperforming the 60.4% baseline of always choosing stocks. Sentiment features proved more informative than mention frequency for several taxonomy categories.
context-aware nlpsentiment polaritymention frequencyasset-class labelingfeature engineering
On the Slow Convergence to Trivial Solutions of Algorithms for Hard Optimization Problems
The paper reveals slow convergence to trivial solutions in algorithms for hard combinatorial optimization problems, challenging negative asymptotic predictions. Through rigorous analysis of large-graph asymptotics and numerical experiments on maximum independent set and maximum $K$-SAT, it demonstrates that local algorithms outperform theoretical bounds in intermediate constraint-density regimes. This finite-size behavior gap underscores the practical relevance of sophisticated algorithm design despite asymptotic failure predictions.
combinatorial optimizationnp-hardaverage-case analysislocal algorithmsconstraint density
A FEM-Based Surrogate Modelling and Optimization Framework for Physics-Constrained Electromagnetic Coil Design
The study presents a surrogate-assisted optimization framework for electromagnetic coil design under physical constraints, using a FEM-GP workflow with sequential Bayesian optimization. A Python–MPh–COMSOL pipeline couples a 2D axisymmetric FEM model with a Matern 5/2 Gaussian-process surrogate, evaluating designs under geometric, manufacturing, and mass constraints. Results from five paired runs show method-dependent performance: EI–BO excels at small budgets, COBYLA leads early, and BOBYQA achieves highest terminal response, though no method dominates across all metrics.
surrogate modelingbayesian optimizationfinite-element methodelectromagnetic designgaussian process
Quantum Tensor Network Learning with DMRG
The study introduces a global normalization condition for matrix product states (MPS) to represent quantum states in tensor network learning. Two optimization methods are investigated: gradient descent and an adaptation of the density matrix renormalization group (DMRG). The effectiveness of these methods in finding locally optimal tensors is compared, leveraging techniques from quantum many-body physics simulations. Results indicate that both methods achieve local optimization, with DMRG adaptation showing particular promise for quantum state representation.
matrix product statequantum stategradient descentdensity matrix renormalization grouptensor network
Graph-Based Approaches to Learning Epileptogenic Zone Localization Using Stereo-EEG Recordings
The study systematically evaluates graph topology choices for epileptogenic zone (EZ) localization from resting-state stereo-EEG, proposing Region-Bridge-$c$ as a sparse alternative to dense connectivity graphs. Using leave-one-patient-out validation on 40 patients, the authors compare dense graphs, anatomical/geometric priors, budgeted sparsification, and learned sparsification while controlling edge counts. Region-Bridge-$c$ achieves 0.371±0.015 PR-AUC (0.743±0.010 ROC-AUC) at 30% edge retention, outperforming dense graphs (0.349±0.014 PR-AUC) with 69% fewer edges. Results show topology performance varies by patient, suggesting graph construction should be treated as an explicit modeling choice rather than fixed preprocessing.
epileptogenic zonestereo-eeggraph topologyfunctional connectivitysparsification
Sharper Regret Bounds for Time-Varying Gaussian Process Bandits with Constant Exploration
(No summary returned.)
Multi-stage neural operator learning with application for convolutions
The paper introduces two multi-stage neural operator learning frameworks for convolution integrals: Deep Collocation Neural Operator (DCNO), a supervised method that iteratively refines approximations using input-output pairs, and Deep Galerkin Neural Operator (DGNO), an unsupervised approach leveraging PDE residuals. Both methods progressively enrich the approximation space through multiple training stages, achieving near-machine precision accuracy for convolution problems and efficiency gains over traditional solvers. Theoretical analysis and numerical experiments validate their performance, with extensions to multi-input scenarios involving density and kernel variations.
neural operator learningconvolution integralsmulti-stage trainingpde residualssupervised learning
GEAR: Generative Expansion and Real Anchoring for Two-Stage Distillation of Tabular Foundation Models
The paper introduces GEAR, a two-stage distillation framework for tabular foundation models (TFMs) that produces lightweight MLP or tree-based predictors deployable on CPUs. Stage 1 trains students on synthetic covariates with soft TFM targets, while Stage 2 re-anchors them to real data distributions using out-of-fold teacher predictions to avoid leakage. Experiments on TALENT and TabArena show AUC improvements of 1.19–2.40 points over baselines, with 57–2866× faster inference and 1.9–3.3× lower memory usage while outperforming non-TFM methods like CatBoost.
tabular foundation modelsknowledge distillationin-context learningsynthetic data generationlightweight predictors
A Unifying Relational Perspective on Expressive Lottery Tickets
The paper generalizes the Strong Expressive Lottery Ticket Hypothesis (SELTH) to multi-relational and temporal domains via relational Weisfeiler-Leman (RWL). It proves that sufficiently parameterized relational GNNs (RGNNs) contain sparse subnetworks preserving 1-RWL expressivity, derives a probabilistic lower bound for random pruning, and shows temporal GNNs (TGNNs) inherit these guarantees. Experiments validate the bound on synthetic data and demonstrate pre-training expressivity's relation to optimization and prediction metrics on molecular and temporal benchmarks.
graph neural networkslottery ticket hypothesisweisfeiler-lemanrelational learningsparse networks
Many Optimizers But Only One Training Path: Repeated Resampling for Adaptive Optimizer Selection
The paper introduces Repeated Optimizer Resampling (ROR), a method for adaptive optimizer selection during neural network training without requiring multiple complete runs. ROR evaluates candidate optimizers via short scouting phases (s epochs) from the current weights every b epochs, continuing with the best-performing optimizer for the remaining b-s epochs. Evaluated on MNIST, Fashion-MNIST, and insurance claim models, one-epoch ROR reduces training cost to 24-35% of exhaustive search while matching the performance of the best fixed optimizer across tasks.
adaptive optimizationneural network trainingscouting phaserepeated resamplinghyperparameter search
Tensor Field Models
The paper introduces Tensor Field Models (TFMs), mathematical structures where a learned operator maps admissible component-section families to time-dependent tangent sections on a generative state manifold. TFMs encode analytic and dynamical restrictions through admissible families, with structured refinements via constructed, component-separable, and tensor bundle variants. Conditional realizations map structured conditions componentwise to reusable representations, combined by a field operator to generate vector fields. Trained using flow matching, experiments demonstrate TFMs improve performance and enable faster generation through amortized sampling with reusable condition representations.
tensor field modelsgenerative state manifoldflow matchingcomponent-separableamortized sampling
A Real-Time Tsetlin Machine-based Non-intrusive Load Monitoring System on MCUs
The paper proposes a Tsetlin Machine-based Non-intrusive Load Monitoring (NILM) system for real-time edge deployment on microcontrollers, addressing privacy concerns by processing data locally. The method reformulates NILM as a classification task, achieving 90% precision and 96% recall for two-appliance classification, and 77% precision with 80% recall for four appliances on the REDD dataset. The model requires only 18 KB of flash memory and achieves 0.43 ms inference latency on an ESP32, demonstrating efficient embedded implementation.
tsetlin machinenon-intrusive load monitoringmicrocontrollersedge computingenergy disaggregation
GraphK: Variable-Size Graph Generation with Efficient Edge Construction
GraphK introduces an encoder-sampler-decoder framework for variable-size graph generation, enabling both upscaling and downscaling of node counts while maintaining structural fidelity. The method learns permutation-invariant latent representations, samples new node embeddings via maximum likelihood estimation, and constructs edges efficiently using KDTree-based top-k neighbor search in latent space. Experiments on synthetic and real-world datasets demonstrate superior performance over existing methods in learning graph structures and generating synthetic graphs without explicit definitions.
graph generationpermutation-invariantlatent representationskdtreemanifold smoothness
MIFR: A Modality-Invariant and Fair Representation Framework for Skin Disease Classification
The paper introduces MIFR, a modality-invariant and fair representation framework for skin disease classification that jointly addresses modality dependence and skin-tone bias. The method employs ViT-based encoders with modality-specific projection heads, trained via a five-component loss combining classification, fairness (confusion and skin-type classification), class alignment (supervised contrastive loss), and modality-invariance objectives. Evaluated on HIBA+Derm7pt, PAD-UFES-20, and ISIC 2019, MIFR achieves competitive accuracy and fairness, with t-SNE confirming geometric alignment of clinical and dermoscopic embeddings.
modality-invariantfair representationvit-based encoderssupervised contrastive lossskin disease classification
To Go Far, Go Together: Diverse Preferences Induce a Curriculum for Reward Optimization
The paper introduces CurriPO, a curriculum-based method for optimizing diverse user reward models in AI alignment. It addresses the challenge of underserved users with hard-to-optimize reward models by constructing a tree-structured curriculum that reuses previously incorporated models. Experiments on personalized continuous control show CurriPO achieves 1.2–2.1× population satisfaction over baselines while reducing training time, particularly benefiting underserved users.
reward optimizationcurriculum learningai alignmentpersonalized controlmulti-user structure
Gradient Mirage: Trainable yet Label-Unidentifiable Gradients in Large Language Model Split Learning
Gradient Mirage defends against gradient matching attacks (GMAs) in LLM split learning by breaking gradient-objective consistency while preserving optimization utility. The method induces inconsistency across three dimensions: (1) Selective Autoregressive Supervision uses a masked surrogate loss, (2) Scale Blinding applies randomized rescaling, and (3) Directional Privatization randomizes gradient direction via von Mises-Fisher mechanism under differential privacy. Dual-Track Backpropagation and Bottom-Gradient Recovery maintain training efficacy. Experiments show superior privacy-utility trade-offs compared to existing defenses, with comparable fine-tuning performance.
gradient matching attackssplit learningdifferential privacyvon mises-fisherautoregressive supervision
Learning Canonical Register Automata over Ordered Data Domains
The paper presents a unified framework for active learning of deterministic register automata (DRAs) over both dense and non-dense ordered data domains, including rationals and integers. The method employs polynomial-time algorithms using membership, equivalence, and memorability queries, extending prior work on domains with identity tests. Key results include decidability of DRA minimization over integers and improved complexity bounds for related decision problems.
deterministic register automataactive learningordered data domainsmemorability queriespolynomial-time algorithms
Enhancing Distance-Based Graph Autoencoders with Structural Penalties for Dynamic Graph Embedding
The paper proposes three distance-based graph autoencoder (GAE) variants incorporating structural penalties to address node-level heterogeneity in dynamic graphs. Methods employ a GCN encoder and Euclidean decoder, augmenting sparsity-corrected loss with two regularization terms: (i) hub penalty based on degree centrality and (ii) NC-LID-based penalty motivated by prior evidence linking high NC-LID to poor embeddings. Experiments on dynamic graphs demonstrate NC-LID regularization consistently outperforms baseline and hub-aware variants in reconstruction tasks, validating NC-LID as an effective structural signal for distance-based GAEs.
graph autoencoderdynamic graph embeddingstructural regularizationnc-liddistance-based reconstruction
Geometric Data Perturbation with Noisy-Anchor Alignment for Privacy-Preserving Collaborative Learning
The paper proposes noisy-anchor alignment for Geometric Data Perturbation (GDP) in privacy-preserving collaborative learning, addressing vulnerabilities under analyst-participant collusion. Unlike prior methods that either degrade utility (independent transformations) or enable exact data recovery (shared-anchor alignment), the approach perturbs only anchor representations while preserving private-data transformations. The solution involves solving a Generalized Orthogonal Procrustes Problem for alignment and provides theoretical analysis of alignment/recovery errors and attack resilience. Experiments on MNIST and CelebA demonstrate superior privacy-utility trade-offs compared to private-data noise, with higher accuracy at equivalent leakage levels under evaluated collusion scenarios.
geometric data perturbationcollaborative learningorthogonal procrustesprivacy-utility tradeoffrepresentation alignment
Decision-Metric Alignment in Latent World Models: Diagnostics and Action-Conditioned Objectives for MPC Planning
The paper introduces decision-metric alignment, a property ensuring latent-world-model costs (e.g., Euclidean distance) rank action sequences by real task progress. It proposes Plan-Real Spearman and CEM-stage Spearman to diagnose alignment gaps, identifying encoder distortion, terminal rollout error, and candidate margins as key factors. DA-LeWM, an augmented latent world model with inverse-dynamics and demonstration-conditioned heads, improves alignment and achieves higher online success than LeWM in experiments, while maintaining similar probe scores. Results demonstrate that action-conditioned objectives enhance Euclidean-cost MPC planning.
decision-metric alignmentlatent world modelsmodel-predictive controlcross-entropy-methodinverse-dynamics
Regularised Iterative Generalised Least Squares with Optimal Selection of the Hyper-Parameter for Identifying Nonlinear Phenomenological Models
The authors propose a ridge regression scheme for estimating parameters in nonlinear phenomenological models with confounded parameters, common in fields like lithium-ion battery state of health prediction. The method incorporates an automated hyper-parameter optimization using information-theoretic measures, solved via fixed-point iteration, and integrates this within a regularized iterative generalized least squares framework for handling heteroscedastic and serially correlated data. Simulation results confirm the method's efficacy in improving numerical conditioning while preserving model structure.
ridge regressionhyper-parameter optimizationphenomenological modelsgeneralized least squaresfixed-point iteration
FedLNS: Leverage LayerNorm Signature Modeling to Mitigate Adversarial Manipulation in Federated LLMs
FedLNS introduces a server-side framework for detecting malicious updates in federated language model training by leveraging normalization-layer parameter changes as signatures. The method screens suspicious updates against a cross-client reference without requiring additional client-server communication, raw data, labeled attacks, or a trusted dataset. Evaluated on GPT-style, BERT-style, and LLaMA-style models with 200 clients under 40% target manipulation, FedLNS achieves lower test perplexity than six baselines across IID and non-IID data partitions.
federated learningnormalization signaturemalicious updatestest perplexityclient-server communication
Visual-Aware Representation of Web Pages for Machine Learning Applications
The paper introduces a visual-aware representation platform for machine learning over web pages, leveraging the open-source FitLayout rendering tool. The system captures visual and structural properties in RDF format, supports REST API control and SPARQL queries, and integrates with Python ML workflows. Demonstrations include graph-based representations for training graph neural networks to recognize key content elements, showcasing reproducibility and applicability. The architecture enables complete dataset preparation from rendering to downstream learning tasks.
visual-aware representationrdf-based representationgraph neural networksweb page renderingsparql queries
Multi-Class Electrical and Mechanical Fault Classification Using Random Convolutional Kernels
The work introduces a multivariate extension of SelF-Rocket, a random convolutional kernel-based method for multi-class fault diagnosis in rotating machinery. The approach is evaluated against leading ROCKET-based methods on MaFaulDa (mechanical faults) and ITSC-UDG (stator inter-turn short circuits) datasets under univariate and multivariate settings. SelF-Rocket achieves the best accuracy-latency trade-off, yielding the highest classification performance on MaFaulDa while remaining competitive on ITSC-UDG.
random convolutional kernelstime series classificationfault diagnosismultivariate extensionrotating machinery
A single design choice determines whether machine learning models of materials make physically impossible predictions
The study identifies a critical design choice determining whether machine learning models of materials predict physically impossible properties: the inclusion of parity labels in feature representations. Using group theory, the authors derive a parity gap criterion to assess which properties and crystals are susceptible to such errors. Evaluations on 2,000 centrosymmetric crystals reveal that parity-labelled models achieve near-zero predictions for forbidden piezoelectric tensors, while rotation-only models predict nonzero values in 90-96% of cases, differing by six orders of magnitude. Training with explicit zeros fails to recover exactness, and frozen backbone symmetry persists in downstream heads. Verification requires only one reflection at initialization.
parity labelsgroup theorycentrosymmetric crystalspiezoelectric tensorsymmetry group
Transforming Heart Disease Prediction with Advanced Machine Learning Techniques
This study advances heart disease prediction by systematically evaluating nine machine learning classifiers on two datasets (UCI and Kaggle) containing 14 cardiac health attributes. Techniques including J48, Naive Bayes, Logistic Regression, Simple Cart, Bagging, Decision Stump, AdaBoost, Artificial Neural Networks, and Support Vector Machine (SVM) were assessed using metrics such as Mean Absolute Error, Relative Absolute Error, accuracy, precision, recall, and F-measure. Results demonstrated that SVM achieved optimal performance on the UCI dataset (highest accuracy, lowest error), while Simple Cart performed best on the Kaggle dataset. The findings indicate that properly tuned ML models can significantly enhance early heart disease diagnosis and clinical decision support.
support vector machinemean absolute errorartificial neural networkslogistic regressiondecision stump
An Empirical Benchmark of Deep Time-Series Models for Smart Meter Energy Forecasting
This paper presents a systematic benchmark of nine deep learning architectures for smart meter energy forecasting, evaluating performance across input length, prediction horizon, and model complexity. The study tests linear, MLP-based, convolutional, and Transformer models on two public datasets, analyzing accuracy-computation tradeoffs and demographic subgroup performance. Key findings show: (1) increasing historical context improves accuracy until saturation (~7 days), (2) longer horizons degrade performance, (3) deep models outperform classical baselines, and (4) architectural differences matter most for long horizons on heterogeneous data, with lightweight models offering competitive accuracy at lower cost.
time-series forecastingsmart meterdeep learningcomputational complexityprediction horizon
Reinforced Planning with Latent World Models
The paper introduces Reinforced Planning (RP1), a novel method that learns both plan evaluation and multi-step plan improvement through offline training on imagined world-model rollouts. RP1 combines a learned critic for outcome evaluation with an optimizer for plan refinement, operating independently of any pretrained latent world model. Evaluations across visual navigation, arm reaching, and robotic manipulation tasks demonstrate RP1's superiority over hand-designed search algorithms, achieving near-perfect success rates while using 1,000× fewer world-model rollouts and being up to 67× faster during concurrent planner inference.
reinforced planninglatent world modelsmulti-step plansoffline trainingplanner inference
Computational Measurement of Team-Process Phase Dynamics in Collaborative Virtual Reality
The article introduces a computational framework for analyzing dynamic team-process phases in collaborative VR using timestamped dialogue. The method employs late chunking for context-aware transcript representations, penalized Gaussian-kernel change-point detection for semantic transitions, and TF-IDF with NMF for phase interpretation, validated by LLM-generated and human-reviewed labels. Results demonstrate coherent phase structures with traceable transcript evidence and alignment with interaction logs, supporting the framework's utility for temporal teamwork analysis.
collaborative virtual realitychange-point detectionnon-negative matrix factorizationlate chunkingteam-process phases
Quantum-Logic Tsetlin Machines: Interpretable Quantum Machine Learning with Commuting Projector Clauses
The Quantum-Logic Tsetlin Machine (QL-TM) bridges Tsetlin Machines with quantum logic by replacing Boolean literals with commuting quantum projectors while retaining classical automata. Clauses activate via Born probabilities of joint projectors, with exact Boolean TM reduction proven for computational-basis contexts. Experiments on Bell states, stabilizer tasks, and context-budget ablations demonstrate that correct non-diagonal contexts recover physical clauses, while diagonal contexts lose phase/syndrome information. Results align with the predicted separability ladder 2^(b-k) as stabilizer generators are removed.
quantum-logic tsetlin machinecommuting projectorsborn probabilitystabilizer semanticscontext-budget ablation
FlashAttention for Scalable Vector Architectures
FlashAttention-V introduces a blocked FlashAttention variant optimized for scalable vector architectures, addressing memory bandwidth bottlenecks in transformer inference on CPUs. The method exploits parallelism across attention heads, employs inter-head packing for efficient vector length utilization beyond head dimensions, and enhances register utilization and memory locality. Evaluated on TinyLlama, Llama 3.2, Qwen2.5, and Pythia-410M using gem5 and Banana Pi BPI-F3, FlashAttention-V achieves 22x-42x speedup in prefill and 8x-11x in decode at 512-bit VL over scalar FlashAttention, with diminishing returns at higher vector widths due to memory-bound execution. Structural bottlenecks in Q8_0 quantized linear layers are identified as limiting factors for long-vector scalability.
flashattentionvector architecturesmemory bandwidthquantization formatstransformer inference
Progressive Experience Fusion for Multi-Task World Model Control in Endovascular Navigation
Proposes Progressive Experience Fusion (PEF) for multi-task control in endovascular navigation, combining TD-MPC2 with adaptive-horizon Model Predictive Path Integral planning and patient-specific fine-tuning. Evaluated across five subtasks in ten vascular anatomies, PEF achieved 74% mean success (vs. 37% for Soft Actor-Critic, p<0.001) and 90% in held-out vasculatures with adaptive planning. In vitro transfer to an unseen stroke patient improved path ratio from 63% to 80% after 40×10³ fine-tuning steps (≈107 min simulated transfer time), demonstrating feasibility for multi-vasculature training and clinical adaptation.
progressive experience fusiontd-mpc2model predictive path integralendovascular navigationpatient-specific fine-tuning
ProxyGuard: Direct Reliability Inference for Randomized Data Release Mechanisms with Shared Targets
ProxyGuard introduces two modes for reliable inference in randomized data release mechanisms with shared targets. Named-release mode corrects multiplicity and certifies specific releases, while direct shared-target mode evaluates independent mechanism draws on a common target, providing finite-sample reliability guarantees without requiring independent target batches. The method includes a sharp mean-only penalty and a smooth-score certificate with additive target concentration. In a registered study, direct mode increased power from 5.6% to 64.2% at reliability 0.95, with named mode performing better under high-signal evidence. Applications include full-pipeline Rice--TVAE and non-tabular text mechanisms.
randomized data releasereliability inferencemultiplicity correctionfinite-sample guaranteeshared targets
Coordination on a Budget: Federated Active Learning with Few Labels
The paper introduces a federated active learning (FAL) framework addressing label scarcity and data privacy in low-budget regimes, demonstrating a 'heterogeneity reversal' phenomenon where IID data requires stronger coordination than heterogeneous data. The method employs federated representation learning to align client data in a shared embedding space, enabling globally coordinated query selection while preserving local annotation. Evaluations show the framework outperforms existing FAL methods even with smaller annotation budgets, highlighting the efficacy of centralized coordination under privacy constraints.
federated active learningheterogeneity reversallow-budget regimerepresentation learningquery selection
Scalable Geospatial Machine Learning for Power-Line Asset Risk: Integrating Remote Sensing for Lightning and Vegetation Risk Modelling
The study presents a scalable geospatial ML framework for power-line asset risk assessment, featuring a modular architecture that accommodates new environmental data sources and failure modes without pipeline redesign. The method integrates multi-source predictors (SRTM topography, MODIS NDVI, LIS VHRMC lightning climatology, OpenStreetMap proximity features, and utility records) into a unified pipeline for vegetation- and lightning-related failure probability modelling. Results demonstrate computational efficiency and operational extensibility, enabling utility-scale deployment for risk-based asset management and climate resilience planning.
geospatial machine learningprobability-of-failure modellingmulti-source predictorsasset-level risk stratificationutility-scale deployment
Off-Manifold Collapse in Guided Protein Language Models
The paper identifies an off-manifold collapse phenomenon in guided protein language models, where optimized sequences degenerate to low-complexity, non-foldable forms despite scoring well on target properties. The authors propose Mahalanobis filtering, a post-hoc method that detects collapse by comparing guided activations to natural protein activation statistics. This training-free approach improves both property scores and structural plausibility across guidance methods, validated by maintaining typical activation densities. The activation statistics are publicly released.
protein language modelsoff-manifold collapsemahalanobis filteringguided inferenceactivation statistics
Infrared Universality of Collective Dynamics across Transformer and State-Space Architectures
The study demonstrates that Transformer and Mamba architectures develop similar near-marginal slow-mode dynamics despite differing microscopic mechanisms. By analyzing Mamba's relaxation dynamics at three levels—intrinsic state-space spectrum, selective rescaling, and full-block Jacobian—the authors show selective dynamics reorganize relaxation hierarchies but yield a reproducible slow-mode continuum. Full-block spectra exhibit power-law scaling $ρ(λ)∼λ^β$, with Mamba ($β_{\rm M}≃-0.17$) and Transformer ($β_{\rm Tr}∼-0.1$) converging to marginal $1/t$ memory decay. Results extend infrared universality beyond Transformers, supporting Cognitive Field Theory's dynamical framework.
infrared universalitycollective dynamicsstate-space modelsmamba architectureslow-mode continuum
Compress and Forget: bitsandbytes Quantization Amplifies Proactive Interference in LLMs
The study demonstrates that 4-bit quantization via bitsandbytes significantly exacerbates proactive interference (PI) in large language models, reducing retrieval accuracy under high interference (e.g., Qwen2.5-7B-Instruct drops from 81.0% to 68.3%). Evaluating FP16, INT8, and INT4/NF4 precision levels across Qwen2.5-7B-Instruct, Mistral-7B-Instruct-v0.3, and Phi-3.5-mini-instruct, the authors find INT4 increases same-key intrusion errors (21.5% to 24.6%, p=4.8e-7) and show the effect originates in the quantized transformer backbone. INT8 also incurs smaller penalties in two models, challenging assumptions of its safety for PI-sensitive tasks.
proactive interferencepost-training quantizationbitsandbytesintrusion errorstransformer backbone
Beyond receptive fields: sequence-pooled normalization can supply most of a sequence labeler's context
The study demonstrates that sequence-pooled normalization layers in convolutional sequence labelers can provide global context beyond the model's receptive field, contradicting conventional assumptions. Through Jacobian analysis and synthetic experiments, the authors show that normalization statistics create a sequence-spanning path that supplies nearly optimal context (within 0.009 of whole-sequence performance) for tasks with long label runs. Ablation studies reveal this path confounds attribution, overstating receptive field contributions by 8.3-16.1×. The effect diminishes with frequent label switches, highlighting a misattribution of performance gains in traditional analyses.
sequence-pooled normalizationconvolutional sequence labelerreceptive fieldjacobian analysisablation study
Continual Reasoning Gym: Diagnosing and Harnessing Shared Reasoning in Continual RLVR
The paper introduces Continual Reasoning Gym, a benchmark for continual reinforcement learning with verifiable rewards (RLVR) across five task sequences combining text and visual reasoning. It identifies shared reasoning as a transferable structure that mitigates forgetting but finds sequential RLVR still underperforms multitask RL (MTRL). The proposed Continual Prompt Replay (CPR) method replays previous-task prompts and regenerates responses with the current policy, achieving MTRL-level performance by leveraging shared reasoning.
continual rlvrshared reasoningmultitask rlprompt replayverifiable rewards
NanoSleep: A Parameter-Efficient Hybrid Temporal Convolutional Network for Single-Channel Sleep Stage Classification
NanoSleep introduces a parameter-efficient hybrid temporal convolutional network for single-channel EEG sleep stage classification, addressing deployment constraints on resource-limited devices. The architecture combines a Sinc-convolutional front end, dual-branch multi-scale feature extraction, a gated dilated TCN with channel recalibration, and conditional random field decoding, optimized via weighted calibrated focal loss. Evaluated on Sleep-EDF and Sleep-EDF-Expanded datasets, it outperforms six baselines in subject-wise cross-validation, with ablation studies validating component contributions. The model achieves accuracy-efficiency trade-offs suitable for wearables and clinical applications.
temporal convolutional networksleep stage classificationsinc-convolutionchannel recalibrationconditional random field
Learning Topological Features of $\widehat Z$-invariants
(No summary returned.)
MARCUS: Missing-Aware Region Representation with Contextual Urban Signals for Rent Prediction
The paper introduces MARCUS, a missing-aware region representation model that treats data absence as contextual urban signals rather than noise. The method employs a three-stage architecture: Intra Learning for joint encoding of observed features and missing patterns, Inter Learning for modality reliability estimation to guide cross-modal interaction, and Fusion with missing-aware and time-aware gating for final embedding generation. Evaluated on rent prediction tasks in Sydney and New York, MARCUS reduces MAE by 51.35% and 12.62% respectively compared to baselines, with ablation studies confirming its robustness to missing data.
region representationmissing-awarecross-modal interactionrent predictionurban signals
LLM-Powered Predictive Decision-Making for Sustainable Data Center Operations
We introduce an LLM-based predictive scheduling system for sustainable data center operations, reducing energy consumption and queuing delays. The system employs an LLM to predict execution time and energy consumption from source code, with potential extensions to metrics like water usage and carbon emissions. A real-time scheduling algorithm allocates GPU resources, optimizing energy efficiency and task throughput. The framework demonstrates fast inference, generalization across task types, and minimal training data requirements. In collaboration with a data center, the system achieved a 32% reduction in energy consumption and a 30% decrease in waiting time, advancing sustainability in AI-driven infrastructure.
llmpredictive schedulingenergy consumptiongpu allocationsustainability
Tianmu-TC: Physics-constraints Generative Artificial Intelligence for Global Tropical Cyclone Forecasting
We introduce Tianmu-TC, a physics-constrained generative AI framework for global tropical cyclone forecasting that addresses computational inefficiency and forecast uncertainty in numerical weather prediction and deep learning models. The framework incorporates physics constraints to generate controllable outputs, trained on Western North Pacific data. Tianmu-TC outperforms deterministic and ensemble meteorological AI models and ECMWF NWP systems across global ocean basins, demonstrating superior performance in data sparsity, anomaly tracks, and rapid intensification scenarios with significantly reduced computational costs. Results indicate that physics-constrained generative AI provides a reliable and efficient approach to tropical cyclone forecasting.
physics-constrained generative aitropical cyclone forecastingnumerical weather predictionensemble meteorological modelscomputational efficiency
WhiteMatter: All-to-All Cross-Layer Connections via KV Mixing
WhiteMatter introduces all-to-all cross-layer connections in Transformers via dynamic KV mixing, enabling each attention layer to attend to representations from all layers of past tokens. The method employs a router that mixes L layer states into k KV channels (k ≤ L) for caching, with consumer layers attending to one channel each, reducing KV-cache memory footprint when k < L. Pretraining experiments show WhiteMatter outperforms a vanilla Transformer with 50% more layers and retains most gains even with 50% KV-cache compression.
transformerkv-cacheautoregressiveroutercross-layer
Atrial Fibrillation Detection with Arbitrary Leads via a Codebook-Based Reconstruction-Classification Framework
The paper introduces DCGCNet, a vector-quantized variational autoencoder for robust atrial fibrillation (AF) detection from ECG signals with arbitrary lead configurations. The model combines a Local-Global Contrastive Module for noise-invariant representations and an Adaptive Codebook Vector Quantizer to prevent codebook collapse, enabling joint AF classification and ECG reconstruction. Evaluations show state-of-the-art performance (AUC > 0.98) in both intra-dataset 12-lead settings and cross-dataset generalization across seven scenarios, with maintained accuracy under noisy conditions like baseline wander and EMG artifacts.
atrial fibrillation detectionvector-quantized autoencoderecg reconstructioncodebook collapsenoise-invariant representations
Adaptive Multi-Agent Feature Selection for Personalized Fall Risk Prevention
The paper introduces PAFIR, a Personalized and Adaptive Feature selection framework for fall risk Identification and pRevention, addressing limitations of static approaches in modeling evolving risk factors. PAFIR formulates feature selection as a reinforcement learning problem over longitudinal multimodal health data, jointly modeling structural dependencies among assessment variables and temporal dynamics in wearable-derived activity data. Applied to the PEER trial data, PAFIR outperforms baselines in capturing longitudinal feature relevance patterns and enables dynamic, subject-specific feature selection for timely fall prevention.
adaptive feature selectionreinforcement learninglongitudinal multimodal datapersonalized fall preventionwearable-derived activity
The Embodiment Gap in Robot Foundation Models
The survey introduces the concept of the 'embodiment gap' in robot foundation models (RFMs), highlighting the discrepancy between reusable models and their practical deployment on specific robot bodies. It categorizes existing methods along two axes—shared structure type and adaptation stage—and analyzes three research directions: shared semantics/perception, shared robot data/interfaces, and cross-embodiment correspondence learning. A reporting framework is proposed to evaluate adaptation work beyond success rates, identifying unresolved challenges and future research questions.
robot foundation modelsembodiment gapcross-embodiment learningvision-language-action policiesadaptation framework
Process Optimization Under Uncertainty for Improving the Bond Quality of Polymer Filaments in Fused Filament Fabrication
The paper presents a computational framework for optimizing fused filament fabrication (FFF) process parameters to maximize inter-filament bond quality under uncertainty. The method combines transient heat transfer analysis with a sintering neck growth model, employs Sobol indices for variance-based sensitivity analysis, and uses Gaussian process surrogate modeling to account for model discrepancy. Experimental validation confirms that the uncertainty-aware optimization yields superior bond quality compared to deterministic approaches.
fused filament fabricationgaussian processsobol indicesprocess optimizationuncertainty quantification
Multi-Objective Optimization Under Uncertainty of Part Quality in Fused Filament Fabrication
The paper presents a data-driven multi-objective optimization method for fused filament fabrication (FFF), simultaneously minimizing geometric inaccuracy and maximizing filament bond quality under parameter uncertainty. A Bayesian neural network (BNN) models process parameter effects while capturing epistemic (model) and aleatory (input) uncertainties. Robust optimization formulations explore tradeoffs between objectives, with Pareto surfaces quantifying performance compromises. Experimental validation confirms the BNN's predictive accuracy and optimization effectiveness for nozzle temperature, speed, and layer thickness parameters.
bayesian neural networkmulti-objective optimizationfused filament fabricationepistemic uncertaintypareto surface
The Road Taken: The Role of Optimizers at the Edge of Stability
The authors propose a reformulation of the edge of stability (EOS) phenomenon in deep learning optimization, addressing systematic violations of classical stability bounds by first-order methods. They derive a new stability threshold based on the directional Hessian and gradient-alignment score, accounting for optimizer-specific update directions rather than maximum curvature. This approach eliminates optimizer-dependent offsets, improves stability predictions, and provides diagnostic tools revealing optimizer roles in balancing temporal and spatial optimization budgets. Experiments show violations of classical bounds by up to 21.1×, demonstrating the method's consistency across optimizers.
edge of stabilityhessian eigenvaluegradient alignmentfirst-order optimizationdirectional hessian
Role-Conditioned Sub-Token Routing for Efficient Vision-Language-Action Policies
The paper introduces Role-Conditioned Sub-Token Routing (RoleSub), a method for efficient compression of Vision-Language-Action (VLA) models by reducing value representations within retained tokens. RoleSub partitions token values into orthogonal groups and uses a lightweight router conditioned on token representation, latent roles, and language context to selectively preserve groups. Evaluated on OpenVLA-OFT-7B across LIBERO suites, RoleSub outperforms token-only compression in 33/36 settings, achieving 9.2-11.3% KV cache size while maintaining control performance. The approach demonstrates complementary benefits to token pruning for aggressive VLA compression.
vision-language-actionsub-token routingkv-cachemultimodal compressionorthogonal partitioning
Algorithms for adaptive and heteroskedastic linear regression at the computational threshold
The paper introduces novel algorithms for finite-sample linear regression under heteroskedastic and adaptive noise models. For heteroskedastic regression, where label noise variances σ_i² are unknown, the authors propose a polynomial-time estimator achieving error rate Õ((nd³/m⁴)^(1/6)) when m ≫ d^(3/4)n^(1/4), with matching lower bounds. In adaptive regression, where noise follows an unknown mixture of k symmetric log-concave densities, they present an inefficient estimator matching the optimal oracle rate with Õ(n/k) samples, and a polynomial-time L_q regression variant for k=1. Computational thresholds are explored via a planted linear regression problem, suggesting an information-computation gap between m=d+1 and m∼d^(3/4)n^(1/4).
heteroskedastic regressionadaptive linear regressionfinite-sample analysisinformation-computation gapstatistical query lower bounds
Inference and Uncertainty Quantification for Streaming $r$-PCA
The paper resolves two open problems in streaming PCA via Oja's algorithm: achieving sharp operator-norm convergence for general rank-r under sub-Gaussian data, and providing distributional inference for subspace estimators. The authors develop a convergence analysis without remainder terms, yielding a minimax-optimal rate (up to log factors) in dense-tail spiked covariance regimes, and prove a matching lower bound under mild nondegeneracy conditions. Key innovations include a linearization of Oja's iterates enabling high-dimensional Gaussian approximation for subspace error, and an online multiplier bootstrap for practical inference. Results extend beyond streaming PCA to nonconvex stochastic approximation.
streaming pcaoja's algorithmgaussian approximationsubspace estimationmultiplier bootstrap
When Does Dynamic Ensembling Pay Off? Diagnosing Regionwise Gains in Regression under Distribution Shift
The paper introduces $\widehat{D}_{\mathrm{CF5}}$, a diagnostic metric that predicts when dynamic (input-dependent) ensembling of regression models outperforms static blending under distribution shift. The method estimates cross-fitted gains of regionwise convex combinations over static blends using small labeled target-domain probes. Evaluated on 12 dataset-shift pairs, $\widehat{D}_{\mathrm{CF5}}$ achieves Spearman +0.98 correlation with realized test gains (95% CI [+0.83, +1.00]), significantly outperforming alternative diagnostics. A Probe-Validated Ensemble Selector leverages this to deploy dynamic ensembles only when beneficial, reducing test risk by 11-16% in successful cases while avoiding catastrophic failures.
dynamic ensemblingdistribution shiftregionwise convex combinationprobe validationregression ensembles
SingularClip: Preventing Spectral Collapse to Maintain Plasticity in Continual and Reinforcement Learning
The paper introduces SingularClip, a method to prevent loss of plasticity in neural networks by addressing growing anisotropy in weight matrices' singular values during training. The authors identify spectral collapse as a novel source of plasticity loss and propose periodic singular value clipping to maintain matrix conditioning. Empirical results demonstrate SingularClip's effectiveness across continual supervised learning and deep reinforcement learning tasks, outperforming baseline methods.
plasticity lossspectral collapsesingular value clippingcontinual learningreinforcement learning
Self-supervised In-context Operator Learning for Stochastic Mean-Field Control
The authors propose NFIST, the first mesh-free, self-supervised neural operator for stochastic mean-field control (MFC), enabling zero-shot generalization across tasks. They address diffusion-term challenges via a normalizing-flow-based transformer that recasts dynamics as a deterministic continuity equation, achieving O(d) cost per particle. The method combines probability-flow ODEs with in-context learning, requiring no precomputed solutions. Experiments on optimal control, Schrödinger bridge, and path planning demonstrate effective generalization, reducing computational costs for MFC families.
stochastic mean-field controlnormalizing flowoperator learningin-context learningfokker-planck equation
A Comprehensive Review of Large Language Models for Nanophotonics: From Surrogate Modeling to Autonomous Design
The review systematically evaluates how Large Language Models (LLMs) enhance nanophotonic design workflows by transitioning from task-specific neural networks to transformer-based universal reasoning. It categorizes LLM applications into two modes: (1) surrogate models framing structure-spectrum mapping as a language task, and (2) agentic systems generating code, orchestrating simulations, and enabling closed-loop optimization. The analysis extends to cross-disciplinary opportunities in materials science and wireless communications, proposing future multimodal foundation models with physical perception for autonomous scientific collaboration.
large language modelsnanophotonicssurrogate modelingtransformer architecturesautonomous design
Model Card for OpenAI Privacy Filter
The OpenAI Privacy Filter introduces a bidirectional token-classification model for detecting and redacting personally identifiable information (PII) and secrets in unstructured text. Derived from an autoregressively pretrained checkpoint, the model employs banded-attention classification and a constrained Viterbi decoder to label sequences across eight privacy categories. With 1.5B total parameters, 50M active parameters per token, and a 128K-token context window, it supports configurable precision-recall tradeoffs for efficient local deployment and domain-specific fine-tuning.
token-classificationbanded-attentionviterbi decoderprivacy categoriesdata-minimization
Rethinking Privileged Information in On-Policy Self-Distillation
The study re-evaluates performance gains in on-policy self-distillation (OPSD) by analyzing whether improvements stem from learning privileged reference information or recovering existing reasoning behaviors in the base model. Experiments on science and mathematics datasets with Qwen3 models (1.7B–8B) separate supervision induced by reference information from base-model supervision, measuring alignment with student predictions. Results show inconsistent benefits from correct references, with students improving even without them and alternative solutions outperforming correct ones. Alignment with base-model reasoning dominates, and controls replicate much of the observed alignment, suggesting performance gains and distributional alignment alone cannot confirm reference information's role in OPSD.
on-policy self-distillationprivileged informationtoken-level supervisionreasoning behaviordistributional alignment
Global Crises and National Policies: A Large Scale Analysis of Political Content in German Language Online Media
The study introduces an automated method for analyzing political bias in German-language online media during 2019-2022, leveraging modern political text analysis techniques on millions of articles and tweets. It examines coverage of COVID-19 and the Ukraine war, revealing thematic convergence in international event reporting between German and Swiss outlets, with divergence in domestically focused content. Newspapers showed stable political discourse, while Twitter exhibited event-driven spikes. Findings suggest international crises synchronize media content across platforms and national boundaries, offering transparency tools to counter algorithmic echo chambers.
political text analysisalgorithmic biasmedia consumptioncontent synchronizationecho chambers
SIGMA: Symmetry-aware, Intelligent, Geometric, Multi-objective Adaptive Control for Robust, Dependable Traffic Management
SIGMA introduces a reinforcement learning framework for adaptive traffic signal control, integrating a large language model (LLM) for dynamic objective tuning and orientation-invariant learning. The method converts natural-language emergency commands into priority vectors for a multi-objective actor-critic controller, employs rotational augmentation for transferability across intersections, and utilizes offline-to-online learning for stable initialization. Reliability properties, including emergency service levels and graceful degradation under LLM failures, are validated via bootstrap statistics. Evaluated in SUMO on four urban intersections, SIGMA outperforms fixed-time, actuated, and DQN controllers, reducing waiting times, queue lengths, and boosting throughput while demonstrating robustness to component failures and geometric rotations.
reinforcement learninglarge language modelmulti-objective controlrotational augmentationoffline-to-online learning
ClosureBench: A Constructive Benchmark for Compositional Graph Reasoning
ClosureBench introduces a constructive benchmark for compositional graph-relational reasoning with programmatically verified ground truth, generating instances on demand via the Ein tensor-logic language. It evaluates models from 1.5B to frontier systems, revealing a 19.3 percentage-point memorization gap in fine-tuned models and showing accuracy degradation with increasing graph size and query depth. A 4B model fine-tuned for program synthesis achieves 94.3% accuracy on held-out instances, demonstrating the efficacy of verified program synthesis across languages.
compositional reasoninggraph-relationalprogram synthesisbenchmarkingmemorization gap
Sobolev Regularized Score Difference Estimation in Diffusion Models
The authors propose a Sobolev-regularized estimator for Stein score differences in diffusion models, addressing limitations of existing methods in statistical consistency and scalability. The method leverages Sobolev regularization to ensure consistency and stability in small-sample regimes, achieving a convergence rate of $O(n^{-\frac{s-1}{d+2s-2}})$ for $d$-dimensional data with smoothness $s$. Theoretical analysis includes a minimax lower bound of $\tilde{\Omega}(n^{-\frac{2(s-1)}{d+2s}})$. Empirical validation on ECG signal generation demonstrates superior stability and downstream classification performance compared to non-regularized estimators.
stein scoresobolev regularizationdiffusion modelstransfer learningminimax lower bound
Classifying Directional Trajectories Near Criticality in the Three-State Majority-Vote Model with Deep Belief Networks and Bidirectional GRUs
The study demonstrates a hierarchical deep learning approach for classifying directional trajectories in the three-state majority-vote model (MV3). Using a Deep Belief Network (DBN) for unsupervised encoding of static equilibrium samples and a Bidirectional Gated Recurrent Unit (Bi-GRU) for temporal sequence analysis, the method achieves near-perfect separation of four trajectory types. Results show the Bi-GRU's hidden state space effectively discriminates between approach/departure from order/disorder, with real-time applicability in detecting dynamical regimes.
deep belief networkbidirectional grumajority-vote modelcritical transitionsagent-based modeling
Allocating Recurrent Compute in Looped Language Models
The paper introduces MixerLoop, a method for selectively applying recurrence in looped language models by repeating only Gated DeltaNet mixer layers while executing dense feed-forward networks (FFNs) once. This approach is motivated by Iterative Transport Rank (ITR) and marginal ITR, which quantify cross-position influence trajectories and nonredundant contributions. Experiments at 15M and 110M parameters show MixerLoop outperforms full-block recurrence (FullLoop) on aggregate CORE at 15M and retains 41.5% of CORE improvement at 110M while reducing recurrent-backbone FLOPs by 45.9%. Results demonstrate that recurrent depth benefits can be achieved without redundant FFN computation.
looped language modelsiterative transport rankgated deltanetrecurrent computationfeed-forward networks
Think Shallow, Solve Deep: Controlling Recurrent Dynamics for Reliable Test-Time Depth
The paper establishes a dynamical systems framework for analyzing test-time depth behavior in recurrent-depth reasoners, demonstrating that finite-time dynamical regimes (settling, marginal, drifting) predict whether additional iterations improve, preserve, or degrade performance. The authors derive a sufficient depth-safety condition based on per-step displacement relative to decoder margin and show empirically that settling operators maintain or improve accuracy with increased depth (e.g., Sudoku performance improves from 0.19 to 0.34 beyond training horizon). They validate the framework through controlled experiments on algorithmic tasks and analyze Huginn-3.5B as a non-settling case study.
recurrent-depth reasonersfinite-time dynamical regimedepth-safetydecoder margintest-time depth
Primitive Representation Learning for Unsupervised Dynamic Contrast Enhanced MRI Reconstruction
The authors propose a primitive-based unsupervised framework for dynamic contrast-enhanced MRI reconstruction that disentangles anatomy, contrast enhancement, and motion into separate temporal basis functions. The method employs multi-dimensional Gaussian and Gabor primitives to achieve scan-specific reconstructions without requiring large training datasets. Results demonstrate competitive performance with conventional methods in both reconstruction quality (PSNR/SSIM) and accuracy of extracted aorta/kidney enhancement curves, while offering geometric interpretability and extensibility to higher acceleration rates.
dynamic contrast-enhanced mriprimitive representation learninggabor primitivestemporal basis functionsunsupervised reconstruction
ChiroEcho: extending automated bat vocalisation classification beyond the learned taxonomy
The paper introduces ChiroEcho, a deep learning framework for automated bat vocalization classification that extends taxonomic coverage beyond trained species by combining genus-level predictions with geographic distributions. The method jointly predicts species and genus, using geographic constraints to resolve species absent from the training taxonomy when only one species of a predicted genus occurs locally. Evaluated on 35 European bat species, the approach increases operational coverage from 73% to 85% (35 to 41 species) and demonstrates how limited evaluation data can obscure rare-species performance.
automated classificationdeep learningecholocation callstaxonomic resolutiongeographic constraints
Safe Domain Adaptation for Physics: Overcoming Nuisances, Label Shifts, and Simulation Priors
The paper introduces adaptive domain adaptation (ADA) to address limitations of standard adversarial domain adaptation in physics applications, where both nuisance parameters and target quantity distributions may differ between simulation and experimental domains. ADA reweights simulated events to isolate genuine physical mismatches during adaptation, avoiding bias from simulation priors. Experiments on a toy air-shower benchmark demonstrate that standard methods erroneously align spectral shifts, while ADA preserves target distributions. A label-free model selection criterion is proposed for optimal operational point selection without ground-truth labels.
domain adaptationadversarial learningsimulation biasspectral shiftmodel selection
rl-triton: High-Performance Triton GPU Kernels for Reinforcement Learning Credit Assignment
The paper introduces rl-triton, an open-source library of high-performance GPU kernels for reinforcement learning credit assignment, implemented using Triton. The key innovation is a unified associative scan framework that generalizes seven RL estimation algorithms (GAE, V-Trace, Retrace(λ), TD(λ) returns, discounted returns, eligibility traces, and episodic prefix sums) as instances of a first-order linear recurrence solvable in O(log T) parallel steps. The framework uses algorithm-specific fused Triton kernels to construct recurrence coefficients on-chip, verified algebraically. Benchmarks demonstrate 1.6-5.70× speedup over a torch-compile baseline in massively parallel simulations, with greater gains at longer sequences due to reduced HBM round-trips.
associative scangpu kernelscredit assignmenttritonreinforcement learning
What Can Artificial Intelligence Learn from Medicine? Generative Analogies and Reliable Machine Learning Systems
The article proposes a generative analogy between clinical translation in medicine and machine learning (ML) system development, arguing that epistemic and methodological standards from medicine can inform ML reliability. Using tools from Hesse's work, it characterizes this analogy and identifies specific warrants of clinical translation applicable to ML. The authors reinterpret these warrants in reliabilist terms, proposing a novel form of ML reliabilism distinct from existing philosophical accounts in AI.
generative analogyclinical translationepistemic warrantsmethodological standardsml reliabilism
📰 Industry Media (10)
Debates over AI consciousness are a trap
The article critiques the emerging discourse on AI consciousness as a corporate liability shield, arguing that anthropomorphic framing distracts from accountability for AI harms. It analyzes how both regulatory advocates (e.g., Hassabis, Amodei) and philosophical proponents (e.g., MacAskill) converge on treating AI as autonomous agents, potentially enabling legal personhood claims. Through case studies of AI-related lawsuits and policy conflicts, the author demonstrates how this narrative could undermine product liability frameworks. The analysis concludes that consciousness debates serve corporate interests by obfuscating negligence in AI development.
anthropomorphic framinglegal personhoodproduct liabilitymoral outsourcingglobal workspace theory
Unlocking hidden revenue streams with market models
Generative AI-powered market models enable dynamic pricing and revenue management by analyzing complex, real-time market variables. These deep learning systems integrate high-resolution numerical data (demand, capacity, competitor positioning) to simulate multi-faceted market environments, outperforming static rule-based approaches. Virgin Atlantic reports improved decision granularity and speed in pricing engines, leveraging the model's ability to process hundreds of interdependent commercial factors.
generative aimarket simulationdynamic pricingrevenue managementdeep learning
Liquid AI Releases LFM2.5-DSpark Draft Models That Deliver Up to 3.18x Faster Decoding Without Changing Model Outputs
Liquid AI introduces DSpark draft models (295.7M–327.7M parameters) for speculative decoding with LFM2.5 family targets (1.2B–8B parameters), achieving up to 3.18x faster decoding on H100 GPUs without altering outputs. The method combines a parallel backbone (DFlash), sequential Markov-chain head, and confidence-scheduled verifier to propose and validate 9-token blocks per forward pass. Results show speedups vary by task (1.04x–3.18x), with identical greedy-decoding accuracy and 57% latency reduction in multi-tool agent scenarios, though MoE models exhibit limited gains (1.18x) on Apple M4 Max due to implementation constraints.
speculative decodingdraft modelgreedy decodingkv headsblock size
Auditing Preference Biases and Fine-Tuning Language Models with Direct Preference Optimization on Anthropic HH-RLHF Using TRL and LoRA
The study presents a Direct Preference Optimization (DPO) pipeline for fine-tuning the Qwen2.5-0.5B-Instruct model on the Anthropic HH-RLHF dataset, addressing preference biases through lexical and structural audits. Methodologically, it employs TRL and LoRA for efficient adaptation, analyzes length-based biases via TF-IDF diagnostics, and evaluates reward accuracy across HH-RLHF subsets (helpful-base, helpful-rejection-sampled, helpful-online, harmless-base). Results include a 0.5B-parameter model fine-tuned with β=0.1, achieving measurable separation between chosen and rejected responses (AUC >0.5) despite detected lexical shortcuts.
direct preference optimizationhh-rlhf datasetlora adaptationlexical shortcut diagnosticsreward accuracy
Amazon’s Prime Air autonomous drones to reach 500 US cities
Amazon's Prime Air drone delivery service is set to expand to nearly 500 US cities by 2026, leveraging autonomous flight software and a Detect-and-Avoid system for real-time obstacle detection and navigation. The drones operate under FAA Part 135 certification, ensuring safety in adverse conditions and enabling rapid scaling without individual waivers. The service delivers items under 5 pounds within 30-60 minutes, with a tiered fee structure based on Prime membership. Current operations span 10 metro areas, with plans to extend to Chicago, Syracuse, and other cities. The UK site marks the first international expansion.
autonomous flight softwaredetect-and-avoid systemfaa part 135 certificationprime airreal-time navigation
Stripe agrees to buy OpenRouter as AI model routing expands
Stripe's acquisition of OpenRouter integrates model routing into its AI billing infrastructure, enabling developers to access over 400 models from 80+ providers via a unified API. OpenRouter employs dual-layer routing, selecting models based on task complexity, price, speed, and reliability, while also optimizing provider endpoints using latency, throughput, and cost metrics. The platform processes over 10 trillion tokens daily, supporting failover, data-handling constraints, and enterprise-specific requirements. This acquisition aligns with Stripe's token-based billing system, which tracks usage metrics like input/output tokens and inference costs. Multi-model environments are increasingly prevalent, with 52% of organizations orchestrating an average of seven models.
model routingtoken-based billingfailoverlatency optimizationmulti-model environments
Agentic AI in government just hit the hard part: deciding what a machine may decide
The United Arab Emirates (UAE) has initiated a strategic program to deploy agentic AI across 50% of federal government operations within two years, marking the first large-scale governmental implementation of autonomous decision-making systems. A workshop involving 100+ officials established implementation priorities, performance indicators, and classification frameworks to determine which tasks can be delegated to AI agents versus those requiring human oversight. Key challenges include defining liability for autonomous system errors and operationalizing the 'human leads, AI enables' principle across seven strategic pillars, with 80,000 employees undergoing training.
agentic aiautonomous systemsgovernment operationsdecision classificationstrategic frameworks
A third of ChatGPT ads appear in irrelevant conversations
The study reveals ChatGPT ads' relevance mismatch, analyzing 11,000 ads in real conversations from July to August 2026. The method pairs each ad with its conversation, grading relevance in three bands. Results show 33% ads unrelated, 27% direct match, and 40% contextual match. Two thirds of ads land in conversations with no purchase intent.
relevance mismatchcontextual matchpurchase intentdirect matchunrelated ads
AI data centre regulation just got a template that needs no new law
Pennsylvania introduced a regulatory framework for AI data centers via Executive Order 2026-05, leveraging existing permitting authority without new legislation. The GRID Requirements mandate developers to sign consent agreements covering energy costs, local hiring, water conservation, and prohibit nondisclosure agreements (NDAs), ensuring public disclosure of energy/water usage. The order prioritizes residential power over data centers during grid strain and removes projects from fast-track permitting. Over 100 projects are tracked publicly, with 5 fully permitted. This approach contrasts with 27 states considering large-load legislation, offering a replicable template for AI infrastructure regulation.
data center regulationpermit fast-tracknondisclosure agreementsenergy curtailmentinfrastructure consent
HoneyBook bets on agentic AI to streamline small business operations with its new Claude connector
HoneyBook introduces MCP, a connector integrating Anthropic's Claude AI with its SME-focused CRM platform to automate business operations. The agentic AI system enables querying customer records (e.g., lead status, invoice tracking) and executing actions (e.g., contract generation, payment reminders) via natural language. Early adopters reported hours saved on administrative tasks, with 68% of SMEs in a Bluevine survey anticipating AI benefits. The solution employs sandboxed sessions for data privacy, addressing adoption disparities where only 29% of sub-$100M firms deploy AI operationally versus 50% of enterprises.
agentic aicrm integrationsandboxed executionsmall business automationnatural language interface
Generated automatically at 2026-08-20 19:55 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.