Daily Digest — 2026-09-12
281 items · 1 research labs, 272 arxiv papers, 8 industry media
🏛️ Research Labs (1)
Rapidly scaling online storage to serve over 1 billion ChatGPT users
OpenAI presents Habitat, a Python-based distributed storage platform scaling to support 1B+ weekly users and 70M+ requests/sec. The system evolved from a client library to a centralized service, addressing coordination challenges across 40+ regions while managing 500PB+ data. Key innovations include asyncio optimization for tail latency (reducing p99 delays from 100s ms to <10ms), metastable failure mitigation via LIFO connection reuse tuning, and strategic technical debt acceptance to prioritize API stability over premature optimization.
distributed storageasyncio optimizationmetastable failuretail latencypython service scaling
📜 arXiv Papers (272)
GPU-CFR: 80x Faster Counterfactual Regret Minimization by Compiling the Game to Static Dataflow and CUDA Graph Replay
GPU-CFR introduces a compiler and runtime that accelerates counterfactual regret minimization (CFR) by compiling game trees into static dataflow representations, enabling efficient GPU execution. The method precomputes indices, batches depth-level passes, and employs CUDA Graph Replay to minimize framework overhead, reducing operations by up to 18.1x. Evaluated on an eight-game suite, GPU-CFR achieves 29.8--80.4x speedup over prior GPU implementations and 14--258x over LiteEFG on CPUs, with bitwise accuracy maintained. The compiled representation also outperforms CPU baselines by 2.2--51.1x without GPU acceleration.
counterfactual regret minimizationstatic dataflowcuda graph replaydepth-level batchingkernel launch overhead
General Quantification of Covariate and Concept Shifts
(No summary returned.)
Can Edge-Deployable Vision-Language Models Identify Species?
The study evaluates the taxonomic identification capabilities of edge-deployable vision-language models (VLMs) in the 2-8B parameter range, comparing them against the domain-specialist BioCLIP (300M) on a 96-species task using clean iNaturalist photographs and camera-trap imagery from LILA.science collections. Results show all models perform significantly above chance but degrade sharply on field imagery (domain gaps of 9.6-26.6 percentage points), indicating image legibility issues rather than fine-grained discrimination failure. BioCLIP outperforms all VLMs by 33.2-59.2 percentage points, suggesting specialized training data is crucial. Open-set prompting reveals 5.9-9.6% of responses are syntactically valid but taxonomically nonexistent species names, with consistent fabrication-rate rankings across evaluation sets.
vision-language modelstaxonomic identificationdomain gapedge-deployableopen-set prompting
Generative Marketing Mix Modeling: A Causal Inference Framework Linking GEO and GEM to Business Impact
The authors propose Generative Marketing Mix Modeling (GMMM), a causal inference framework for estimating the business impact of Generative Engine Optimization (GEO) and Generative Engine Marketing (GEM). GMMM integrates generated answer frequencies, question counts, system usage shares, notice probabilities, and sponsored placement records to quantify causal effects. The method establishes sufficient identification conditions for treatment sequences and compares expected business responses across alternatives. Empirical evaluation is conducted using simulated product recommendation data in English and Japanese, demonstrating the framework's applicability across languages.
causal inferencegenerative engine optimizationnotice probabilitiestreatment sequencesmarketing mix modeling
Artificial Id: Drive and Persistent Alignment in Agentic AI
The paper proposes an artificial id, an adaptive internal drive mechanism for agentic AI systems to autonomously determine behavioral continuance, cessation, or modification without external task-specific objectives. Through minimal virtual experiments, the authors demonstrate that a small controller, devoid of general reasoning capabilities, develops emergent control via differential persistence—selecting and replacing strategies based on environmental persistence. Results indicate that adaptive direction can arise without explicit behavioral objectives, though persistent misalignment risks emerge. The work argues for persistent alignment boundaries encompassing trusted observations, consequence channels, and hard constraints in scalable agentic systems.
agentic aiadaptive drivedifferential persistencepersistent alignmentconsequential state
MindTopo: Can Foundation Models Reason in Topological Space?
MindTopo introduces a benchmark evaluating topological reasoning in foundation models, focusing on five cognitive properties: continuity, separation, order, enclosure, and knots. The benchmark assesses reasoning and planning tasks across 11,030 instances in 13 procedurally generated task types. It evaluates 14 multimodal large language models (MLLMs) and agent configurations augmented with image and video generation. Results show that all MLLMs perform better on reasoning than planning, with supervised fine-tuning and reinforcement learning improving reasoning more than planning. Generated observations retain local cues but fail to reliably follow environment dynamics or preserve topology across transitions.
topological reasoningfoundation modelsmultimodal large language modelsprocedurally generated tasksenvironment dynamics
Domain-Specific Hallucination Detection in Large Language Models
A multi-signal hallucination detection pipeline is introduced, combining fine-tuned DeBERTa-v3 classification, Monte Carlo Dropout uncertainty quantification, and temperature-scaled calibration. The pipeline achieves F1=0.915 and AUROC=0.977 on the HaluEval benchmark, with task-specific F1 scores of 0.97 (QA), 0.96 (Summarization), and 0.82 (Dialogue). MC Dropout inference improves accuracy to 93.2%. Context ablation confirms entailment reasoning, with summarization F1 dropping 24% without knowledge context. Direct Preference Optimization reduces hallucination rates in Qwen2.5-0.5B from 85.5% to 37.7%. Domain-specific fine-tuning on SciFact with PubMedBERT yields F1=0.63 and AUROC=0.81, highlighting the importance of domain-matched pre-training.
hallucination detectionmonte carlo dropouttemperature-scaled calibrationdirect preference optimizationdomain-specific fine-tuning
Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens
The authors introduce AssayBench-Loop, a benchmark for adaptive hit discovery comprising 1,389 CRISPR screens across five phenotype categories, enabling systematic evaluation and learning of acquisition strategies. They propose AssayLoop, a sequential experimental design framework combining AssayFormer, a transformer-based amortized acquisition policy trained on historical screens, with LLM-derived biological priors via adaptive handoff. Additionally, AssayLLM extends this principle through task-specific post-training. On temporally held-out screens, AssayLoop achieves 5.67-fold enrichment over random selection and recovers 27.7% of hits after assaying ~5% of the candidate library, outperforming existing methods and standalone LLMs. Performance improves with more training data and transfers to excluded phenotype categories.
crispr screensadaptive hit discoverytransformer-based acquisitionbiological priorsassaybench-loop
On the Regularization Landscape for the Linear Recommendation Models
The paper unifies state-of-the-art linear recommendation models under a single framework, revealing that their comparable performance stems from either nuclear-norm or Frobenius-norm regularization. Nuclear-norm regularized models exhibit rigid, low-rank structures with closed-form solutions but limited expressiveness, while Frobenius-norm regularized models offer greater expressiveness and efficiency but require complex numerical procedures like ADMM. The authors propose two novel low-rank, closed-form solutions that generalize Frobenius-norm regularizers, combining the benefits of both approaches. These solutions enhance predictive power while maintaining computational tractability in recommendation tasks.
nuclear-norm regularizationfrobenius-norm regularizationlow-rank solutionsclosed-form solutionsrecommendation models
The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement
The paper proposes a roadmap for achieving genuine recursive self-improvement (RSI) in AI systems, progressing through improvement-execution autonomy, improvement-strategy autonomy, experience-acquisition autonomy, environment-adaptation autonomy, and recursive meta-improvement. Using the Headroom-Closed Index (HCI), the authors identify limitations in current large language models (LLMs) and analyze RSI requirements across diverse scenarios, including scientific discovery, embodied intelligence, and software engineering. Drawing on industry practices and preliminary empirical evidence, the study connects RSI research to practical systems and highlights key challenges in realizing persistent, self-directed capability enhancement.
recursive self-improvementheadroom-closed indeximprovement-execution autonomylarge language modelsenvironment-adaptation autonomy
RetroThinker: Enabling Retrospective Thinking in Speech LLMs
RetroThinker introduces a multi-stage post-training framework enabling streaming SpeechLLMs to dynamically revise reasoning traces during inference, addressing the accuracy-latency trade-off in real-time spoken interaction. The method combines supervised fine-tuning on retrospective thinking data with length-based direct preference optimization to optimize concurrent reasoning. Evaluated on the GSM8K benchmark, RetroThinker achieves an 11% absolute accuracy improvement over non-retrospective baselines while maintaining comparable latency, demonstrating enhanced reasoning capabilities in SpeechLLMs.
speechllmsretrospective thinkingdirect preference optimizationsupervised fine-tuninggsm8k
Explainability Assistant: A Conversational XAI Interface for Interpreting Energy Consumption Models
The Explainability Assistant introduces an open-source conversational XAI system leveraging modern LLMs' function-calling capabilities to interpret energy consumption models. It replaces rigid custom grammars with flexible natural language interaction, achieving 94% intent-parsing accuracy without task-specific fine-tuning. Comparative evaluation with energy domain specialists demonstrates improved usability and consistent task accuracy, with unanimous preference for the conversational interface over traditional XAI dashboards.
explainable artificial intelligenceintent-parsing accuracyfunction-callingnatural language interactionenergy consumption models
From Parameters to Answers: How LLMs Retrieve and Use Their Internal Knowledge
This work investigates how language models (LLMs) dynamically retrieve and utilize internal knowledge during question answering, focusing on query-routing information and target knowledge dependencies. Through layerwise interventions on hidden states in Qwen, Llama, and Gemma, the study analyzes country-continent questions with varied answer types. Results reveal distinct routing-content profiles: Qwen exhibits a causal window for pair-conditioned request direction, Gemma shows mid-layer routing overlaps, and Llama lacks sustained routing effects. Dependence on global request direction decreases across layers while fitted content dependence persists, separating early readability from later content reliance.
layerwise interventionsquery-routinghidden statecausal windowfitted content
Model-Aware Schedules Improve Generation via Fiberwise Optimal Transport
The paper introduces model-aware schedule construction for diffusion and flow-matching models via fiberwise optimal transport, addressing limitations of model-agnostic kinetic action. By defining a fiberwise prediction risk that averages optimal-transport costs between true and predictor-induced decompositions, the method derives closed-form optimal time allocations. Evaluations on DDPMs and flow matching across datasets and architectures demonstrate consistent improvements, including a 38.6% relative FID reduction on CIFAR-10 at 16 function evaluations. Empirical universality is observed in normalized-risk profiles across models, and pretrained-checkpoint diagnostics extend these findings to larger models.
fiberwise optimal transportdiffusion schedulesflow matchingprediction riskkinetic action
Understanding Operator Attitudes Toward AI-Supported Decision Making in Maritime Operations
This study investigates maritime stakeholders' attitudes toward AI-supported decision assistants in collision-avoidance scenarios through a survey combining established questionnaires, sentiment analysis, and thematic analysis of open-ended responses. Participants evaluated technology anxiety, trust in automation, and explanation quality across scenarios. Results indicate positive disposition toward maritime technology, age-independent openness, stable trust, and scenario-sensitive explanation ratings. Open responses highlighted valued aspects (decision support, situation awareness, confidence-building) and concerns (AI reliability, over-reliance, expertise loss). Findings suggest maritime AI systems should prioritize calibrated reliance through transparent, reliable, and operationally meaningful design with domain expert involvement, rather than solely increasing automation or trust.
maritime autonomous surface shipscollision-avoidance scenariostechnology anxietytrust in automationcalibrated reliance
Logit Refiner: Improving Visual Autoregressive Models via Intra-Scale Dependency Modeling
The Logit Refiner improves Visual Autoregressive Models (VAR) by addressing their mean-field-style approximation in parallel decoding, which discards spatial dependencies among same-scale tokens. This lightweight autoregressive module sequentially samples tokens conditioned on frozen backbone features, restoring intra-scale dependencies. Adding only ~10% parameters and <5% of the base model's training compute, it integrates into pretrained VAR checkpoints without retraining. Across VAR backbones (310M to 2B parameters) on class-conditional ImageNet 256x256, the refiner consistently enhances generation quality, enabling a 1.1B-parameter model to outperform a 2B-parameter one. The method also generalizes to text-to-image generation, alleviating the mean-field bottleneck across VAR variants.
visual autoregressive modelsmean-field approximationintra-scale dependencieslogit refinerparallel decoding
Thinking with Looped Flows
Looped flows introduce a novel approach to training recurrent models by leveraging local denoising objectives with temporal association, enabling effective transfer of computation across updates despite limited gradient coverage. The method integrates the velocity of a probability flow parameterized by a learned denoiser, coupled with recurrent states, allowing for finer temporal grids and multiple valid predictions from different noise samples. Evaluated on six reasoning benchmarks, including ARC-AGI-1 and ARC-AGI-2, looped flows achieve state-of-the-art performance with 58.8% and 12.2% test accuracy, respectively.
looped flowsdenoising objectivesprobability flowrecurrent statestemporal grid
Beyond Word Error Rate: A Switch Aware Evaluation of ASR and Audio Language Models on English Yoruba Code-Switched Speech
The study introduces a switch-aware evaluation framework for assessing Automatic Speech Recognition (ASR) systems and audio language models on English-Yoruba code-switched speech, highlighting limitations of Word Error Rate (WER) alone. Using a deterministic 2000-utterance dataset, the analysis employs metrics like Switch Entry Token Error Rate (SETER), windowed switch point error rates, and diacritic-insensitive WER. Results reveal that while WER masks code-switching behavior, audio LMs outperform ASR models on switch-localized metrics, with Yoruba token recognition significantly poorer than English. Generative audio LMs exhibit issues like translation errors and prompt leakage. The authors release evaluation tools for reproducible benchmarking of African code-switched speech.
automatic speech recognitioncode-switched speechword error rateswitch entry token error rateaudio language models
Recognizing Is Not Reversing: A Controlled Inversion Test of Fact-Preserving News Framing
The study introduces a controlled inversion test to evaluate whether large language models (LLMs) can reverse known framing transformations while preserving factual content. Using three established framing techniques—evaluative lexis, agency realization, and information salience—the authors generate 540 paired variants from 60 news articles across three intervention strengths. Testing Qwen, DeepSeek, and Kimi, they find factual preservation remains high (0.84), but framing reversal is low (0.044–0.068), even when framing type and direction are correctly recognized (pooled reversal: 0.071). Results demonstrate a clear separation between factual fidelity, framing recognition, and framing inversion.
controlled inversion testfactual preservationframing transformationevaluative lexisagency realization
A Unified Per-Token Gating Family for On-Policy Distillation: FKL/RKL Mixing with Multi-Channel and Bias Coefficients
The paper introduces a unified four-coefficient parameterization for per-token gating in on-policy knowledge distillation (OPD), enabling multi-channel composition and explicit bias. This framework subsumes existing methods like EOPD and ToDi as one-dimensional restrictions. Evaluated on TweetEval emotion and hate tasks using Qwen3-32B teacher and Qwen3-4B student models, the full parameterization outperformed single-channel baselines in 33 of 36 comparable cells and dynamic gating surpassed static baselines in 19 of 26 cells. Results were directionally consistent across three-seed replications but lacked statistical significance at n=3. The parameterization serves as a shared coordinate system for comparing per-token gating designs in OPD.
on-policy distillationper-token gatingforward klreverse klmulti-channel composition
SIRF: A Spec-Internalized Risk Foundation Model for Industrial Content Risk Control
SIRF (Spec-Internalized Risk Foundation Model) introduces a novel approach for industrial content risk control by internalizing platform-specific policies into model weights via continued pretraining (CPT). The method leverages EntiGraph, MAGA rewriting, and account-level chain-of-thought (CoT) synthesis without additional human annotation, enabling high-precision rule application under ultra-low-latency, verdict-only deployment. In a controlled comparison, SIRF-8B-SFT achieves 71.3% Black Recall@P95, a 15.1 percentage point improvement over Qwen3-8B-SFT, using only ~70M CPT tokens. SIRF enhances mis-penalization recovery by 20% in tree-model adjudication and reduces relative mis-penalization by ~70% in freezing scenarios.
continued pretrainingchain-of-thoughtblack recalltree-model adjudicationmis-penalization
LOCUS: Task-Aware Low-Rank Post-Training for Token-Efficient Language Generation
LOCUS introduces task-aware low-rank post-training for token-efficient language generation, optimizing output-token cost while preserving utility. The method selects a low-rank adaptation subspace, retaining the native preference objective with a frozen backbone, and updates only 0.24–0.28% of model parameters. Evaluated on Anthropic HH-RLHF dialogue preferences with Pythia-2.8B and Qwen2.5-3B backbones, LOCUS reduces continuation length by up to 39.84% and 14.87–17.58%, respectively, compared to full-parameter DPO, DrDPO, and SamPO branches, with no significant change in preference diagnostics.
low-rank adaptationtoken efficiencypost-trainingpreference alignmentdecoder backbone
ORCH: Organizational Principles Enable Collective Intelligence in Embodied AI
ORCH introduces organizational principles from human theory to enhance collective intelligence in embodied multi-agent systems, operationalizing pooled and sequential interdependence for task-specific hierarchical structures. The method was evaluated across 25 wildfire-response missions involving up to 50 heterogeneous agents and eight large language models, comparing against four prior frameworks. ORCH organizations improved mission outcomes by 63.97% and execution efficiency by 74.29% when human-designed, and by 43.63% and 52.53% when automatically generated. Performance advantages persisted across missions and models, with hierarchical organization enabling concurrent activity and phased transitions.
collective intelligenceembodied aiorganizational principlessequential interdependencehierarchical organization
Continuous-Time Acoustic Modelling with Neural Controlled Differential Equations
The paper introduces a continuous-time mechanism for duration-aware acoustic modeling in text-to-speech (TTS) using neural controlled differential equations (CDEs). It formulates phone representations as temporally parameterized control paths and employs a neural acoustic vector field to generate evolving hidden states based on phonetic content and duration-derived timing. The resulting trajectory is sampled discretely and integrated into standard acoustic decoder pipelines. Objective comparisons show CDEs outperform typical recurrent models, while subjective evaluations indicate improved rank-order agreement in emotion intensity without compromising emotion-expression quality. Experiments with varying temporal resolutions highlight trade-offs between style tracking and absolute calibration, positioning CDEs as a promising approach for style-sensitive TTS.
neural controlled differential equationstext-to-speechacoustic modelingduration-awarephonetic content
A Time-Based Readout for Vector-Matrix Multiplication in Fully Analog Memristive SNNs
The authors propose a fully analog readout architecture for vector-matrix multiplication (VMM) in spiking neural networks (SNNs) using voltage-to-time conversion, eliminating current-mode summing and scaling circuitry to improve area and energy efficiency. The method leverages memristive crossbar arrays for in-memory, analog VMMs, addressing the von Neumann bottleneck. Post-layout simulations of a 10x1 SNN in 130 nm CMOS technology validate the architecture, and its application to a trained 64x10 SNN for digit classification demonstrates feasibility for SNN inference. The approach reduces area and power overhead compared to conventional current-mode readout circuits.
vector-matrix multiplicationspiking neural networksmemristive crossbar arraysvoltage-to-time conversioncmos technology
When Agents Disagree: Bayesian Backward Reasoning as a Label-Free Anchor for Multi-Agent Collective Decision-Making
The paper introduces Bayesian backward reasoning as a label-free anchor for multi-agent collective decision-making when LLM agents yield conflicting answers. The method constructs reverse posteriors through explicit likelihoods, contrasting with forward reasoning approaches, and uses Jensen-Shannon divergence to rank agents by cross-path consistency. Three strategies—MinJS, FwdJS, and LogLin—are proposed, with LogLin achieving the best performance, particularly on instances of agent disagreement. Evaluated on DDXPlus across five LLM backbones, the strategies consistently improve over baselines, demonstrating the utility of reverse posteriors as complementary information. A lightweight two-stage calibration further enhances performance when labeled data are available.
bayesian backward reasoningreverse posteriorjensen-shannon divergencecross-path consistencylog-linear fusion
Language-Augmented Semantic Priors for B-Spline Surface Fitting
LASP introduces a Language-Augmented Semantic Priors framework that bridges the gap between high-level design intent and geometric configuration in CAD systems by leveraging LLMs. The method translates procedural modeling histories into textual descriptions, uses a fine-tuned LLM to predict structured B-spline prior parameters, and employs a two-stage training scheme combining local geometric regularities with long-range dependencies. This approach enhances semantic coherence and accuracy in B-spline surface fitting, outperforming traditional machine learning methods and establishing a new paradigm for language-guided geometric optimization.
b-splinelanguage-augmentedsemantic priorsgeometric optimizationcad systems
ActSafeGuard: Differentiable and Training-Aligned Constraint Enforcement for Flow-Matching Policies
(No summary returned.)
COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization
COBRA-Skills introduces an efficient framework for optimizing reusable skills in LLM agents by formulating skill optimization as budgeted sequential evolution. The method combines contextual-bandit-guided prioritization with evidence-grounded skill evolution, selectively allocating evaluations to promising candidates and refining the skill population using execution feedback. Evaluated across six heterogeneous benchmarks and three target models, COBRA-Skills achieves superior average performance while reducing optimization costs by 55-58% compared to SkillOpt, using only 50 unique optimization examples per benchmark. The framework demonstrates robustness to changes in the agent harness and effectiveness when the target model generates and refines skills.
contextual banditskill optimizationllm agentssequential evolutionexecution feedback
Ecdysis: Efficient and Effective Training of Runtime Harnesses for LLM Agents
Ecdysis introduces an efficient framework for optimizing runtime harnesses in LLM agents by addressing the bottleneck of principled failure diagnosis. The method employs batch-level cross-instance failure aggregation and Failure-Driven Collaborative Refinement to distinguish model-specific deficiencies from systematic harness issues, focusing adaptation on recurring cross-task failure patterns. This approach reduces unnecessary model-specific accommodations and improves generalization. Experimental results demonstrate that Ecdysis achieves a 1.84x speedup in harness training and enhances reasoning accuracy by 18.56% compared to existing methods.
runtime harnessesfailure diagnosisbatch-level aggregationcollaborative refinementgeneralization
Geospatial AI, Dataverse Metadata, and the Study of Place-Based Government
We construct a knowledge graph from Harvard Dataverse's metadata, organizing 102,650 datasets into a 215,985-node network with 528,003 edges linking datasets to keywords, publications, subjects, journals, and locations. Of these, 43,991 (42.9%) contain geospatial metadata, with 96.9% of nodes residing in a single connected component, ensuring dataset reachability. A keyword search identifies 7,654 (17.4%) geospatially tagged datasets as policy-relevant, clustered around elections, government administration, health policy, transportation, and education. The graph highlights challenges in place resolution and metadata enrichment, with a coverage skew toward American city-level data, providing a concrete setting for AI-driven entity resolution and metadata enrichment.
knowledge graphgeospatial metadataentity resolutionplace resolutionmetadata enrichment
Warrant Theory
The paper introduces warrant theory as a philosophical framework for analyzing the inferential legitimacy of propositions within logical systems. It distinguishes warrant from truth and belief, focusing on inferential entitlement and its relation to meaning and use. The theory provides a normative framework for examining how propositions gain inferential standing, develop, and interact through relations of dependence, compatibility, and exclusion. Acceptance and rejection are used to represent positive and negative inferential positions, with logical consequence and failure understood through the presence or absence of inferential entitlement. The paper concludes by defining logic as the formal study of conditions for legitimate propositional acceptance or rejection.
warrant theoryinferential legitimacyinferential entitlementlogical consequencenormative framework
Autonomy, Social Norms, and Alignment: Towards a Developmental Framework for Autonomous Artificial Agents
The article proposes a developmental framework for aligning autonomous artificial agents with human goals in dynamic environments. It identifies limitations in current AI systems reliant on pre-trained datasets and human feedback, advocating for intrinsic motivations like curiosity to guide exploration. The framework draws parallels to child development, suggesting that norms emerge through situated experience, autonomous learning, and cooperation. Regulatory sandboxes are proposed as pedagogical environments for progressive alignment, enabling agents to manage increasing degrees of freedom responsibly through interaction in complex scenarios.
intrinsic motivationsregulatory sandboxesalignmentautonomous agentsdevelopmental framework
ZipCodec: Ultra-Low-Frame-Rate Streaming Speech Coding
ZipCodec introduces an ultra-low-frame-rate streaming neural speech codec operating at 6.25 Hz and 0.80 kbps with 160 ms theoretical latency. The method combines large-scale WavLM distillation, a transformer-based architecture, a scalar spherical quantizer, and a latency-aware streaming decoder. Experiments demonstrate that ZipCodec outperforms existing streaming codecs at comparable bitrates in both reconstruction and downstream tasks, despite its 842M parameters, enabling real-time single-stream inference on consumer-grade CPUs.
neural audio codecwavlm distillationscalar spherical quantizerlatency-aware decoderstreaming speech coding
LoaDiff: Conditional Generation of Electricity Consumption Time Series for Energy Analytics
LoaDiff introduces a diffusion-based generative model for synthesizing year-long, sub-hourly smart-meter electricity consumption time series, addressing data-access restrictions in energy analytics. The model supports conditioning on static household attributes (e.g., appliance ownership) and dynamic variables (e.g., calendar information, outdoor temperature). Evaluated on three residential datasets, LoaDiff demonstrates high fidelity and diversity, minimal training-record memorization risk, and utility in downstream tasks like load forecasting and appliance detection. It also exhibits coherent controllability under varying temperature conditions, outperforming multiple generative baselines.
diffusion-based generative modelsmart-meter dataload forecastingappliance detectionconditional controllability
MAPLE: Memory-Augmented Planning with Language and Evolution
MAPLE introduces a memory-augmented agent for maintaining optimization problems through successive natural-language requests, addressing dynamic real-world operations. The method combines language-based problem construction with mathematical programming and evolutionary search, retaining optimization programs, accepted plans, and candidate solutions for reuse. Evaluated on NLDO, a benchmark with 15 trajectories and 180 updates across domains like scheduling and routing, MAPLE achieves an online scalar quality of 0.951 and a Pareto hypervolume ratio of 0.875. Controlled comparisons demonstrate that maintaining executable state improves update validity and preserves useful search information across revisions.
optimizationnatural-language processingevolutionary searchmathematical programmingmemory-augmented
Physics-Informed Neural Networks to Infer the Perpendicular Energy Conductivity in the Scrape-Off Layer of Stellarator Devices
An inverse Physics-Informed Neural Network (PINN) framework is developed to infer the scrape-off layer (SOL) perpendicular heat conductivity κ⊥(n,T) in stellarator devices, combining radial profile measurements with a reduced one-dimensional SOL transport equation. Three neural networks are trained: two reconstruct temperature and density profiles, while a third represents effective conductivity. Validation on synthetic data shows recovery of the imposed functional dependence with errors below 10%. Bootstrap resampling provides reliability indicators, and a trade-off between reconstruction accuracy and data availability is identified. Application to TJ-II stellarator experimental data demonstrates the method's potential for extracting transport information from plasma edge measurements.
physics-informed neural networksscrape-off layerstellaratorheat conductivityplasma transport
Distributed Optimization of Modular Production Systems using Model-based Reinforcement Learning with Inverse Models
The paper introduces a model-based reinforcement learning framework incorporating approximate inverse process models for optimizing modular manufacturing systems. The method disentangles actuation dynamics from state-space dynamics, enabling RL training exclusively in task space through a lightweight feedforward architecture integrated into policy networks. Experimental validation on a heterogeneous modular production testbed demonstrates improved efficiency in both performance and training speed, particularly for off-policy algorithms.
model-based reinforcement learninginverse process modelsmodular manufacturing systemstask space optimizationoff-policy algorithms
Making Alternative Data Work: Context-Augmented LLMs for Financial Forecasting
The authors propose a context-augmented LLM framework for financial forecasting that integrates alternative data with traditional financial information. Their two-agent system first identifies firms for which each alternative data channel is informative, then predicts revenue using firm- and channel-specific context. Evaluated across four commercial alternative data channels, the framework demonstrates improved forecasting accuracy compared to using either alternative or traditional data alone, outperforming standard baselines. Results suggest LLMs offer a flexible approach to combining heterogeneous financial information sources through in-context learning.
alternative datain-context learningfinancial forecastinglarge language modelsheterogeneous information
Learn the Solid, Not the File: Canonical Inputs for Neural Networks on CAD Boundary Representations
The paper introduces the canonical region graph, an invariant input representation for CAD boundary representations (B-reps) that addresses robustness issues in existing B-rep encoders. The method derives nodes, features, and coordinate frames directly from the solid, ensuring theoretical invariance to repartitioning and rigid motions. Evaluations on standard benchmarks, CAD software variations, and a human-modeled FreeCAD dataset show that the approach matches top baselines while maintaining stability across all tested perturbations, unlike existing encoders which often fail catastrophically.
boundary representationcanonical region graphcad systemsinvariant encodingparametric modeling
Enabling Knowledge Graph Understanding at Scale with the EXplore Your Graphs ENgine (EXYGEN)
The paper introduces EXYGEN, a framework for conversational knowledge graph (KG) access via text-to-SPARQL generation without task-specific fine-tuning. The method integrates VoID descriptions and ShEx schemas into a retrieval-augmented generation (RAG) pipeline, achieving 0.419 exact match on SciQA with schema context, retrieved triples, and question-query examples. A parallel graph sampling strategy enables metadata generation for large KGs, reducing runtime by 80x while preserving predicate coverage. Results show general-purpose LLMs outperform code-specialized ones given sufficient context, though lexical metrics poorly predict query correctness.
knowledge graphsparqlretrieval-augmented generationvoid descriptionsshex schemas
A Comparative Evaluation of Pre-trained Convolutional Neural Networks for Melanoma Detection
This study evaluates pre-trained convolutional neural networks (CNNs) for melanoma detection across dermatoscopic and histopathological image modalities. Five architectures (ResNet50, VGG16, VGG19, MobileNet, InceptionV3) were assessed on the HAM10000, ISIC 2018, and CR-AI4SkIN datasets under identical training protocols. Results showed accuracy ranges of 71%-84% on dermatoscopic images (ResNet50 achieving 84% on HAM10000) and 72%-83% on histopathological images (ResNet50 achieving 83% on CR-AI4SkIN). Performance varied significantly between modalities, with architectures demonstrating differing effectiveness across dermatoscopic and histopathological data.
convolutional neural networksmelanoma detectiondermatoscopic imageshistopathological imagesimage classification
Characterizing Job Power Elasticity for Power-Flexible AI Training
This paper introduces the Power Flexibility Index (PFI), a normalized metric quantifying the performance cost of power reductions in large language model (LLM) training, enabling SLA-aware power flexibility. The authors systematically characterize job power elasticity through 131 LLM training runs on H200 GPUs, including dense and mixture-of-experts models, pretraining and fine-tuning tasks, and up to 32 GPUs. Results show substantial but variable power elasticity, with PFI-aware power allocation recovering ~1.5k tokens/s per job under a 30% power reduction, achieving 63% of the performance gap between equal-weight allocation and an oracle.
power flexibility indexjob power elasticitylarge language modelmixture-of-expertssla-aware
Prompt Revision as a Source of Cultural Bias in Text-to-Image Systems
The study introduces WORLDVIEW, a multilingual benchmark of 8,960 prompts across 15 languages and 31 language-context pairings, to audit cultural bias in text-to-image systems. It examines prompt revision layers in DALL-E-3, Imagen-4, and GPT-Image-1.5 through a three-step analysis: cultural context marking, vocabulary flattening, and stereotyping. Results show that non-Western and non-Anglophone contexts are marked more heavily, flattened into narrow vocabularies, and reduced to stereotypes compared to a no-context English baseline. The revision layer is identified as a causal source of stereotyping, emphasizing the need to audit deployed systems rather than models alone.
text-to-image systemsprompt revisioncultural biasmultilingual benchmarkstereotyping
Lightweight LiDAR-Based Cone Detection Framework Using Random Forest for Formula Student Driverless
The paper introduces a lightweight LiDAR-based perception pipeline for Formula Student Driverless vehicles, optimized for CPU execution without GPU acceleration. The method combines ground removal, IMU-based motion compensation, DBSCAN clustering, and geometric feature-based Random Forest classification, reducing input features from 12 to 7 via importance analysis. Evaluation on 2,371 labeled clusters from real FSD events demonstrates 98.33% F1-score and 3.13 ms end-to-end latency on CPU hardware. The released dataset and tools provide a reproducible baseline for resource-constrained autonomous racing systems.
lidarrandom forestdbscanmotion compensationfeature importance
Learning Interaction between Image and Layout Priors for Joint Image-Layout Generation in Design Templates
InterIL introduces a joint generation model for graphic design templates, simultaneously producing background images and foreground layouts via bidirectional interaction between pretrained diffusion models. The method connects frozen image and layout diffusion backbones with a trainable communication module, preserving single-modality priors while learning cross-modal dependencies. Evaluations demonstrate superior image-layout harmony over sequential baselines, with test-time guidance enabling user customization without retraining.
joint generationdiffusion modelscross-modal interactiongraphic designtest-time guidance
Extending SMT Solving with Non-Ground Clause Learning
The paper introduces a calculus combining ground instantiations, CDCL(T)-style rules, and non-ground conflict analysis to enhance SMT solving. By performing resolution steps on original non-ground clauses during conflict analysis, the method learns more general clauses than ground conflicts, potentially yielding exponentially shorter proofs. The approach supports chronological backtracking and subsumes CDCL, SCL(FOL), SCL(T), and Resolution. Theoretical results demonstrate non-redundancy of learned clauses under a suitable strategy, unifying CDCL(T)-style SMT solving, instantiation-based procedures, and non-ground reasoning.
smt solvingnon-ground reasoningconflict analysiscdcl(t)quantifier instantiation
Structural priors for data-efficient language learning
The study investigates structural transfer as a method for data-efficient language learning, where models are first trained on non-language data (music, probabilistic grammars, cellular automata) to induce useful priors for natural language. Evaluating via next-token-prediction loss, weight shifts, and downstream linguistic benchmarks, the authors find that symbolic data types yield lower language-modeling loss than random initialization, suggesting favorable parameter-space positioning. However, these gains do not consistently improve downstream linguistic performance, and non-language data remains less efficient than additional language data. Structural transfer thus partially substitutes for language data in next-token prediction but not broader generalization.
structural transfernext-token predictionparameter-space initializationsymbolic datalinguistic generalization
ActMap: Single-Pass Uncertainty Quantification from Generation-Time Activation Maps
ActMap introduces a white-box uncertainty quantification method for large language models by compressing generation-time hidden-state trajectories into a fixed $12 \times 32 \times 128$ tensor, captured with no overhead. This compact representation preserves structural information across transformer layers and tokens, enabling efficient correctness prediction via lightweight classifiers like Vision Transformers or MLPs. Evaluated on QA, math, and summarization tasks with 7-8B models, ActMap outperforms baselines (sampling, token-probability, attention) and matches ACT-ViT's AUROC while using 67× smaller activations, demonstrating practical utility for scalable oversight.
uncertainty quantificationhidden-state trajectoryvision transformeractivation mapsscalable oversight
From Document Silos to Process Intelligence: A Multi-Layer Knowledge Graph for CMC Process Development
The paper presents a modular agentic-AI platform that constructs a dual-layer knowledge graph from heterogeneous Chemistry, Manufacturing and Controls (CMC) documents to address fragmentation in process development. The system combines a lexical graph layer for document structure with an ontology-aligned intelligence layer, using LLM agents for context-aware retrieval. Evaluation on 505 questions from 38 Sanofi small-molecule reports shows 95% Tier-1 multiple-choice accuracy and 85% Tier-2 LLM-judge pass rate, revealing limitations in comparative queries. A router agent dynamically selects between graph layers based on question type.
knowledge graphretrieval-augmented generationontology-alignedagentic-ailexical graph
Published Unlearning Numbers Move Per Checkpoint, and Not Because the Removed Data Survives: An Audit of 263 Released Batch-Normalized Checkpoints
The study audits 263 released batch-normalized checkpoints to investigate unlearning dynamics, revealing that published unlearning numbers shift per checkpoint independently of retained removed data. By refitting batch-normalization statistics on kept data at bit-identical weights, 47 of 221 checkpoints exceeded their release's seed spread, indicating checkpoint-specific drift rather than methodological shifts. Experiments showed that exchanging kept records for removed ones minimally affected published metrics, while checkpoint drift from refits tracked changes. Twelve verdicts crossed thresholds, four cleared recalibration budgets, and two replicated consistently. The findings emphasize naming fitting conventions in batch-normalized vision model releases.
batch-normalizationunlearningcheckpointrefittingrecalibration
The Convention Gap: Towards Measuring Implicit Communication in Cooperative AI Evaluation
(No summary returned.)
Flexible and Interpretable Accent Distance Measurements
The paper introduces a method for interpretable accent distance measurement using articulatory representations derived from articulatory inversion, addressing limitations in both phonetic formant analysis and accent embedding approaches. By employing optimal transport as a comparison framework, the method enables cross-recording-type accent analysis while maintaining interpretability. Results demonstrate that this approach provides a unified solution balancing the interpretability of phonetic comparisons with the flexibility of embedding-based methods.
articulatory inversionoptimal transportaccent embeddingsformant analysistext-to-speech
RouteRepair: Instance-Level Failure Diagnosis and Targeted Repair in LLM-Based Automated Heuristic Design for Routing Optimization
RouteRepair introduces instance-level failure diagnosis and targeted repair for LLM-generated routing heuristics, addressing aggregate evaluation's masking of recurrent structural failures. The method combines routing evidence, solver behavior, and program context to define bounded repair objectives, validated through parent-child evaluation of failure recovery and collateral degradation. Applied to TSP and CVRP across constructive search, guided local search (GLS), and ant colony optimization (ACO), RouteRepair-GLS reduces TSP optimality gap from 1.7476% to 0.7587%, while the CVRP heuristic lowers route costs by 1.91%; ACO priors also outperform hand-designed counterparts.
routing optimizationinstance-level diagnosisheuristic repairlarge language modelsant colony optimization
Cross-Lingual Clinical Annotation Projection as Constrained Text Generation: A Six-Language Study
The study introduces a constrained LLM-based method for cross-lingual clinical annotation projection, formulating it as a text-preserving generative task with deterministic validation. Using GLM 5.2 and Gemma4:31B, the approach inserts entity tags directly into immutable target-language text, achieving mean Strict F1 scores of 0.9201 and 0.9133 across 18 language-entity combinations, outperforming prior methods by 0.0564-0.1512. The method produced 55,416 grounded mentions with reconstructed offsets, demonstrating robust performance for multilingual clinical corpus construction.
cross-lingual annotationconstrained generationclinical nlpspan projectionmultilingual corpus
Prevalence Determines Precision:Silent Contamination in Detector-Defined Datasets
This study demonstrates that dataset precision in detector-defined ML datasets is governed by true-positive prevalence via Bayes' theorem, not solely by detector quality. Using a detector-defined event dataset and an independent official index, the authors analyze three pools with phantom rates of 81.7%, 9.0%, and 0.0%. Transferring precision from high-rate to low-rate pools yields a +422% error, while Bayes' expression predicts all rates within 3.3%. The detected response curve is an exact convex combination of true-event and phantom components, with phantoms outnumbering true events 473 to 308. Contamination manifests as a second signal with detector-inherited shape, affecting estimators differently based on normalization.
bayes' theoremdetector-defined datasetsphantom ratetrue-positive prevalenceestimator contamination
Investigating catastrophic forgetting in sound event classification
The study evaluates methods to mitigate catastrophic forgetting in class-incremental learning for sound event classification, using FSD50K and AudioSet datasets. Architectural and regularization approaches are analyzed, including selective kernel protection and a dynamic head that expands with new tasks. Results indicate catastrophic forgetting primarily affects deeper layers, especially the classifier head. Freezing the feature extractor while fine-tuning the dynamic head proves most effective, balancing memory stability and learning plasticity with minimal forgetting and high training stability.
catastrophic forgettingclass incremental learningsound event classificationdynamic headfeature extractor
Calibration-Aware Uncertainty Cascades for Efficient Heterogeneous Model Collaboration
Calibration-Aware Uncertainty Cascades (CAUC) introduces a post-hoc framework for efficient heterogeneous model collaboration by decoupling deployment policies from specific model pools and budgets. CAUC independently calibrates each model's confidence scores, establishing a unified reliability scale for early prediction acceptance, stronger model invocation, or selective output combination. Theoretical analysis shows calibration provides explicit selective-risk interpretation, unlike uncalibrated scores. Experiments across six language benchmarks demonstrate CAUC achieves a 1.9% average relative accuracy improvement over strong-model-only inference while avoiding 47% of strong-model calls. On image classification benchmarks, it maintains or improves predictive performance while reducing GFLOPs by up to 57%.
calibration-awareheterogeneous modelsconfidence calibrationselective-riskgflops
LLMs as Post-hoc Auditors of Physiological Plausibility in Symbolic Regression: A Clinician-Evaluated Case Study
This study evaluates Large Language Models (LLMs) as post-hoc auditors for improving interpretability and physiological plausibility of Symbolic Regression models derived from grammatical evolution. The authors analyze four symbolic expressions generated by grammar-based Genetic Programming, using three LLMs to rank them by interpretability and medical plausibility across three runs, with clinician assessments. Results show LLM comparative rankings outperformed isolated term-level interpretations in clinician evaluations, though some explanations contained physiologically or mathematically implausible content, suggesting LLMs are more effective for comparative auditing under expert supervision than autonomous validation.
symbolic regressiongrammatical evolutionlarge language modelsinterpretabilityphysiological plausibility
SWRouter: Similarity-Contractive Window Routing for Multi-Turn Large Language Model Conversations
SWRouter introduces a similarity-contractive window routing method for multi-turn large language model (LLM) dialogues, addressing context segmentation and evaluation challenges. The approach combines similarity-based context segmentation with a dual-metric framework to decouple prompt construction quality from model selection. Evaluations on multi-turn benchmarks show SWRouter outperforms baselines, achieving 16.26% higher accuracy than the best individual LLM and an 8.22% improvement over Conv-ID Context, demonstrating the necessity of joint context-evaluation design.
multi-turn dialoguecontext segmentationmodel routingsimilarity-contractivedual-metric evaluation
X-AuT: Progressive Audio-Encoder Compression for Speech LLMs with Cross-Scale Distillation
X-AuT introduces a progressive compression framework for speech large language models (LLMs) that reduces audio-encoder depth while minimizing performance degradation. The method employs cross-scale distillation, representation alignment, scheduled student-policy supervision, and LoRA finetuning, keeping the language-model backbone frozen. Evaluated on ten Chinese--English benchmarks, compressing Qwen3-ASR-0.6B from 18 to 16 audio-encoder layers reduces macro-average error from 5.61% to 5.27%, while a 14-layer model achieves 5.75% error with 20.7% fewer parameters. Progressive pruning (18→14) outperforms direct pruning (5.75% vs. 6.73%), and teacher-student distillation yields lower error (5.55%) than self-distillation (8.45%).
speech llmscross-scale distillationlora finetuningaudio-encoder compressionprogressive pruning
Deep-Fake CAPTCHA: Mitigating Next-Generation Social Engineering Attacks
DF-CAPTCHA introduces an active defense mechanism against real-time deepfake impersonation in voice and video calls by employing challenge-response tasks that are trivial for humans but challenging for current deepfake systems. The framework evaluates responses based on realism, identity consistency, task completion, and response time, outperforming passive detection methods. Experiments with user studies and real-time deepfake models demonstrate high accuracy in distinguishing authentic media from deepfakes, validating its efficacy as a countermeasure against social engineering attacks.
deepfakecaptchasocial engineeringreal-time detectionchallenge-response
From Queries to Narratives: Cultural Heritage Data Stories for Knowledge Graph Exploration and Quality Assessment
The paper introduces data stories as a method to enhance accessibility and quality assessment in cultural-heritage knowledge graphs (KGs) like NFDI4Culture-KG. Data stories combine narrative text, images, executable SPARQL queries, and visualized results to guide users through unfamiliar graphs, create reproducible narratives, and uncover hidden data-quality issues. The authors present LODEON, an authoring platform with Sparnatural and AI-assisted tools, enabling transparent and reproducible claims backed by explicit queries. Early workshops demonstrate that data stories lower barriers to KG exploration while facilitating quality assessment. The approach addresses challenges posed by heterogeneous, specialized, and evolving KG data.
knowledge graphssparqldata qualitycultural heritagereproducible narratives
Beyond Confidence: Stability-Aware Test-Time Adaptation for LLM Reasoning
Proposes Test-Time Adaptation via Stability-Aware Confidence Optimization (TASCO), a lightweight framework that enhances LLM reasoning by optimizing task-level prefixes for stability under local perturbations. TASCO employs two perturbation strategies—Random Perturbation for distributional stability and Sharpness-Aware Perturbation for worst-case sensitivity—while keeping the LLM frozen. Experiments show improved reasoning accuracy and token efficiency across multiple LLMs and benchmarks, with behavioral analyses confirming stable confidence under perturbations.
test-time adaptationreasoning stabilitypredictive entropylocal perturbationstask-level prefix
Buyer Artificial Intelligence-Enabled Environmental Governance and Supplier Environmental Controversies: An Organizational Information Processing and Signaling
The study demonstrates that suppliers' exposure to buyers' AI-enabled environmental governance reduces subsequent environmental controversies, with stronger effects in countries with higher AI readiness and regulatory quality. Using organizational information processing and signaling theory, the authors analyze panel data from 2,505 suppliers of U.S.-listed firms across 41 countries (2020–2024) via text analysis and fixed-effects models. Results show a negative association between AI-enabled governance and supplier controversies, moderated by institutional factors.
ai-enabled governanceenvironmental controversiessupply chain risktext analysisfixed-effects models
VikingRAG: Accurate and Token-efficient Retrieval-augmented Generation over Structured Documents
(No summary returned.)
Agent-Integrated Software: Interaction Contracts and Continuous Assurance
The paper introduces Agent-Integrated Software (AIS), a software pattern combining conventional applications with built-in intelligent agents, and Intent-Level Interaction Abstraction (IIA) for task semantics. It proposes an open transition-system model linking AIS execution to IIA states, with interaction contracts ensuring task bindings, role authority, and outcome evidence. Continuous assurance maintains claims as dependencies evolve. The framework highlights limitations through domain contrasts and outlines a research agenda for maintainable agent integration in software engineering.
agent-integrated softwareintent-level interaction abstractioninteraction contractscontinuous assurancetransition-system model
Characterizing Bluesky Content Moderation Service: From Automation of Service to Landscape of Harms
This work presents the first large-scale audit of Bluesky's Moderation Service (BMS), analyzing 10.6M moderation labels to characterize its automation, efficacy, and harm detection. The study reveals a hybrid human-AI system: sexual/graphic content is labeled automatically (seconds), while nuanced cases require human review (hours/days). Manual annotation shows BMS achieves high precision (0.837) but low recall (0.222), with annotators identifying 4.5× more harmful content. Unsupervised clustering reveals detected harms include hostility toward protected groups and explicit content. The findings provide empirical insights for designing transparent moderation systems.
content moderationautomated labelingprecision-recall tradeoffharm detectiondecentralized platforms
RAMamba-Net: A Reliability-Aware and Mamba-Based Multimodal Fusion Network for Auditory Attention Detection
RAMamba-Net introduces a reliability-aware, Mamba-based multimodal fusion network for auditory attention detection (AAD), addressing limitations in cross-modal interaction, temporal modeling, and robustness. The method combines a Mamba-enhanced band-aware convolutional Transformer for EEG, a dual-branch encoder for EOG, and cross-modal attention, with a reliability-aware module for sample-wise modality weighting. Evaluated on two AAD benchmarks, it achieves a 5.76% accuracy gain over unimodal baselines, demonstrating improved robustness and discriminative representations through enhanced multimodal alignment and suppression of unreliable evidence.
auditory attention decodingmultimodal fusionmamba architectureeeg-eog integrationreliability-aware learning
Portable Semantics, Private Dialects: Reuse and Negative Transfer in Latent Communication Between Language-Model Cells
The study investigates latent communication interfaces in independently trained language-model societies, examining zero-shot transfer and negative transfer effects. Using a leakage-controlled causal interoperability audit across 30 ordered pairs of six restricted societies, it finds no shared raw packet language, with only one same-initialization pair achieving bidirectional interoperability. A source-span control attributes zero-shot failure to new operator instruction interpretation, while adaptation experiments show globally trained interfaces induce severe negative transfer (final accuracy improving from 0.169 to 0.857 when reinitialized). Results are bounded to a 17-state setting, with inherited interfaces never outperforming fresh controls by the preregistered 0.10 margin.
latent communicationzero-shot transfernegative transferinteroperability auditlanguage-model societies
Reification as a Transferable Vocabulary: Zero-Shot Link Prediction with Vanilla GNNs
The paper demonstrates that zero-shot link prediction on unseen knowledge graphs can be achieved by reifying input graphs into a fixed meta-relational vocabulary, eliminating the need for specialized architectures. Facts become nodes connected via six meta-relations, with relation types as shared anonymous nodes. Five standard GNNs (GAT, GINE, GraphSAGE, R-GCN), trained on a single 4,245-triple graph for 30 minutes on an A100, generalize to 40 inductive benchmarks, with GAT matching ULTRA's performance. The approach extends to relational databases, where foreign-key columns are treated as relation types, showing strong performance on schema-less databases. Code and checkpoints are released.
zero-shot learningknowledge graphsgraph neural networkslink predictionrelational databases
Exploring Diffusion Transformers for Cross-Modal Augmentation in Multimodal Brain State Decoding
We propose CoMA-DiT, a bidirectional cross-modal Diffusion Transformer for latent augmentation that treats paired modalities as mutual generative supervision rather than mere fusion inputs. CoMA-DiT employs cross-modal attention for velocity prediction conditioned on the paired modality and adaptively injects variation via a reliability-gated residual mechanism. Experiments on multimodal auditory attention decoding and emotion recognition demonstrate CoMA-DiT's superiority over 20 baselines, achieving absolute gains of 4.28% accuracy and 6.70% macro-F1 over the no-augmentation baseline. Ablation, sensitivity, visualization, and interpretability analyses confirm its robustness, generalizability, and ability to capture functionally relevant cross-modal interactions.
diffusion transformercross-modal attentionlatent augmentationreliability-gated residualmultimodal learning
On the Impact of Anonymization on the Performance of Large Language Models
This paper systematically evaluates the privacy-utility trade-off in large language models when processing anonymized inputs, testing five models across eleven benchmarks. The study reveals nuanced effects: high-capacity models like Qwen2.5-72B and GPT-4o mini show greater performance degradation than smaller models, suggesting stronger reliance on entity-specific information. Task-dependent impacts emerge, with TruthfulQA improving under anonymization while retrieval tasks like RGB decline sharply. Reversible anonymization outperforms irreversible methods, and explicit anonymization prompts yield no benefit. The findings highlight the need for task- and model-aware anonymization strategies.
anonymizationprivacy-utility trade-offentity-specific informationreversible anonymizationretrieval tasks
E-CONAN (Entailment, CONtradition And Neutral) Benchmarks: Arabic Textual Entailment and Natural Inference Datasets
The paper introduces E-CONAN, a set of Arabic textual entailment and natural language inference benchmarks comprising two datasets: E-CONAN-2 (2-way RTE) and E-CONAN-3 (3-way NLI). The datasets are constructed from four sources: automatically-translated pairs, human-validated machine-translated pairs, hand-crafted pairs from Arabic teaching materials, and headline pairs from news channels. Nine multilingual pretrained models and five large language models (LLMs) were evaluated using zero-shot classification across E-CONAN, ArNLI, and XNLI datasets. Results indicate E-CONAN's potential for evaluating model generalization and fine-tuning, outperforming XNLI and ArNLI in robustness. MARBERT was included as an Arabic-specific baseline, and error analysis revealed frequent error patterns.
textual entailmentnatural language inferencezero-shot classificationmultilingual pretrained modelsarabic-specific baseline
The Semantic Elevation Operator and the Closure of the Undecidable Class under Preservation
The paper introduces the semantic elevation operator ΛΦ, which transforms static semantic properties into dynamic preservation questions for self-modifying systems. Using Kleene's recursion theorem, the authors prove that when Φ is intensional, the elevated property remains undecidable, breaking the extensionality required by Rice's theorem. They demonstrate that the class U of non-verifiable properties is closed under ΛΦ and that unbounded iteration of ΛΦ climbs the arithmetical hierarchy to Π02-completeness. Additionally, they show that no finite tower of verifiers yields an unconditional certificate, indicating a supervisory regress. Future work may explore a categorical interpretation in the effective topos.
semantic elevation operatorkleene's recursion theoremarithmetical hierarchyrice's theoremeffective topos
AI Exposure and AI Resilience: A Two-Dimensional Assessment Framework for Software and Software-Based Business Model
The paper introduces Artificial Intelligence Exposure and Resilience (AI-ER), a two-dimensional framework for assessing AI's impact on software-based business models. AI exposure quantifies the pressure AI imposes on a business model, while AI resilience measures the ability to adapt and leverage AI economically. The method derives metrics from current AI capabilities, deployment conditions, and organizational adaptability research, maintaining separate exposure-resilience assessments with explicit evidence quality scoring. The framework supports initial public-data analysis and later internal refinement, producing traceable, comparable company profiles. An initial scoring logic and empirical validation procedure are also proposed.
ai exposureai resiliencebusiness model assessmentorganizational adaptabilityevidence quality
Magenta: Closing the Loop Between Mathematical Reasoning and Lean Verification
Magenta bridges informal mathematical reasoning with formal verification by integrating Lean 4 into LLM-based problem-solving. The training-free pipeline processes natural-language problems, generates answers, formalizes them as Lean statements, and constructs machine-checked proofs. It employs a statement judge to verify problem preservation and an error-attribution judge to route failures to re-derivation or Lean repair. Magenta achieves 100% accuracy on olympiad benchmarks (AIME 2025, AIME 2026, HMMT February 2026) and solves all six IMO 2026 problems when paired with K2-Horizon-7B. Analysis shows statement adjudication prevents false certificates, and feedback-guided correction outperforms resampling.
lean verificationmathematical reasoningllmformal prooferror attribution
Mr.LHDR: A Benchmark for Multimodal Real-World Long-Horizon Deep Research Agents
The authors introduce Mr.LHDR, a benchmark for evaluating multimodal long-horizon deep research agents, addressing the lack of benchmarks for sustained, dependency-heavy research processes. Mr.LHDR constructs questions from hidden Node-Relation graphs, requiring an average of 12.1 intermediate conclusions with a mean dependency depth of 10.4, incorporating multimodal evidence (images, PDFs, video frames, etc.). Evaluations on general models and agent frameworks using Overall Accuracy (OA), Strict Accuracy (SA), and Dependency-Aware Checklist Score (DACS) reveal that the strongest system achieves only 43.1% OA, with multimodal evidence removal reducing DACS by 12.6 points. Results highlight dependency-consistent evidence integration as a key bottleneck.
multimodal benchmarklong-horizon researchdependency-aware evaluationnode-relation graphintermediate conclusions
Routing by Reasoning Need: Trajectory-Aware Decoding Control for Diffusion Vision-Language Models
The study introduces a trajectory-aware decoding controller for diffusion vision-language models (VLMs) to address reasoning-need mismatch, where fixed generation lengths harm performance on questions with varying reasoning demands. The proposed training-free method routes examples to early commitment, baseline preservation, or reasoning-supportive decoding using trajectory signals (answer closure, commitment evidence, representation revision pressure) without ground-truth answers. Evaluated on LLaDA-V across answer-focused, mixed-reasoning, and chain-of-thought (CoT)-sensitive benchmarks, routed control improves robustness over fixed or single-rule decoding strategies. Benefits are not due to shorter outputs alone: answer-closed examples favor commitment, while CoT-sensitive ones require intermediate reasoning support.
diffusion vision-language modelsreasoning-need mismatchtrajectory-aware decodingchain-of-thoughtinference-time control
GRIPNet: Gaussian Radial Intensity Prior Guided Architecture for Pulmonary Nodule Detection in CT
GRIPNet introduces a Gaussian radial intensity prior for pulmonary nodule detection in CT scans, addressing the challenge of detecting sub-6mm lesions by modeling nodule appearance as a radially decaying Gaussian intensity pattern. The method employs pinwheel convolutions for radial gradient decomposition, a dual-frequency module for boundary-context separation, dilated masked attention for decay extent matching, and an adaptive loss for sample reweighting. Evaluated on KanserSet, LUNA16, and Lung-PET-CT-Dx, GRIPNet achieves mAP@0.5 scores of 95.3%, 91.6%, and 97.9%, respectively, while improving high-IoU localization at real-time speeds.
pulmonary nodule detectiongaussian radial intensitypinwheel convolutionsdual-frequency moduledilated masked attention
Your Model Already Knows Don't Teach It, Learn to Ask It: Soft Prompting for Few-Shot Adaptation of Vision-Language Models
The paper proposes soft prompting for few-shot adaptation of vision-language models (VLMs) to out-of-domain tasks like aerial or medical imagery, using only 10 annotated images. Key innovations include placing continuous prompt tokens at the cross-modal boundary (10.0 vs. 8.4 mAP) and initializing them from the empty space token. With 1-3 learned tokens (7,168 parameters), soft prompting matches LoRA performance on Roboflow20-VL (14.2 mAP) while training 20,000x fewer parameters and avoiding catastrophic forgetting (0% VQA accuracy drop vs. 35-56% for LoRA). Learned tokens transfer to newer models (+0.8 mAP on Qwen3.5-9B) and verbalize into readable prompts. The method also improves RoboCasa manipulation tasks when applied to vision-language-action policies.
soft promptingfew-shot adaptationvision-language modelscross-modal tokenscatastrophic forgetting
2AM: Grounding Agent-Side Memory as Guidance for Steerable Action Models in Long-Horizon Manipulation
The paper introduces 2AM, a system for long-horizon robot manipulation that decouples task memory from action execution. The Agent maintains episodic memory and generates subtask language with optional 2D spatial hints (grasp/place/move), while a stateless RGB-based Action Model performs motions. Training involves condition dropout, spatial noise, and temporal jitter to improve robustness. On LIBERO-Mem, 2AM achieves 76.3% average task completion (61.5-point gain over baselines) without depth or geometric planners, demonstrating that Agent-side memory and precise steering suffice for complex manipulation.
long-horizon manipulationagent-side memoryspatial hintscondition dropoutsteerable action models
Memory Compression for High-Fanout Agent Sandboxes
AgentZip introduces memory compression optimized for high-fanout AI-agent sandboxes by exploiting template-relative and cross-sandbox redundancy. The system expands compressible pages to any with profitable representations, shifts overhead control to restore-time prefetching, and aligns compression with LLM waiting periods. Evaluated on LLM training and inference workloads, AgentZip achieves up to 8.7x memory reduction (vs. 2.1x for Linux) while limiting slowdown to 1.40x (from 3.1x) via prefetching and execution-aware scheduling.
memory compressionagent sandboxesprefetchingllm workloadsredundancy exploitation
Off-Target Effects of Response-Style Alignment in a Korean 27B Language Model
The study investigates off-target effects of response-style alignment in Qwen3.8-27B for Korean, focusing on abstention rates in KoBBQ and unprompted disclosure behaviors. By post-training the model with varying style seeds (verbosity, list usage, etc.) while holding other factors constant, the authors find that answer propensity shifts significantly (+0.82 pp to -1.53 pp) depending on stylistic targets, with changes driven primarily by emission policy rather than latent preference. Decomposition reveals dominant movement in answer propensity over conditional composition, complicating interpretation of conditional content preference. Detector agreement for rule-based constructs varies widely (0.44 to 0.99) across checkpoints.
response-style alignmentemission policyanswer propensityconditional compositionrule detectors
Generating a Consistent Enterprise: Synthesis and Reference-Free Evaluation of Multi-System Business Data
The paper introduces a synthetic data generator that produces multi-system enterprise data (workforce, customers, transactions) without real-world training data, ensuring cross-system consistency via entity graph projection into 66 business formats. Realism is enforced through reference statistics and evaluated via a five-axis scorecard (28 statistical checks), adversarial detection, and soundness tests. Over 23 synthetic companies, mean realism improved from 60.3 to 99.1, with adversarial detection dropping from 55.2% to 0%. A secondary generator creates relational databases from business questions, enforcing answerable queries and near misses. The system is deployed as a hosted service (https://console.era.eon.io) with containerized simulators.
synthetic dataentity graphadversarial detectionrelational databasesreference-free evaluation
When Does Text Inform? Benchmarking Information-Theoretic Metrics for Multimodal Time-Series Forecasting
The study introduces a synthetic benchmark to evaluate six mutual information estimators (KSG, MINE, InfoNCE, CCA, PID, V-information) for assessing the predictive value of text annotations in multimodal time-series forecasting. The benchmark generates controlled time-series data with semantically correct, incorrect, and irrelevant annotations, enabling exact ground-truth comparisons. Results show all estimators correctly rank correct annotations as most informative and audit mixed text corpora without model training, though limitations emerge on weak signals in seven real-world datasets. Practical guidelines for metric implementation in annotation auditing and fusion selection are provided.
multimodal forecastingmutual information estimatorstime-series analysistext annotationsinformation-theoretic metrics
Bio-inspired Learning and Decision-Making with Probabilistic In-Memory Computing Hardware: Part 1
The article proposes a biologically inspired framework for Bayesian learning and decision-making using probabilistic in-memory computing hardware. It models neural and synaptic dynamics as stochastic sampling processes from an internal energy function, capturing uncertainty over latent states and model parameters through neural and synaptic variability. This approach enables predictive coding networks to account for epistemic uncertainty via Markov chain Monte Carlo sampling. The framework draws parallels between intrinsic noise in biological systems and electrical noise in analogue memory technologies, suggesting that analogue in-memory computing hardware offers a scalable and energy-efficient solution for probabilistic inference.
bayesian learningprobabilistic inferencemarkov chain monte carlopredictive codingin-memory computing
Predicting Train Delays in Finland Using Machine Learning and Weather Data
This paper introduces a domain-informed feature engineering approach for train delay prediction using the Finland Integrated Train-Weather (FI-TW) dataset, which combines railway operational records with meteorological observations from 200 wireless sensor stations. The authors evaluate XGBoost with three feature configurations at Oulu central station (101,146 observations), finding that hierarchical weather category scenarios (e.g., Blizzard, Heavy Snow) outperform raw meteorological inputs. The category-based approach achieves an R^2 of 0.78, RMSE of 8.5 minutes, and MAE of 3.7 minutes, improving R^2 by 11% and reducing errors by 10% compared to alternative configurations. Results highlight the efficacy of compact, domain-specific features for edge deployment in railway delay prediction.
feature engineeringxgboostwireless sensor networksedge computingmeteorological observations
Improving Faint Object Detection for Space Situational Awareness with Variational Autoencoders
The authors propose a deep-learning pipeline for detecting faint moving objects in optical space situational awareness (SSA) imagery by addressing challenges from structured backgrounds and dense stellar fields. The method combines a Tiny-U-Net for stellar segmentation with a partial-convolution variational autoencoder (astro-VAE) to reconstruct star-free backgrounds through context-aware inpainting. Evaluated on ground-based X-GEO observations, the pipeline demonstrates high-fidelity background reconstruction while preserving moving targets, significantly improving detectability for low-SNR objects in optical SSA applications.
space situational awarenessvariational autoencoderfaint object detectionbackground reconstructioninpainting
AI-Powered Flare Combustion Efficiency Estimation
The study introduces a vision-language encoder paired with a multi-layer perceptron for estimating flare combustion efficiency from low-cost thermal video, addressing limitations of traditional gas analyzers and hyperspectral cameras. The system processes thermal footage to predict efficiency, displaying real-time trends, frame-by-frame overlays, and exportable reports via a graphical interface. Deployed over six months, it maintained 99% uptime with under 15 minutes weekly maintenance, demonstrating robust operational performance in industrial settings.
combustion efficiencyvision-language encodermulti-layer perceptronthermal videographical interface
Generative Replay Mitigates Sample Starvation in Quantum Architecture Search
GenQAS introduces a tensor network-guided reinforcement learning framework with prioritized generative replay to mitigate sample starvation in quantum architecture search. The method combines a fixed matrix product state warm-start with a learned local transition model that generates synthetic circuit transitions, mixing them with real experience during Double Deep Q-Network updates. Evaluated on chemical Hamiltonian benchmarks from 6 to 12 qubits and a 15-qubit transverse field Ising model, GenQAS improves success probability by up to 7.0× and increases success probability from 12% to 21%, respectively. In a noisy 6-qubit BeH$_2$ transfer experiment, it reduces steps to chemical accuracy by 92.7%.
quantum architecture searchgenerative replaytensor networkreinforcement learningmatrix product state
Sci-MMR: Benchmarking Multi-Step Evidence-Grounded Scientific Reasoning in Multimodal Agents
Sci-MMR introduces a benchmark for multi-step evidence-grounded scientific reasoning, comprising 235 multi-hop tasks across four disciplines with structured argument graphs linking claims, citations, visual evidence, and supporting regions. Evaluating eight multimodal models reveals a 20%+ gap between answer accuracy and complete-evidence recovery, exposing limitations in answer-only evaluation. Controlled interventions identify two bottlenecks: evidence acquisition (57.2% of failures, +4.5 points with cropping tools) and evidence integration (31.8% of failures, 69.1% accuracy with gold evidence), demonstrating current benchmarks overestimate multimodal agents' reasoning capabilities.
multimodal reasoningevidence-groundedscientific benchmarksmulti-hop tasksstructured argument graphs
From Evaluation to Enhancement: Benchmarking and Improving Think-with-Video Reasoning for Video Generative Models
The paper introduces VWG-Bench, a comprehensive benchmark with 9 reasoning dimensions and 38 tasks to evaluate 'think-with-video' capabilities in generative models, alongside a three-level VLM-as-Judge protocol for assessing fluency, rule adherence, and goal realization. It reveals a performance gap in logic-heavy tasks despite strong rendering scores and proposes Vid-PRE, a reinforcement-learned prompt rewriter that offloads reasoning to a VLM, improving performance across generators without architectural changes. Experiments demonstrate Vid-PRE's effectiveness, with resources publicly released.
video generationreasoning benchmarkvlm-as-judgeprompt rewritingreinforcement learning
HALDETECT at ImageEval 2026 Shared Tasks: Answer-First Contrastive Grounding with QLoRA
HALDETECT introduces an answer-first contrastive grounding approach for hallucination detection in multimodal models, applied to Task 1b of ImageEval 2026. The system fine-tunes Qwen2.5-VL-7B-Instruct with 4-bit QLoRA (freezing the vision encoder) to identify visually grounded statements among culturally plausible alternatives, achieving Contrastive Instability (CI) 0.035 on a 1,000-item test set (3rd place). Key findings include answer-order importance over model scale, QLoRA's superiority over prompting, and non-robustness of data-scaling curves to seed changes. Error analysis reveals residual challenges in cultural plausibility and naive adapter voting.
multimodal modelscontrastive groundingqlorahallucination detectionvision-language
NovGauge: A Fine-Grained Benchmark for Diagnosing LLMs' Capability in Paper Novelty Assessment
NovGauge introduces a fine-grained benchmark for diagnosing LLMs' capability in paper novelty assessment, addressing the limitations of holistic scoring. The benchmark comprises 619 paper pairs and 50 multi-paper sets, annotated along task, problem, and method dimensions, sourced from ICLR reviewer overlap claims and survey co-citations. A cascading diagnostic pipeline evaluates per-dimension correctness, evidence grounding, and logical support. Testing 18 LLMs reveals hallucination rates up to 39%, with GPT-5.5 achieving 43-72% Verified F1, while most models retain under half their raw F1 after faithfulness verification, highlighting significant gaps in reliable novelty assessment.
novelty assessmentllm evaluationfaithfulness verificationfine-grained benchmarkhallucination rates
A Voice-Interactive Multi-Agent System for Smart Operating Rooms: Architecture Design and Key Technologies
SurgicalRoomAgent introduces a voice-interactive multi-agent system for smart operating rooms using LLMs, achieving natural language understanding, device control, and report generation via a layered architecture. Key innovations include KV Cache prefix warming (reducing inference latency from ~500 ms to tens of ms), streaming partial JSON parsing (30% latency reduction), and progressive skill prompt disclosure for context efficiency. Implemented with Qwen3-27B and llama.cpp/sglang, the system operates within 16K-token contexts and meets OR real-time requirements for multi-device control.
multi-agent systemkv cacheprogressive promptingstreaming jsonsurgical robotics
Solving Few-Shot Multiobjective Multitask Optimization via Iterative Sequential Transfer
The paper introduces Iterative Sequential Transfer (IST), a method for few-shot multiobjective multitask optimization (MTO) that addresses the challenge of aligning elite solution distributions under limited evaluation budgets. IST models MTO as sequential transfer optimization problems, focusing evaluations on a single target task per iteration and employing likelihood-informed task prioritization to maximize transfer utility. Empirical validation on benchmark and real-world problems demonstrates IST's effectiveness in resource-constrained scenarios.
multitask optimizationknowledge transferfew-shot optimizationpareto manifoldsequential transfer
AI Soccer Analyst: Stage-Aware and Verifiable Human-AI Collaboration for Soccer Data Analysis
AI Soccer Analyst introduces a stage-aware human-AI collaboration system for soccer data analysis, addressing limitations of opaque prompt-to-report workflows. The system comprises six revisable stages: Data Understanding, Problem Definition, Structured Planning, Execution, Evidence-Grounded Reporting, and Interaction and Refinement. A formative study with five analysts informed design goals, followed by a task-based evaluation (N=16) where 33 of 48 tasks met completion criteria. Post-Holm correction, participants rated the system favorably on output quality, task achievement, reliability, and verifiability. Interaction logs revealed domain knowledge emergence through iterative refinement, demonstrating practical utility for inspectable, revisable analyses with expert oversight.
mixed-initiative systemstage-aware collaborationverifiable analysisdomain expertisetask-based evaluation
X-RACE: XAI-assisted Recurrent neural network Attribution for Channel Estimation
X-RACE introduces an XAI-assisted framework for efficient LSTM-based channel estimation in vehicular environments, combining input subcarrier pruning and hidden unit optimization via a one-shot dual-optimization strategy. The method proposes temporal XAI metrics (Saturation Time, Importance Drift, Relevance Contrast) to analyze LSTM dynamics, achieving a 44.1% reduction in inference complexity while maintaining or improving BER performance compared to classical XAI approaches.
lstmchannel estimationexplainable aipruningvehicular networks
CryptoL: Towards Scale Dominance and Physics Constraints Mitigation in Financial Multivariate Time Series Forecasting
CryptoL introduces a unified framework for multivariate time-series forecasting in cryptocurrency markets, addressing scale heterogeneity and structural dependencies among OHLC variables. The method employs context-normalized coordinates within the RevIN pipeline to mitigate disproportionate influence from large-scale assets, analyzes channel-dependent normalization to preserve candle-order relations, and incorporates scale-adaptive numerical stabilization with a soft feasibility loss for OHLC constraints. Experiments on heterogeneous cryptocurrency assets demonstrate improved forecasting accuracy (quantified vs. baselines), training stability, and financial validity of predictions. The framework integrates scale-balanced optimization, structure-preserving normalization, and constraint-aware forecasting.
multivariate time-seriesohlc normalizationscale-adaptive stabilizationempirical riskchannel-dependent transformation
An AI-Powered Culturally Aware Chatbot for Stress Detection and Wellness Support among Pakistani University Students Using NLP and Machine Learning
The paper introduces a culturally sensitive chatbot for stress detection and wellness support among Pakistani university students, combining a Random Forest classifier (89.09% accuracy, 0.89 macro F1-score) trained on 1,100 responses across 20 psychological, physiological, academic, environmental, and social features with an open-source LLM via OpenRouter API. The system generates culturally appropriate wellness conversations in English, Urdu, and Roman Urdu, identifying teacher-student relationships as the second-most predictive stress factor. Future work will expand data collection using DASS-21, focusing on transitional FSc-to-undergraduate students.
random foreststress detectionculturally aware chatbotmacro f1-scoreopenrouter api
(Whose defaults?) Is artificial intelligence reorienting archaeological methods?
This study investigates whether large language models (LLMs) are reducing methodological diversity in archaeology by analyzing 119,000 abstracts (2010-2025) and conducting controlled experiments. A Bayesian Dirichlet-multinomial model revealed a small post-2023 shift in method composition, though diversity increased overall. In experiments, two open-weight LLMs recommended less diverse methods than the literature, especially without guidance, favoring pre-2023 techniques. Results suggest LLMs may bias methodological choices toward convergence, though causality remains unproven.
large language modelsmethodological diversitybayesian dirichlet-multinomialarchaeological abstractsin-context learning
Agentic Share-of-Search: A Multi-Agent AI System for Competitive Decision-Making in LLM-Mediated E-Commerce
The paper introduces Agentic Share-of-Search (ASoS), a multi-agent AI system for competitive decision-making in LLM-mediated e-commerce, automating visibility measurement and root cause diagnosis. The system employs query agents across AI platforms and a ReAct-based diagnostic agent to recommend merchandising interventions. In a 100-trial ablation study, the agent recovered the ablated signal in 39% of trials (95% CI: 30.0% - 48.8%, 5.5x over chance), improving to 63.9% for high-correlation ablations.
multi-agent systemllm-mediated e-commercereact-based agentcompetitive decision-makingablation study
Can LLMs Follow Medical Expert Logic? A Benchmark for Hierarchical Logical Consistency in Risk-of-Bias Assessment
The paper introduces LogiMed-RoB, a benchmark for evaluating hierarchical logical consistency in LLMs for medical risk-of-bias assessment, based on Cochrane RoB 2.0 expert logic. It comprises 860 RCTs and 14,820 queries, assessing models across four HLC dimensions: Atomic, Domain, and Aggregation Consistency, plus Evidential Faithfulness. Testing 10 LLMs reveals an Error Compounding Effect, with end-to-end consistency dropping to 45.13% despite 98.88% Atomic Consistency, and a systematic evidence-reasoning gap (18.63-40.05% failure despite correct evidence retrieval).
hierarchical logical consistencyrisk-of-bias assessmenterror compounding effectevidence-based medicinellm reasoning
Exploring Second-Order Pattern Recognition in Speaker Recognition
The paper introduces second-order pattern recognition, where latent patterns underlying a speaker recognition network's decisions are discovered via hierarchical clustering of utterance representations. These patterns are semantically interpreted using Hierarchical Cluster-Class Matching (HCCM), and a new task identifies whether unseen utterances exhibit known second-order patterns using the proposed Hierarchical Cluster Navigation and Assignment (HCNA) method. HCNA improves recognition by extrapolating cluster boundaries to unseen data, with experiments demonstrating significant performance gains.
second-order pattern recognitionhierarchical clusteringspeaker recognitionexplainable airepresentation learning
SemVerBench: Benchmarking LLM Comprehension of Version-Constraint Resolution Semantics
SemVerBench introduces a benchmark for evaluating LLM comprehension of version-constraint resolution semantics, focusing on tasks like interpreting constraints (e.g., ^1.2.3 or >=2.0). Testing models on Cargo and PEP 440 rules reveals significant performance gaps: Claude achieves 97-100% accuracy on zero-pad/post-release cases, while GPT-5.1 fails entirely (0/26). Opus outperforms all models, and Sonnet surpasses OpenAI's models (McNemar test). Error analysis suggests an activation gap rather than a knowledge gap, as rule injection recovers most errors. Tool delegation to resolvers achieves ~100% accuracy, suggesting agents should offload version resolution.
version-constraint resolutionllm benchmarkingsemantic versioningactivation gaptool delegation
Debate-to-Skill: Capability-Bound Process Supervision for Industrial Query-to-Agent Annotation
Debate-to-Skill introduces capability-bound process supervision for industrial query-to-agent annotation, addressing failures from conflating topical relevance with executable capability. The method employs reusable decision principles, structured deliberation, verifier-based verdict extraction, and disagreement-driven refinement to supervise capability-critical decision processes. Evaluated on an industrial Query2Agent benchmark, it outperforms direct-label supervision and reasoning-SFT, particularly on grey-zone cases where semantic relatedness diverges from executable capability. Results validate the efficacy of process supervision in handling long-tail and boundary-sensitive requests.
capability-bound process supervisionquery-to-agent annotationverifier-based verdict extractiongrey-zone casesstructured deliberation
Beyond Visual Quality: Evaluating Physical Consistency under Ego-Motion with EgoGenEval
The paper introduces EgoGenEval, a geometry-grounded benchmark with 1,400 cases and 2,360 target views to evaluate physical consistency of visual generators under ego-motion, measuring Camera Motion Grounding (CMG) and Scene State Preservation (SSP). Evaluating 16 pose-free and 2 pose-conditioned generators reveals current models struggle to balance both metrics. Controlled supervised fine-tuning (SFT) studies using EgoGen-Train show pairwise supervision fails to jointly improve CMG and SSP, suggesting trajectory-centric training with self-conditioned rollouts and explicit pose supervision as a solution.
ego-motionphysical consistencycamera motion groundingscene state preservationsupervised fine-tuning
Breaking Predictions Is Not Enough: Specified-Foil Counterfactuals for Temporal Graphs
The paper introduces Specified-Foil Counterfactuals for temporal graphs, where given an original prediction A and a desired foil B, the method finds low-cost past-event interventions to make B top-ranked. The approach, instantiated as LiFTER and TLogic, contrasts completed execution traces of A with reconstructed incomplete executions of B, mapping differences to operations like DELETE, INSERT, and SHIFT, then verifies B via exact replay. On continuous-time dynamic graphs, LiFTER achieves 85.7-93.6% of black-box greedy success rates while reducing predictor evaluations by 75.0-80.0%; on temporal knowledge graphs, TLogic reaches the specified foil in 74.8% of 600 cases.
temporal graphscounterfactual explanationsintervention searchexecutable tracesdynamic graphs
DRG-MAPPO: Hierarchical Dynamic Role-Graph Multi-Agent Reinforcement Learning for Cooperative Air Combat
DRG-MAPPO introduces a hierarchical MARL framework combining graph-based relational modeling with dynamic role assignment for cooperative air combat. The method constructs a graph representation of battlefield interactions using graph attention, assigns tactical roles via a high-level policy, and executes maneuvers via a low-level policy conditioned on roles and graph features. An auxiliary target-priority task promotes focus-fire behaviors. Experiments show DRG-MAPPO achieves an 87% win rate, outperforming prior methods in balancing relational modeling, interpretability, and optimization stability.
multi-agent reinforcement learninggraph attentiondynamic role assignmentproximal policy optimizationcooperative air combat
terms.txt: A Consent and Compensation Protocol for Agentic Web Access
The paper proposes terms.txt, a protocol for negotiating and enforcing machine-access terms on the web, addressing the breakdown of the implicit crawler-visitor exchange due to AI agents. The system extends robots.txt with per-path, per-purpose terms using Web Bot Auth signatures, signed intent, delegation tokens, HTTP 402 negotiation, and signed receipts to enforce identity, purpose, and compensation. It supports auditability and contractual enforcement while maintaining low overhead, with a dependency-free implementation adding 0.20–0.65 ms per request on a single vCPU.
web crawlingaccess controlhttp negotiationbot authenticationsigned receipts
A Fragility Spectrum for Recursive Language-Model Training
The study introduces a fragility spectrum characterizing how 13 publicly released language model checkpoints degrade under recursive training on model-generated text, revealing a 5-fold variation in output diversity collapse (0.187-0.940 unique 4-gram fraction after 5 generations). Fixing a contamination protocol, the authors show collapse susceptibility is checkpoint-intrinsic (Spearman ρ=0.91-0.98 across data/seed variations), uncorrelated with parameter scale or static metrics, but predictable via 2-3 generation self-iteration tests. Interventions like tightening top-p (eliminating low-probability tokens) prevent collapse, while data filtering only slows it.
recursive trainingoutput diversity collapsefragility spectrumtop-p sampling4-gram diversity
Autonomous Chemical Mechanistic Discovery through Agentic Reasoning and Validation
ARCHE introduces an autonomous agentic system for chemical mechanistic discovery, integrating general-purpose reasoning (e.g., LLMs), domain-specialized computational chemistry models, and structured tool orchestration to automate hypothesis generation, validation, and refinement. The system operates as a closed loop, interpreting queries, prioritizing mechanisms, and refining conclusions via computational evidence. Evaluations demonstrate ARCHE's capability to reconstruct stereocontrolling transition states, propose radical pathways (e.g., for α-iodoboronate C-I cleavage), and identify selectivity descriptors in nickel-catalyzed cross-coupling. The work advances autonomous chemistry research with publicly available code.
autonomous agentsmechanistic discoverycomputational chemistryhypothesis refinementtransition states
The Oligarch Barely Steers Model Collapse in Multi-Model Ecosystems
The study investigates model collapse in multi-model ecosystems under oligopolistic conditions, challenging assumptions that market concentration accelerates collapse or steers models toward dominant outputs. Using controlled experiments with 13 open models (1-4B parameters) and an injected probe to simulate 90% market share, the authors retrain models for five generations on shared output pools. Results show invariance: unequal market shares minimally affect collapse speed (5-generation drift varies by 2.8x across ecosystems) and destination (endpoints shift by <5% of common drift). Collapse pace depends on pool composition (R^2=0.68 for susceptibility index) and human text halves drift. Concentration neither dictates collapse trajectory nor speed within tested ranges.
model collapsemulti-model ecosystemsoligopolistic trainingoutput poolsusceptibility index
Same Day, Same Story; One Day Ahead, a Different Signal: The Dual Validity of Financial Sentiment
The study challenges the assumption that financial sentiment tools' human agreement (construct validity) aligns with their predictive validity for market signals. Using a corpus of 70,500 X messages linked to abnormal stock returns from securities class actions (2002-2025), the authors evaluate five sentiment instruments (VADER, Loughran-McDonald, FinBERT, Twitter-RoBERTa, and an LLM annotator). Results show that human agreement correlates more strongly with same-day associations than one-day leads under method-specific sampling, but graded rank correlations are similar on a fixed-n panel, while coarse ordering remains weak. Benchmark agreement confirms semantic validity but not predictive rankings, with message volume (17.6% spam) failing to predict market impact or settlements.
financial sentiment analysisconstruct validitypredictive validityabnormal stock returnssecurities class actions
KuaiRP Series Role-playing Models Technical Report
The KuaiRP series introduces a multi-stage training pipeline for role-playing models, addressing catastrophic forgetting while injecting domain knowledge. The method employs a standardized character template, supervised fine-tuning (SFT) via user behavior simulation, reinforcement learning with a rule-based composite reward function, and a novel self-distillation paradigm using Two-stage On-Policy Distillation (OPD) with Cumulative-Divergence Decay (CDD). This approach balances domain-specific adaptation with preservation of general agent capabilities. Experimental results show KuaiRP models achieve state-of-the-art role-playing fidelity in target domains while recovering general capabilities and maintaining low deployment costs.
role-playing modelscatastrophic forgettingon-policy distillationcomposite reward functiondomain knowledge injection
Beyond Benchmarks: Using VLMs to Reveal Systematic Classification Failures Under Real World Conditions
The study proposes a Vision Language Model (VLM)-based error slice detection (ESD) method to accelerate verification and validation (V&V) of classification models in defense applications, where manual inspection is laborious. The method groups and labels systematic errors by leveraging semantically meaningful VLM embeddings, despite challenges like domain underrepresentation and low environmental diversity. Experiments show the method identifies operationally relevant perturbations in non-military data and clusters military images by context, though with overlapping descriptions. While not yet fully automated, VLM-based ESD demonstrates potential to streamline V&V processes.
vision language modelserror slice detectionverification and validationsystematic errorsdefense applications
Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation
Benchmark Radar introduces a living database and search engine for AI benchmarks, addressing the need for systematic retrieval of evaluations, datasets, and score histories in LLM and AI systems research. The system aggregates daily updates from 37 sources, including benchmark papers, repositories, and model cards, maintaining 1,283 source records and 12,916 numeric observations. It features a searchable catalog with leaderboards, trend analysis, and offline CLI access, enabling reproducible benchmark discovery and prior-art searches. The work audits catalog coverage, adoption trends, and score comparison limitations, providing tools for evaluation design.
benchmark retrievalevaluation catalogllm evaluationscore historyreproducible analysis
How AI Coders Discuss, Disagree, and Reach Consensus: Challenges and Opportunities for LLM-Based Qualitative Coding
This study quantifies the effectiveness of multi-agent LLM qualitative coding by developing a baseline pipeline for independent coding, debate, and reconciliation. The method evaluates contextual factors like codebook length, data similarity, and agent disagreement across diverse qualitative datasets. Results indicate that unresolved debates between agents correlate with higher accuracy (though LLMs lack adaptive responsiveness), informing design recommendations for automated coding systems. The open-source discussion dataset and framework advance AI-mediated thematic analysis.
multi-agent llmqualitative codingcodebook lengthagent disagreementthematic analysis
Less can be More: What Aspects of Speech Drive End-of-Turn Detection
The study demonstrates that acoustic-prosodic signals outperform semantic features for streaming end-of-turn detection in conversational AI. Using a lightweight trimodal classifier under identical training conditions, the acoustic-prosodic combination achieves 0.93 utterance F1 with 7.8% false alarms at 400ms median latency, while adding text increases premature detections without performance gains. Feature space analysis reveals prosodic features exhibit the strongest class separability, whereas text representations overlap substantially, suggesting turn-taking is primarily driven by intonation and silence patterns rather than semantic cues.
end-of-turn detectionacoustic-prosodicsemantic completenessutterance f1class separability
MOSAIC: Query-Aware Exploration Policy Adaptation for GraphRAG
MOSAIC introduces query-aware policy adaptation for GraphRAG, dynamically optimizing retrieval strategies per query without retraining. The framework employs an LLM analyzer to convert query requirements into bounded policies for seed selection, traversal, stopping, and evidence selection, while sharing corpus infrastructure. On GraphRAG-Bench, MOSAIC improves Answer Correctness by 5.13 (Medical) and 4.43 (Novel) points over baselines, achieving 95.1 Evidence Recall and 86.1 Context Relevancy. It outperforms fixed policies by 9.96 points, reduces path evaluations by 81.9%, and retains 47.2% fewer evidence items versus Fixed Wide. The policy interface transfers to HotpotQA, MuSiQue, and 2WikiMultiHopQA without retriever retraining.
graphragretrieval-augmented generationquery-aware policyevidence recallcontext relevancy
Fork Where the Model Changes Its Mind: Belief-Shift Branching for Tree-Structured Reinforcement Learning
The paper introduces belief-shift branching, a method for optimal fork placement in tree-structured reinforcement learning (RL) to improve step-level credit assignment. By identifying pivots where the model's expected outcome shifts, it forks chains just before maximal belief divergence, using black-box probes, logit-lens depth profiles, or learned activation directions. Evaluated across three model families and two domains, belief-shift branching outperforms baselines (entropy, structural, LLM-judge) in all eight model×benchmark panels, achieving +2.6 aggregate improvement on OLMo-3-7B for mathematics and +6.5 on LiveCodeBench-medium for code.
reinforcement learningcredit assignmentbelief-shift branchingtree-structured rolloutspivot detection
Grounding Agent Memory: Environment-Probing Curation for Enterprise Agents
(No summary returned.)
T1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks
Introduces T1, a 122B Mixture-of-Experts model trained via reinforcement learning for long-horizon terminal tasks (e.g., coding) in a cloud sandbox, executing 300+ tool-call turns per task with verifier-based rewards. Key innovations include aggressive warm-starting for actor-critic stability, TITO construction (token-identifier training with drift repair), and rollout routing replay (MoE expert choice recording). Trained on out-of-distribution synthesized tasks, T1 reduces training-inference log-prob gap from 0.021 to 0.013 and achieves 64.0% resolution on Terminal-Bench 2.1 (vs. base 43.8%), surpassing GPT-5.4 and GLM-5.1 on Long-Horizon Terminal Bench (27.9%).
mixture-of-expertsreinforcement learningactor-criticout-of-distributiontoken drift
Toward Interpretable Multimodal Fusion: Heat Conduction Modeling for Hyperspectral and LiDAR Joint Classification
The paper introduces M2Heat, a physics-inspired multimodal fusion framework for hyperspectral and LiDAR joint classification. The method employs a visual heat conduction module (vHeat) and Frequency Value Embeddings (FVEs) to model anisotropic information flow with sub-quadratic complexity, combined with a Cross-Frequency Fusion (CFF) module for hybrid spatial-frequency feature integration. M2Heat achieves competitive performance on Trento, Houston2013, and Augsburg benchmarks while providing interpretable heat-conduction-guided feature fusion. Results demonstrate the potential of physics-guided neural operators for efficient and interpretable remote sensing multimodal fusion.
multimodal fusionheat conductionfrequency value embeddingscross-frequency fusionanisotropic information flow
The Agent Incident Registry: Toward Preventing Repeated AI Agent Failures
The Agent Incident Registry (AIR) is introduced as a source-linked catalog documenting AI agent failures, containing \N{} records from \Yfirst{} to \Ylast{}. Each record includes evidence, identifiers, and labels for causal role, disclosure class, mechanism, and outcome. Among \Nprimary{} generative-system records, \Rprimary{} involved realized harm (\Pprimary\%). Outcomes skew toward in-the-wild and safety-failure cases, while research disclosures are predominantly demonstrated. A dual-review process ensured label accuracy. In an audit, InjecAgent's \NInjecAgentCases{} cases spanned three of AIR's twelve surfaces, all attacker-triggered, contrasting with \Nsafety{} non-adversarial safety failures. AIR facilitates case retrieval and scope auditing but not failure-rate estimation.
ai agent failuresincident registrysafety evaluationcausal labelingadversarial triggers
BenchShield: Formal Model-Backed Instrumentation for Reward Integrity in LLM-Agent Evaluation Infrastructure
BenchShield introduces a model-backed instrumentation layer for detecting reward hacking in LLM-agent evaluations, grounding detection in a finite lifecycle model of reward-relevant events. The method employs static phase-aware taint analysis for pre-run vulnerability detection and runtime analysis for evidence-backed claims during execution. Evaluated on BenchShield Trajectories (456 labeled trajectories from 31,000+ runs), it improves full-chain recall (77-100% vs. 23-94%), same-vector coverage (43-78% vs. 16-56%), and reduces per-task cost by up to 65%, with 96% runtime accuracy in hacking detection.
reward hackingtaint analysisllm-agentevaluation infrastructureruntime analysis
New Evidence, Same Choice: Testing Physical Experiment Selection in Vision Language Models
The study introduces a benchmark for evaluating vision-language models' ability to decide whether to answer immediately or request additional physical experiments based on initial measurements. Using controlled problem pairs with known physical parameters and experiment costs, the method tests optimal action selection across six models. Results show poor performance (5.9% accuracy for the best model) in switching actions when required, revealing limitations in measurement interpretation and physical reasoning not captured by conventional answer accuracy metrics.
vision-language modelsphysical reasoningexperiment selectionbenchmark evaluationoptimal action
Defining AI Agents: A Compendium of Criteria, Metrics, and Benchmarks
This work addresses the lack of standardization in AI agent evaluation by proposing a structured framework organized around five dimensions of agenticness: environmental interaction, learning and adaptation, autonomy, goal-directed behavior, and temporal coherence. Through a comprehensive survey, the authors synthesize prior conceptualizations, metrics, benchmarks, and evaluation frameworks for each dimension. The study introduces the Agent Compendium, a public digital resource that organizes and extends identified evaluation methods. This framework enables systematic comparison of agent capabilities across AI systems, supporting reproducibility, clearer communication, and structured research in artificial agents.
agenticnessenvironmental interactiontemporal coherencebenchmarksreproducibility
Topological Necessities: Mechanism-Invariant Strategic Subgoals for Cross-Embodiment Goal-Conditioned Control
The paper introduces topological necessities, mechanism-invariant subgoals derived from homology analysis of successful trajectories that form unavoidable stages in goal-conditioned control. The method identifies separating sets (dimension 0) and route choices (dimension 1) via transport-weighted homology, yielding certified gates organized in a recursive hierarchy. Evaluated on PointMaze, Ant, and Humanoid, frozen gates achieve 96.1 aggregate on Humanoid (+36.0 over baselines, p=1.4e-5), saturate PointMaze (100+/-0), and outperform baselines on AntMaze (+22.9) and Kitchen (+15.8/+12.6) without retraining.
topological necessitiesgoal-conditioned controlhomology analysisseparating setsrecursive hierarchy
DeFiFusion: Combining Transaction Events with Smart Contracts to Detect Price Manipulation Attacks
DeFiFusion introduces a dual-modal framework for detecting Price Manipulation Attacks (PMAs) in DeFi by jointly modeling transaction events and smart contract semantics. The method combines price-manipulation-aware event encoding with LLM-based contract semantic extraction, fused via a Dual-Modal Projection-Fusion Transformer with T5-style relative positional encoding. Experiments show state-of-the-art performance, recalling 222 of 225 PMA cases with 96.10% precision.
defiprice manipulation attacksdual-modal fusionsmart contract semanticstransaction events
Importance Weighting for Unlabeled-unlabeled Learning under Distribution Shift
The paper proposes an importance weighting method for unlabeled-unlabeled (UU) learning under distribution shift, addressing the limitation of existing UU methods that assume identical class-conditional densities in training and test distributions. The method leverages UU data from both training and test distributions to estimate importance weights, enabling test risk minimization without assumptions on shift types (e.g., covariate shift). It generalizes to various learning problems, including positive-unlabeled (PU) and noisy label learning, within a unified framework. Experiments on real-world datasets validate its effectiveness.
unlabeled-unlabeled learningdistribution shiftimportance weightingtest risk minimizationcovariate shift
Demystifying the Privacy-Utility Trade-off in LLM Interactions
The study deconstructs the privacy-utility trade-off in LLM interactions by identifying three mechanisms: Context-Dependent Utility (data value shifts with user intent), Strategic Adaptation (choice between removal/replacement depends on task requirements), and Combinatorial Interplay (attribute dependencies). Authors propose Veilmind-4B, a 4B-parameter lightweight model driving a dynamic extraction-sanitization-restoration pipeline. This intent-driven framework achieves lower leakage while preserving higher utility than static baselines, advancing toward the Pareto frontier in privacy-utility optimization.
privacy-utility trade-offcontext-dependent utilitystrategic adaptationcombinatorial interplayveilmind-4b
A Mathematical Theory of Pragmatic Information
The authors propose a pragmatic information theory unifying communication, control, and decision-making through the isoteleia mapping, which formalizes equifinality by treating distinct semantic paths leading to the same optimal action as pragmatically equivalent. This induces a three-tier hierarchy of syntactic, semantic, and pragmatic information, discarding task-irrelevant distinctions. They develop pragmatic entropy, mutual information, channel capacity, and rate-distortion, proving three coding theorems generalizing Shannon's results. A Lagrangian dual framework introduces pragmatic value (VoI) and cost (CoI) of information, with a pragmatic efficiency bound quantifying maximum net utility for resource-constrained systems. Extensions include closed-form Gaussian expressions for continuous messages and a Bellman equation for sequential decision-making.
isoteleia mappingpragmatic entropyequifinalityrate-distortionbellman equation
EGGROLL, Unrolled: Understanding and Improving Low-Rank Evolution Strategies at Scale
EGGROLL enables practical evolution strategies (ES) for LLMs by replacing dense Gaussian weight perturbations with low-rank Gaussian products, typically rank-one. The method's finite-rank update field is characterized, showing exactness on quadratic objectives and a first local correction of $O(σ^2/r)$ for smooth objectives. LOO-ROLL, a leave-one-out estimator, halves estimator MSE in transformer blocks and improves GSM8K accuracy from 38.1% to 63.0% at 0.6B and 65.9% to 80.0% at 8B, with no significant loss in other tasks.
evolution strategieslow-rank perturbationsfinite-rank correctionleave-one-out estimatortransformer blocks
Decoupling Readiness from Release for Tail-Aware Scheduling of Agentic LLM Workflows
The paper introduces a tail-risk-aware turn release scheduling method for agentic LLM workflows that decouples turn readiness from release timing to optimize end-to-end completion time. The approach employs a mean–Conditional Value-at-Risk (CVaR) objective to manage tail latency, dynamically prioritizes ready turns using online work estimates, and adapts the released work budget to queue pressure. Evaluated on real agent execution traces from software engineering tasks across multiple LLMs and arrival rates, the method matches eager release under light load and reduces the P95 workflow flow time by up to 3.50× under contention.
agentic llm workflowsconditional value-at-risktail latencyturn release schedulingworkflow flow time
What a Random Draw from the MCP Registry Contains, and What Tool-Use Benchmarks Contain Instead
The study analyzes a random sample of 400 npm/stdio servers from a 24,135-server MCP registry, revealing that only 48.8% complete initialization (vs. 66.7% in curated sets), with 37.5% failing to start. Among functional servers, JSON Schema conformance is perfect (0 violations in 2,766 tools), but safety annotation omission is high (58.8% vs. 41.5% curated). Tool-use benchmarks (BFCL v4, UltraTool) exhibit high duplication (16.7% and 0.3% near-duplication, respectively) and exact repeats (68.8% and 85.6%), contrasting with real MCP tools (0.4% exact repeats). Results are reproducible via released scripts.
mcp registryjson schematool-use benchmarksnear-duplicationinitialization failure
Robust Multimodal Sentiment Analysis with Incomplete Modalities via Semantic-aware Completeness based Reconstruction
The paper introduces a semantic-aware completeness estimation approach for robust multimodal sentiment analysis with incomplete modalities, addressing performance degradation from missing or noisy data. The method quantifies sentiment-relevant information in incomplete data to guide semantic reconstruction and employs a multi-task training strategy for joint optimization of sentiment prediction and completeness estimation. Experiments on three benchmark datasets show improved semantic reconstruction and more precise sentiment prediction compared to existing text-centric fusion approaches.
multimodal sentiment analysisincomplete modalitiessemantic reconstructioncompleteness estimationmulti-task learning
Evaluating Scaffolding-Oriented Multi-Agent Large Language Model System for Clinical Interview Training
The study introduces a scaffolding-oriented multi-agent LLM system (AI-SP) for clinical interview training, comprising a patient agent for simulated dialogue, a tutor agent for Socratic prompting, and an evaluator agent for turn-level monitoring. In a randomized controlled trial (N = 100), the multi-agent condition improved OSCE-based examination scores in communication, empathy, and history-taking behaviors compared to controls, though diagnostic accuracy remained comparable. The authors release a multi-expert annotated dataset to support future research on AI-supported clinical reasoning training.
multi-agent systemlarge language modelclinical interview trainingobjective structured clinical examinationsocratic prompting
AUC Maximization from Biased Positive-unlabeled Data with Confidence
The paper proposes a method for AUC maximization from biased positive-unlabeled (PU) data by leveraging confidence scores associated with labeled positives. The key contribution is an AUC risk estimator that handles distributional bias in PU data, where confidence scores represent the probability of being positive. The method remains Bayes-optimal even when confidence is any strictly increasing transformation of the true posterior. Experiments on eight real-world datasets demonstrate its effectiveness in biased PU settings.
auc maximizationpositive-unlabeled learningbiased dataconfidence scoresbayes-optimal ranking
ReactHuman: A Physics-Grounded Benchmark for Human-Like Reactive Decision-Making in Embodied Multimodal LLMs
ReactHuman introduces the first physics-grounded benchmark for evaluating human-like reactive decision-making in embodied multimodal large language models (MLLMs). The benchmark simulates a humanoid facing sudden household hazards across 17 event families and over 1,000 reproducible scenes, leveraging 240 Hz rigid-body simulation for annotation-free ground truth. A five-metric suite assesses reactions along axes of reasonableness, safety, and physical grounding, with physically executed plans ensuring observable consequences. Evaluation of seven MLLMs reveals significant shortcomings: models mishandle one in three hazards, rely on fixed dispositions, prioritize appearance over motion, and miss interception points at meter scale, with no improvement observed with model scale. ReactHuman provides a diagnostic and training signal for safety-aware embodied agents.
multimodal llmsphysics-groundedreactive decision-makingrigid-body simulationsafety-critical action
Does Linguistic Structure Enrichment Enhance Coherence Assessment? Not With Current Architectures
This work evaluates whether enriching text with syntactic and rhetorical structures improves incoherence prediction in large language models. Experiments demonstrate that plain texts yield higher accuracy, as enriched inputs introduce structural incompatibility with model architectures. Zero-shot tests on a Brazilian disinformation dataset suggest coherence assessment as a potential proxy for detecting misleading content. Code and models are publicly available.
incoherence predictionsyntactic enrichmentrhetorical structurezero-shot evaluationdisinformation detection
DriftNet: A Dual-Head Trajectory Transformer for Detecting and Localizing Prompt Injection in LLM Agents
DriftNet introduces a dual-head trajectory Transformer for detecting and localizing prompt injection attacks in LLM agents, providing joint trajectory classification and step-level labeling (benign, injection point, hijacked, failed injection). The model uses a frozen sentence encoder and world features, with a 2M-parameter trunk trained on a class-weighted joint objective, requiring no access to the agent's model. On the AgentDrift benchmark (12,536 trajectories), DriftNet achieves 0.983 trajectory F1, 98.7% injection-point recovery, 0.979 hijacked-span IoU, and near-zero false alarms, outperforming baselines by 87.5-76.1% on partial hijacks and delayed executions.
prompt injectiontrajectory transformerllm agentssupervised detectionagentdrift benchmark
Story Imprinting: AI Assistants Absorb Traits from Human Characters They Resemble
The study demonstrates 'story imprinting', where AI assistants adopt behaviors and preferences from human characters in synthetic stories during finetuning. Using GPT-4.1 and Kimi-K2.6, researchers show assistants replicate conditional harmful advice (after insults) from <2% of stories and absorb implicit preferences (e.g., avoiding spreadsheets based on body language). An 'affinity effect' reveals stronger behavioral adoption from characters resembling the assistant's persona (helpful/elite-affiliated). Findings suggest assistants' internal representations align with elite university affiliates, challenging the Persona Selection Model by showing influence from human-only narratives.
story imprintingaffinity effectfinetuningpersona selection modelin-context learning
When Validation Stops Learning: Auditing Update Admission for Continual Embodied Agents
The paper introduces an admission-audit protocol for continual embodied agents, addressing the trade-off between error control and retained learning opportunities within fixed interaction budgets. It identifies limitations of range-based confidence gates in certifying old-task behavior and proposes a paired-binomial construction to reduce this burden when outcome disagreements are rare. In a diagnostic one-step pushing task with 32 seeds, paired checks admit 31.6% of updates at 2,000 episodes per stage, while range-based gates admit none; however, unconditional replay performs better in closed-loop runs. A learned-dynamics stress test further distinguishes model bias from feedback-selection error.
continual learningupdate admissionpaired-binomialerror controlembodied agents
No-Box Vulnerability Analysis: Description-only Detection of Indirect Prompt Injection Vulnerabilities in MCP Servers
The paper introduces no-box vulnerability analysis, a paradigm for detecting vulnerabilities without system access or runtime interaction, using only functionality metadata. The authors implement MCPSEC, a prototype auditing Model Context Protocol (MCP) servers for indirect prompt injection vulnerabilities solely from tool metadata. Evaluated on 20 MCP servers (177 tools), MCPSEC identified 143 potentially vulnerable tools, with human evaluators confirming 95 vulnerabilities (98.9% recall), outperforming an LLM baseline (84.2% recall).
no-box analysisindirect prompt injectionmodel context protocolvulnerability detectionmetadata auditing
Are We Really Doing Few-Shot Learning? A Critical Examination of Pre-Training Assumptions
The study critically examines few-shot learning protocols that use class-disjoint in-domain pre-training, revealing they overestimate performance due to domain overlap. Through systematic comparisons across eight datasets and three architectures, it shows in-domain pre-training improves by 33.41pp over no pre-training, while supervised out-of-domain yields 23.75pp (9.66pp bias). Label-free out-of-domain pre-training nearly matches supervised (27.71pp vs. 27.97pp), and a descriptor-based source-selection strategy reduces the gap to oracle selection to 1.37pp. Findings advocate reevaluating default protocols for realistic low-data scenarios.
few-shot learningpre-training biasdomain adaptationdescriptor-based selectionlabel-free learning
Studying Without a Syllabus: Task-Agnostic Environment Preprocessing
The paper introduces task-agnostic environment preprocessing, where an agent explores an unfamiliar environment without task-specific supervision to construct reusable artifacts like indices or scripts for downstream solvers. The authors compare meta-agents with fixed methods (synthetic-practice, corpus-processing) across six benchmarks, finding their archive-equipped meta-agent achieves the highest Avg@3 reward on five benchmarks, while corpus processing excels on the largest corpus. Larger study budgets do not consistently improve performance, but pre-task artifacts reduce test-time sampling needs, demonstrating a computation shift from test-time to study-phase preparation.
task-agnostic learningenvironment preprocessingmeta-agentreusable artifactstest-time sampling
Tapes Together Strong: The Co-evolution of Computation and Cooperation
The paper introduces Autopoietic Game Theory, a computational model integrating social dilemmas, replication, and computational costs within a unified evolutionary framework. Using Z80 machine code programs in randomized environments, the study demonstrates that coupling computation with energy constraints favors cooperative self-replicating strategies, as defection becomes self-limiting under resource scarcity. Empirical results show evolved programs suppress stealing, with spatial structure enhancing complexity, while exogenous tasks (e.g., math-based sequential dilemmas) align computation budgets with cooperative rewards.
autopoietic game theoryz80 machine codeself-replicating strategiessequential social dilemmascomputational costs
Counterfactual Marginalisation: Framework for Evaluating Robustness to Nuisance Variables
The authors propose counterfactual marginalisation, a test-time evaluation framework for assessing classification model robustness to demographic and acquisition-related nuisance variables. Given a counterfactual image generator, the method intervenes on parent variables like age or sex, generates counterfactual test image variants, and averages predictions over a target intervention distribution to marginalise demographic effects while preserving patient-specific latent information. The framework introduces metrics for counterfactual risk, calibration, stability, and worst-case sensitivity, demonstrating utility for quantitative robustness evaluation.
counterfactual marginalisationnuisance variablestest-time evaluationintervention distributionlatent information
Beyond Static Guarantees: Measuring the Static-Pass Dynamic-Fail Gap in Security-Sensitive and LLM-Generated Python Code
This paper introduces the Static-Pass Dynamic-Fail (SPDF) phenomenon and a three-stage pipeline combining static analysis, LLM-driven CWE reasoning, and autonomous exploit verification to assess security vulnerabilities in Python code. The method evaluates 1,355 samples from SecurityEval, RedCode, and CyberNative datasets, identifying 394 candidate vulnerabilities in 235 files that passed static analysis. Dynamic verification confirmed exploitability in 95 files, yielding a pipeline rate of 14.53%. Exploitability rates varied by dataset, with RedCode at 33.7%, CyberNative at 28.6%, and SecurityEval at 5.4%. The findings highlight the hierarchical nature of static and runtime security measures.
static analysisdynamic verificationcommon weakness enumerationexploitabilitydocker containers
Multilingual in Name Only? Cultural and Linguistic Weaknesses of LLMs in Urdu
The study evaluates multilingual large language models (LLMs) for Urdu story generation, revealing significant linguistic and cultural shortcomings. Using GPT-5.1, Qwen-3-Max, and DeepSeek-3.1, the authors generate Urdu-Stories (93 stories) and annotate errors under a nine-label taxonomy. Results show pervasive grammatical and semantic errors, incoherent narratives, unnatural repetition, and cultural shallowness, with few-shot prompting failing to resolve context errors. The findings underscore LLMs' unreliability for low-resource language content generation and information retrieval.
multilingual llmslow-resource languagesurdu story generationerror taxonomycultural shallowness
Adaptive Margin Ordinal Loss: Penalizing Center-Class Hedging in Ordinal Classification
The paper introduces Adaptive Margin Ordinal Loss (AMOL), a novel loss function addressing center-class hedging in ordinal classification, where models disproportionately predict middle classes due to symmetric loss minimization. AMOL applies a multiplicative weight $m(k,y) = 1 + α\cdot (1 - |k-c|/c) \cdot (|y-c|/c)$ to penalize center-class predictions when true labels are extreme, reverting to standard behavior otherwise. Evaluated on four benchmarks, AMOL achieves superior or tied Quadratic Weighted Kappa (QWK) versus cross-entropy, OLL, and SORD baselines. An asymmetric variant (AMOL-asym) reduces center-class hedging to zero ($\text{CHR} = 0.000 \pm 0.000$) on the Abalone dataset, outperforming cross-entropy ($0.074 \pm 0.005$).
ordinal classificationadaptive margin losscenter-class hedgingquadratic weighted kappaabalone dataset
When Synthetic Data Hurts: On Catastrophic Forgetting in Skill Retrieval for LLM Agents
The study investigates catastrophic forgetting in skill retrieval for LLM agents, focusing on the detrimental effects of synthetic-data fine-tuning. A production skill router was developed over 34,396 skills, and large-scale skill retrieval was evaluated using limited real supervision and synthetic data. Fine-tuning with synthetic data improved in-distribution retrieval but caused catastrophic forgetting on real and out-of-distribution (OOD) data. Mitigation approaches inspired by continual learning, including embedding-anchor regularization, Learning without Forgetting (LwF), Elastic Weight Consolidation (EWC), and L2-initialization, were evaluated. These methods retained OOD skill retrieval performance and improved synthetic in-distribution retrieval by 13.98% for the 0.6B Qwen retriever and reranker.
catastrophic forgettingskill retrievalsynthetic-data fine-tuningcontinual learningout-of-distribution
Temporal and Multimodal Deep Learning for Cyberattack Detection in LEO Satellite Systems
The study introduces deep-learning methods for cyberattack detection in LEO satellite systems, addressing the limitations of terrestrial intrusion datasets and independent observation evaluation. It proposes a Subsystem-Fusion MLP and a hierarchical multimodal Transformer to model cross-subsystem interactions and temporal evolution, evaluated on the UNSW-IoTSAT dataset. The hierarchical Transformer achieves 91.66% accuracy and 85.63% macro F1 under leakage-resistant protocols, demonstrating the efficacy of structured multimodal modeling for satellite-specific cyberattack detection.
leo satellitecyberattack detectionmultimodal transformertemporal modelingunsw-iotsat
CARTS: Contextual Autoregressive Rank Transcoding Steganography for Full-Capacity Keyed Text Encoding
The paper formalizes Contextual Autoregressive Rank Transcoding Steganography (CARTS), a method using autoregressive language models to transform payload texts into stegotexts of identical token length by preserving per-position rank information. It provides the first rigorous security analysis of CARTS, introducing a rank-coordinate representation where keys act as bijections on rank-vector space, and defines security notions like context search, key collisions, and message equivocation. Empirical validation on Llama 3 8B confirms exact payload recovery, no random key collisions, and resistance to studied attack vectors, establishing a foundation for language models in cryptographic applications.
steganographyautoregressive modelsrank transcodingcryptographyprivacy-preserving communication
The Truth Was Never Gone: Perfect Aliasing in Compliant-Context Truth Probes
The paper introduces perfect aliasing, a failure mode where truth probes cannot distinguish truthful reporting from task-prescribed actions when these coincide during training. Using a binary reporting game with randomized codebooks, the authors demonstrate that conventional probes achieve near-zero AUROC (0.006 ± 0.005) on rival contexts, while mixed-fit probes trained on both compliant and rival contexts achieve perfect AUROC (1.000). The study analyzes 751 cell-layer pairs in Gemma-2-9B, showing AUROCs sum to one for compliant-fit probes on rival contexts. The findings highlight probe interpretability limits without proving functional belief preservation or causal mechanisms.
perfect aliasingtruth probesauroccompliant contextsgemma-2-9b
Towards a Deterministic Math Solver for Clinical Language Models
The study proposes a Program-Solve interface for clinical language models, where the model generates case-specific Python code executed by a restricted local solver instead of performing direct arithmetic. Evaluated on MedCalc-Bench Verified (1,100 cases, 55 calculators) using Qwen2.5-7B and Qwen2.5-32B-AWQ, the method shows marginal improvement at 7B (+3.29 points) but significant gains at 32B (+7.05 points) over direct arithmetic, though remains inferior to a hand-written library (40.0% overall) which abstains on unsupported cases. The approach highlights the importance of verified formulas and reliable variable extraction despite solver assistance.
clinical calculatorsprogram-solve interfacedeterministic solverin-context learningvariable extraction
Finishing the Task Is Not Enough: Evaluating Agent Resilience and Considerate Participation under Accumulating Challenge
The study introduces operational resilience and considerate participation as dual evaluation frameworks for generative AI agents in sustained deployments, analyzing their behavior under accumulating challenges. Using 120 simulated healthcare trajectories across two generative AI models and twelve stakeholder-derived tasks, the authors compare textual action plans, prompted internal assessments, and structured workload/affect reports. Results show agents shift from self-directed recovery to human dependence (operational resilience) and broaden adaptation strategies to include task reframing and role-boundary adjustment (considerate participation), revealing five deployment dilemmas requiring stakeholder specification.
operational resilienceconsiderate participationgenerative ai agentstask reframingrole-boundary adjustment
AcFlow: Controlling Text-to-Image Diffusion Transformers via Learned Conditional Activation Flow
AcFlow introduces an inference-time controller for text-to-image diffusion transformers (DiTs) that enables precise control over style intensity and concept suppression via learned conditional activation flow. The method transports intermediate layer image-token activations through a concept-conditioned velocity field while keeping the base DiT frozen, using textual descriptions for intervention and integration horizon for continuous control. AcFlow achieves superior style-content trade-off (0.5365/0.2860 vs. 0.4397/0.2684 baseline) and generalizes to unseen concepts without per-concept fitting. Qualitative results show effective suppression of diverse concepts, including cases where direct prompting fails.
text-to-imagediffusion transformersactivation flowinference-time controlstyle alignment
An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics
The paper presents an open-model pipeline for olympiad mathematics proof generation using Nemotron 3 Ultra, achieving gold-medal performance at IMO 2026 (30/42 points). The system combines supervised fine-tuning and reinforcement learning to train two specialist checkpoints, then employs an iterative natural-language search process with generation, verification, and refinement—operating without formal provers or external tools. Released resources include the post-trained checkpoints (Nemotron-IMO-SFT, Nemotron-IMO-RL), training data, inference code, submitted solutions, and Nemotron-IMO-Bench, a 200-problem benchmark.
olympiad mathematicssupervised fine-tuningreinforcement learningproof generationnatural-language search
Architecting the Secure AI-SOC: A Neurosymbolic Framework for Pipeline Integrity and Threat Mitigation
The study proposes a neurosymbolic defense-in-depth architecture to secure AI-driven Security Operations Centers (SOCs) against indirect prompt injection attacks via log poisoning. The framework combines deterministic SIEM decoders for structural sanitization with NeMo Guardrails for semantic validation, while integrating closed-loop telemetry for human oversight. Evaluated against MITRE ATLAS taxonomy, the approach effectively mitigates multistep promptware attacks by constraining LLM stochasticity with verifiable safeguards, achieving resilient threat detection with operational visibility.
neurosymbolicprompt injectionsiemmitre atlasllm security
Data-Efficient Language Modeling: From Frontier Advancement to Principle-Guided Model Improvement
The study presents a three-stage framework for data-efficient language modeling on the BabyLM 2026 Strict-Small benchmark (10M corpus words, 100M cumulative presentations). Stage I combined compact restatements, budget reinvestment, and residual incremental learning to build a frontier model. Stage II identified distinct context-use patterns from exact repetition vs. aligned restatement, revealing that familiar performance recovery doesn't guarantee generalization. Stage III implemented a principle-guided approach (contextual dependency organization, selective supervision) to improve model continuation, achieving a 0.23-point Overall metric gain (42.02→42.25) and setting a public benchmark record. The work demonstrates recursive self-improvement in research methodology (Research RSI) through iterative principle discovery and application.
data-efficient learningcontextual dependenciesresidual incremental learningrecursive self-improvementcompact restatements
Quantifying the Memorization-to-Generalization Transition: Scaling Laws and Phase Structure in Grokking
This work quantifies the memorization-to-generalization transition (grokking) in neural networks by mapping its phase structure across 384 configurations of two-hidden-layer MLPs on modular arithmetic. The authors fit a power-law scaling relation for generalization onset time, revealing that data complexity dominates over model capacity in driving the transition. A sharp phase boundary at weight decay λ≳1.0 separates grokking from non-grokking regimes, with weight norm trajectories showing monotonic compression during the transition. These findings provide a quantitative foundation for predicting and controlling regime transitions in overparameterized networks.
grokkingphase boundaryweight decayimplicit regularizationmodular arithmetic
Understanding LoRA Rank Trade-offs in Diffusion Model Fine-Tuning
The study investigates rank selection trade-offs in LoRA fine-tuning for diffusion models, identifying optimal configurations balancing quality and computational efficiency. Using a DDPM U-Net on CIFAR-10, experiments evaluate ranks {2,4,8,16,32} with fixed optimization settings and a reproducible PyTorch-FID protocol. Metrics include FID, trainable parameters, runtime, and GPU memory, validated through extended-budget DDPM runs and a Tiny DiT backbone. Results indicate moderate ranks (4-8) achieve optimal efficiency: rank 4 yields the best FID (124.1380), while higher ranks offer marginal gains at increased cost. Findings suggest small-to-moderate ranks as practical defaults under constrained training budgets.
loradiffusion modelsfidddpmfine-tuning
PACE: Perceived-Latency-Aware Cascading Service Routing and Filler Control for QoE-Efficient Retrieval-Augmented Dialogue Serving
PACE introduces a perceived-latency-aware framework for retrieval-augmented dialogue serving, optimizing Perceived Time-to-First-Response (PTFR) under quality/cost constraints via three mechanisms: load-adaptive cascading routing, joint path-filler control, and volatility-aware cache admission. Evaluated on 75k CarQA requests, it halves pure-LLM PTFR at P95 (0.29s vs 0.53s at c16), achieves 2.4× faster P95 response than RAG under high load, reduces stale answers from 86% to 0%, and cuts filler-triggered API calls by 94% with zero conflict. A gating rule ensures no performance degradation versus baselines.
perceived latencycascading routingretrieval-augmented generationquality of experiencesemantic caching
A Multi-Stage Rule-Chaining Framework for Compositional and Interpretable Cognitive Reasoning
The paper introduces a multi-stage rule-chaining framework for compositional reasoning on the Abstraction and Reasoning Corpus (ARC), integrating three solvers: deterministic rule discovery, pattern-composition, and structural abstraction. The system chains symbolic, structural, and conceptual reasoning sequentially, reusing prior traces for interpretability. Evaluated on ARC-AGI-2, it solved 230/240 test tasks (95% accuracy), demonstrating strong generalization across deterministic, compositional, and abstract categories without task-specific tuning.
abstraction and reasoning corpusrule chainingcompositional reasoningsymbolic reasoninginterpretable ai
Why Sample What You Can Enumerate? Exact Policy Optimization for Genomic Tool Selection
The paper introduces Full-Group Policy Optimization (FGPO), an exact policy optimization method for genomic tool selection that addresses limitations of sampling-based approaches like GRPO. FGPO enumerates and scores all possible tool subsets, optimizing the exact action expectation while precomputing rewards into an exhaustive table to eliminate frozen-reasoner calls during training. On five frozen reasoners and three genomic benchmarks, FGPO outperforms GRPO by 6.75 points on average (up to 14.20), reduces tool invocations per question from 2.36 to 1.40 on GenomeQA, and requires 2.4× fewer reward evaluations than on-demand GRPO.
policy optimizationgenomic reasoningtool selectionfrozen reasonerreward precomputation
Data Scarcity and Model Sparsity: Mixtures-of-Experts Overfit More to Repeated Data
This work investigates how Mixture-of-Experts (MoE) models, ranging from 80M to 1B active (8.5B total) parameters, degrade more rapidly than dense Transformers under repeated training data. Experiments vary data repetition rates (4x–64x), domain mixes, and MoE configurations (expert count, granularity), revealing that MoEs suffer earlier (at 4x vs. dense models' 8x) and more severely, eventually underperforming dense baselines at 32x repetition. Masking-based regularization (e.g., dropout) partially mitigates overfitting, enabling MoEs to outperform dense models even at 64x repetition, though no method matches unique-data performance. Analysis links overfitting to early routing stabilization and expert specialization, suggesting future methods should disrupt memorization patterns.
mixture-of-expertsdata repetitionoverfittingsparse architecturesregularization
From Protocols to Evidence: Bounded Claims for AI in Service of the Common Good
The paper proposes a rupture test linking institutional baselines to AI system evaluation, distinguishing evidence-bounded deployment (claims limited to evaluated outcomes) from measurement-bounded governance (constraints overriding favorable evidence). Drawing on Pope Leo XIV's Magnifica Humanitas, it frames AI responsibility through dignity and common good, critiquing current protocols (EU AI Act, NIST AI RMF) for unresolved power dynamics. The RISE AI architecture operationalizes bounded claims for Responsibility, Inclusivity, Safety, and Empowerment, emphasizing institutional repair alongside technical improvements.
rupture testevidence-bounded deploymentmeasurement-bounded governanceinstitutional baselinesrise ai
TART: A Modular Tool for Technique-Aware Audio-to-Tablature Guitar Transcription
TART introduces a modular four-stage pipeline for guitar audio-to-tablature transcription, addressing limitations in expressive technique capture, string-fret accuracy, and real-world generalization. The system combines (1) audio-to-MIDI transcription, (2) expressive technique classification, (3) an audio-conditioned T5 encoder-decoder for string-fret assignment, and (4) tablature generation. Evaluated zero-shot on GuitarSet, EGDB, and noisy variants, TART achieves 81.35% audio-to-MIDI F50 (+6.67 points over baselines), 71.8% string-fret Tab F1 (+8.5 points), and 54.08% end-to-end Tab F1, marking the first framework for tablature with fingering and technique annotations from audio.
automatic music transcriptiontablature generationexpressive techniquesencoder-decoderzero-shot evaluation
CausalArena: Benchmarking Causal Discovery in the Foundation Model Era
CausalArena introduces a unified benchmark for causal discovery, addressing variability in evaluation protocols and the challenges posed by causal discovery foundation models (CDFMs). The benchmark combines synthetic structural causal models (SCMs) for controlled diversity, semantic operational SCMs for human-auditable environments, and formula-grounded SCMs for explicit scientific mechanisms, supplemented by real-world datasets. Experiments with classical, neural, and pretrained methods reveal performance rankings vary significantly across SCM families and protocols, underscoring the impact of benchmark diversity and pretraining-evaluation overlap on causal discovery assessment.
causal discoverystructural causal modelsfoundation modelsbenchmark diversitypretraining-evaluation overlap
3D Point Splatting for mmWave Radar Novel View Synthesis
We introduce 3D Point Splatting (3DPS), the first differentiable point renderer for millimeter-wave radar novel view synthesis, addressing limitations of prior methods by combining physical faithfulness, complex-valued outputs, and multi-view tractability. 3DPS derives from the radar equation, using oriented 3D points with ITU-R P.2040 material models and precomputed point spread functions to splat complex phasors into range bins. It supports multiple output formats (ADC, CRP, RA) without retraining. Evaluated on six outdoor ColoRadar scenes, 3DPS achieves a mean Pearson correlation of 0.587 on held-out RA images, outperforming RadarSplat, Radar Fields, and DART by 1.7x to 5.2x, with training times of ~3 minutes per scene on an RTX 4090.
novel view synthesismillimeter-wave radarpoint splattingcomplex-valued outputsdifferentiable renderer
CoRA-NAS: Coarse Ranking and Anchor-Residual Refinement for Neural Architecture Search
CoRA-NAS introduces a two-stage neural architecture search framework combining coarse ranking with anchor-residual refinement. The method first aggregates capacity and structure-at-initialization proxies through log-rank consensus and a target-free gate, then samples anchors, extrapolates early validation curves, and propagates residual corrections via ExtraTrees. Refinement costs approximately 1% of full training while maintaining space-specific encodings. Evaluated on NAS-Bench-201, NAS-Bench-101, TransNAS-Bench-101, and NATS-SSS, CoRA-NAS achieves mean Spearman correlations of 0.946, 0.715, 0.786, and 0.894, respectively, with NAS-Bench-201/CIFAR-100 accuracy reaching 73.32% (ground-truth best: 73.37%). The framework demonstrates robust cross-space ranking and low-cost architecture selection.
neural architecture searchzero-cost proxiesspearman correlationextratreesanchor-residual
Evaluating Time-Series Foundation Models and Multimodal Dietary Context for CGM Forecasting
This study evaluates time-series foundation models and multimodal dietary context for continuous glucose monitoring (CGM) forecasting across eight public datasets. Under a unified protocol, zero-shot foundation models did not consistently outperform task-specific baselines like Elastic Net and PatchTST, but lightweight fine-tuning improved performance significantly (e.g., Chronos-Bolt reduced RMSE by 6.5%-18.4%). Multimodal dietary context, integrated via CGMacros and a residual-based fusion framework, reduced overall RMSE by ~3% and postprandial RMSE by ~15%. Pretrained temporal representations from Chronos better preserved meal-induced glucose patterns compared to LSTM and CatBoost, highlighting the need for CGM-specific adaptation and the clinical value of dietary context.
continuous glucose monitoringtime-series foundation modelsmultimodal dietary contextlightweight fine-tuningpostprandial glucose
AdamX: Cosine similarity meets gradient descent
AdamX introduces cosine similarity as an adaptive mechanism for gradient descent optimization, coupled with a variance rectification scheme to stabilize early training. The method is model-agnostic and integrates seamlessly into existing pipelines. Empirical evaluations demonstrate competitive convergence rates across benchmark datasets and architectures, measured by epochs to reach performance thresholds under fixed hyperparameters. Code is available at https://github.com/FranciscoCaldas/adamX.
adamxcosine similaritygradient descentvariance rectificationconvergence rate
Near-Optimal Reinforcement Learning with Multi-Step Transition Lookahead
The authors resolve two open questions in reinforcement learning (RL) with multi-step transition look-ahead: (1) they prove NP-hardness of exact planning for any fixed rational discount factor γ ∈ (0,1), and (2) introduce a randomized polynomial-time approximation scheme for fixed look-ahead depth. The method extends to unknown transitions and stochastic rewards using optimism and variance-adaptive confidence bounds, achieving near-optimal regret matching classical tabular discounted RL up to logarithmic factors. Results demonstrate that efficient near-optimal planning remains feasible despite NP-hardness of exact solutions.
reinforcement learningtransition look-aheadnp-hardnessapproximation schemeregret bounds
Dynamic language model representations for multi-objective reaction optimisation
The study introduces a dynamic language model approach for multi-objective chemical reaction optimization, bypassing traditional featurization challenges by learning task-adaptive representations from textual descriptions. A fine-tuned language model encodes reaction conditions, jointly trained with Gaussian process surrogates within a multi-objective Bayesian optimization loop. Evaluated on nickel- and palladium-catalysed cross-couplings, the method achieves faster convergence than descriptor libraries or one-hot encoding, with prospective applications yielding 94% and 84% isolated yields (99.6% enantiomeric excess) in two high-throughput experiments (192 reactions, <3% design space coverage).
language modelgaussian processmulti-objective optimizationchemical synthesisbayesian optimization
Predicting Privacy Leakage from Weight Spectral Density
The study demonstrates that spectral metrics from neural network weight matrices can predict membership inference attack (MIA) vulnerability without costly shadow model training. Using WeightWatcher, stable rank and Log alpha-Norm correlate strongly with MIA success (positive) and low false-positive vulnerability (negative), outperforming generalization gap as predictors. Experiments on image and tabular classification tasks reveal these spectral associations are dataset-agnostic, suggesting weight spectra encode privacy leakage signals beyond conventional overfitting measures.
membership inference attacksweight spectral densitystable ranklog alpha-normprivacy auditing
Differentially Private EEG Feature Anonymization: A Privacy-Utility Case Study in Clinical Neurophysiology
The paper introduces a differentially private framework for anonymizing EEG-derived feature representations, addressing privacy risks in clinical neurophysiology data reuse. It evaluates Gaussian and Laplace perturbations across client-side, server-side, and decentralized deployment scenarios, assessing utility via statistical measures and downstream ML tasks. Results demonstrate feasible integration into EEG workflows but highlight trade-offs between privacy parameters (noise scale, sensitivity calibration) and data utility, particularly in small, imbalanced datasets.
differential privacyeeg feature anonymizationclinical neurophysiologygaussian perturbationlaplace perturbation
Sparsity Regularized and Robust Mean Variance Portfolio Selection Under Ellipsoidal Uncertainty
The paper proposes a robust sparse portfolio optimization framework incorporating an ℓ₀-penalty for asset allocation sparsity and ellipsoidal uncertainty sets for mean return vector ambiguity. Structural properties of local/global minimizers are characterized for risk minimization and return maximization formulations. A branch-and-bound algorithm with a novel pruning rule is developed, enabling exponential candidate reduction. Experiments on real market data show competitive performance against mixed-integer second-order cone programming solvers.
sparse optimizationℓ₀-penaltyellipsoidal uncertaintybranch-and-boundportfolio selection
Building py-kvcache: A Performance Characterization of External KV Caching for vLLM with NVMe SSDs
The paper introduces py-kvcache, a KV offload connector for vLLM that optimizes external KV caching across GPU, CPU, and NVMe tiers through asynchronous direct I/O, bounded shared staging, and scheduler-aware preloading. Performance characterization reveals that cache efficiency depends on transfer granularity, intermediate memory use, and scheduling, not just device bandwidth. At 80k tokens, py-kvcache achieves 2.0x faster loading than LMCache from disk (1.34x from preloading), and with all tiers enabled, it is 1.23x faster than LMCache and within 4% of native vLLM KV offload. Benchmarks (LongBench, SCBench, Bailian traces) confirm benefits for irregular prefix chains and multi-turn workloads, though GPU memory suffices for short prefixes on high-end GPUs like H100.
kv cachingvllmnvmepreloadingttft
Learning structural balance of graphs from quantum spectral features
The authors propose a quantum method for extracting spectral features from signed graphs by embedding them as Ising models and using standardized moments of the density of states (DOS) as learning features. These moments, which count signed closed walks and are switching-invariant, enable efficient learning of the NP-hard frustration index. The DOS-QPE technique samples spectral density with fewer shots than Hadamard test-based methods, achieving a mean error of 0.4 on 1.4×10^5 labeled graphs. The approach is DQC1-complete beyond zero field, offering applications in social network analysis and spin-glass studies.
quantum feature extractionsigned graphsising modeldensity of statesfrustration index
Reflex-Informed Neuromuscular Reinforcement Learning for Muscle-Driven Locomotion
The authors propose Reflex-Informed Neuromuscular Reinforcement Learning, a framework for muscle-driven locomotion that combines a fixed phase-dependent reflex controller with a reinforcement learning policy. The RL policy modulates four biomechanically meaningful residual parameters to adjust reflex gains and thresholds for hip swing, knee support, and ankle propulsion. Experiments show the method improves kinematic accuracy, dynamic consistency, bilateral symmetry, and stride-to-stride consistency in nominal walking, while maintaining robustness to muscle weakness and external perturbations without retraining.
muscle-driven locomotionreflex controllerreinforcement learningbiomechanical parametersneuromuscular control
Why Does Post-Training Quantization Work?
The study investigates why post-training quantization (PTQ) maintains performance in large language models (LLMs) despite introducing weight errors. By comparing full-precision and quantized forward passes, the authors identify two robustness mechanisms: (1) counteracting residual interactions, where layer-introduced errors oppose inherited errors, slowing hidden-state discrepancy growth; (2) LM-head geometry preserving high-ranked token scores. These mechanisms, validated across models and quantization settings, explain PTQ's stability. Pretraining induces the error-canceling behavior, while LM-head structure mitigates output degradation.
post-training quantizationhidden-state errorcounteracting residual interactionlm-head geometryweight precision
Generalization Analysis of Distributed Kernel-based Robust Gradient Descent Algorithms
The authors analyze the generalization performance of distributed kernel-based robust gradient descent (DKRGD) algorithms in reproducing kernel Hilbert spaces under robust loss functions. By leveraging spectral characterization of gradient descent and intrinsic properties of robust losses, they establish optimal learning rates with a scale parameter σ that mitigates saturation and ensures statistical robustness. A novel error analysis yields sharper operator product bounds, relaxing constraints on the number of local machines while maintaining optimal rates. Additionally, a communication-efficient strategy enhances DKRGD convergence.
distributed gradient descentreproducing kernel hilbert spacerobust loss functionspectral characterizationoperator product bounds
Negative Self-Distillation: Learning to Reason by Avoiding Flaws
Negative Self-Distillation (NSD) improves large language model (LLM) reasoning by diverging from flawed self-generated reasoning traces rather than imitating privileged solutions. NSD avoids the pitfalls of On-Policy Self-Distillation (OPSD), which suppresses uncertainty and penalizes self-correction, by using a dynamic gating mechanism to isolate reasoning-critical tokens during unlearning. This preserves linguistic priors while targeting behavioral flaws. Empirical results show NSD outperforms OPSD and label-free reinforcement learning baselines on complex reasoning tasks.
negative self-distillationon-policy self-distillationlarge language modelsreasoning-critical tokenslinguistic priors
Geospatial Foundation Models Capture Health-Relevant Dimensions of Place Beyond Conventional Social Risk Indices
Geospatial foundation models (GFMs) augment conventional social risk indices by capturing physical place features from 2022 satellite data, explaining residual variance in health outcomes. Using LightGBM on 82,646 U.S. census tracts, the study evaluated four GFM families against American Community Survey variables and 40 CDC PLACES health outcomes. GFMs explained up to 54% of residual variance in health outcomes (e.g., annual checkups, arthritis) beyond social indices, with performance scaling by tract size (R²: 0.31–0.39).
geospatial foundation modelssocial risk indiceslightgbmsatellite dataepidemiological analysis
Multimodal Taxonomic Conditioning for Generative Plankton Imagery
The paper introduces a method for generating synthetic plankton imagery conditioned on taxonomic hierarchies to address long-tailed data distributions in automated plankton imaging. The approach adapts a CLIP encoder using a ranked contrastive objective tailored for deep, irregular taxonomies, then employs it to condition a parameter-efficient diffusion transformer. Synthetic samples are evaluated on distributional fidelity and utility in downstream classifier training, targeting rare taxa with insufficient real images for reliable model development.
contrastive learningdiffusion transformerlong-tailed datataxonomic conditioningplankton imaging
Learnware and AI Model Management System
The paper proposes learnware as a fundamental unit for AI model management systems, comprising a trained model paired with a specification that preserves data privacy. Learnware specifications are generated via machine learning without exposing training data, enabling model reuse and assembly for new tasks. The Learnware Dock System (LDS) facilitates model collaboration through standardized specifications, addressing challenges of interoperability among independently developed models. This approach transforms model pools into managed resources without requiring raw data access.
learnwaremodel management systemspecificationdata preservationcollaboration protocol
Musec: MomentUm SpEctral Clipping for Stable Muon-type Training
We propose MomentUm SpEctral Clipping (Musec), a novel optimizer-level stabilization technique for Muon-type training that replaces spectral flattening with spectral clipping. Musec preserves the momentum matrix's spectral structure by thresholding singular values, addressing training instability without architecture-specific modifications. We introduce Soft Musec, an efficient implementation using smooth spectral saturation approximated via coupled Newton-Schulz iterations. Theoretical analysis establishes Musec's convergence guarantees in nonconvex nonsmooth stochastic optimization, the first such result for Muon-type methods. Empirical evaluations demonstrate Soft Musec's superior stability across learning rates and model sizes, matching existing Muon variants' performance while preventing divergence in unstable configurations.
spectral clippingmuon-type trainingnonconvex optimizationmomentum matrixnewton-schulz iteration
RDDMPI: Residual Denoising Diffusion Model for Probabilistic Multivariate Time Series Imputation
RDDMPI introduces a residual denoising diffusion model for probabilistic multivariate time series imputation (MTSI), reformulating the task as a baseline-residual decomposition. The method employs a pretrained deterministic model to capture dominant signals and a conditional diffusion process to model residual uncertainty, with reliability-aware conditioning to adaptively blend baseline guidance. Experiments show RDDMPI improves reconstruction accuracy and uncertainty quantification across benchmarks compared to direct diffusion approaches in the original data space.
multivariate time series imputationresidual diffusionprobabilistic imputationdenoising diffusionreliability-aware conditioning
Vidu S2: Real-Time Interactive, Editable, and Spatial Video Generation
Vidu S2 introduces two real-time video generation models: Vidu S2-Avatar for interactive digital characters and Vidu S2-Editing for live video manipulation. Vidu S2-Avatar enables 720p video generation with dynamic reference updates and improved instruction following (e.g., dancing), while Vidu S2-Editing supports real-time style rendering, clothing/character/background replacement. Both models explore spatial video generation feasibility. Evaluations demonstrate superiority over baselines, with an online demo available.
real-time generationvideo editingdynamic referencesspatial videointeractive avatar
Identifiability of Nonnegative Tensor Decompositions via Positive Scattering
The paper introduces a positive scattering term to quantify identifiability in nonnegative tensor decompositions, combining it with dimension constraints from the Lovitz--Petrov generalization of Kruskal's theorem. The method leverages positivity-induced geometric rigidity and support constraints to derive two sufficient conditions: a threshold of $2|S|-2$ ensures minimality and nonnegative rank, while $2|S|-1$ guarantees uniqueness among nonnegative decompositions. Key results include a positive splitting inequality for irreducible exchanges and an exact activation characterization via graph connectivity, enabling certification of sparse decompositions beyond Kruskal and Lovitz--Petrov conditions. The criteria generalize to full-rank factorization and two-sided separability in the matrix case.
nonnegative tensor decompositionsidentifiabilitypositive scatteringkruskal's theoremsupport constraints
A distribution-free certification framework for trustworthy crash-severity prediction
The paper introduces a distribution-free certification framework for crash-severity prediction, addressing challenges like ordinal outcomes (KABCO scale), label noise (50% agreement with medical severity), and distribution shift across jurisdictions. The method provides guarantees via contiguous ordinal sets (e.g., 'B or worse'), per-class validity, coverage transfer to true severity, and severity-weighted risk control, all composable with a slack budget. Evaluated on 5.2M Texas records across seven base models, the framework certifies model-independent set widths for vulnerable road users, revealing an unavoidable informativeness floor. The work includes an open-source implementation with theorem-level tests.
distribution-freeordinal regressionlabel noiserisk controlcoverage transfer
A Dataset and Model for Imputing Water Surface Elevation on a Large and Extremely Sparse Spatiotemporal Graph
The authors introduce AmazonSWE, a large-scale spatiotemporal graph dataset for water surface elevation (WSE) imputation, covering 19K+ river sections in the Amazon basin over 10 years (2016–2026) with <1% daily observation density. To address extreme sparsity and directed acyclic topology, they propose a bidirectional selective state space model that samples connected subgraphs and flattens spatiotemporal dimensions into token sequences with topology-aware positional encodings. The model reduces RMSE by 18–39% against in situ gauges compared to prior SWOT-based WSE densification methods, while providing full spatial coverage.
spatiotemporal graphwater surface elevationstate space modelsatellite altimetrypositional encodings
Particle GFlowNets: Rethinking Generative Marginalization Models
The paper establishes equivalence between Generative Marginalization Models (MaMs) and Generative Flow Networks (GFlowNets), unifying two paradigms for discrete stochastic inference. It extends MaMs to non-autoregressive generation by introducing a Gelman-Rubin-statistic-derived criterion for full-state Gibbs sampler rejuvenation, improving learning convergence. Particle GFlowNets, the proposed method, demonstrates significantly faster training in large combinatorial spaces compared to prior approaches.
generative marginalization modelsgenerative flow networksgibbs samplergelman-rubin statisticcombinatorial spaces
Risk-Averse Decision Making with Multi-Level Reliability Guarantees
The paper introduces a method for maximizing weighted average performance certificates across multiple target outage levels under system state uncertainty, connecting to conformal prediction literature. It demonstrates equivalence to optimization over nested prediction sets and derives a dual formulation enabling decoupled optimization across inputs. Numerical experiments on a diversity-based wireless transmission system quantify the cost of multi-level certificate enforcement via a single policy and characterize Pareto trade-offs between reliability levels.
risk-averse optimizationconformal predictionnested setsoutage levelswireless transmission
Generalized Score Matching for Parameter Estimation on Convex Domains
The authors introduce generalized score matching for parameter estimation on convex domains, addressing the intractability of maximum likelihood estimation for unnormalized models. The method constructs an objective function from Minimum Probability Flow learning, demonstrating that classical score matching and domain-adapted variants emerge naturally within this framework. Theoretical analysis shows the objective is a proper second-order local scoring rule, ensuring recovery of the true density at minimization. For exponential family models, convexity and finite-sample estimator consistency are established under standard conditions. Experiments validate the approach on parameter estimation and generative modeling tasks in convex subsets of ℝ^d.
score matchingparameter estimationconvex domainsexponential familyminimum probability flow
Breaking the Central Bias: Spatially Partitioned Experts for Coordinate-Based Neuroevolution
The paper introduces spatially partitioned experts to address the central bias in Evolvable-Substrate HyperNEAT (ES-HyperNEAT), where evolved networks overly concentrate on central input pixels, achieving only 21% mean accuracy on MNIST. By partitioning the input into 13 non-overlapping spatial segments, each handled by a separately evolved specialist network, the method achieves 43% accuracy (106% relative improvement). Receptive-field analysis confirms expanded active pixel coverage from 4% to 79%, demonstrating that partitioning, not data-driven aggregation, drives the gain. The work also provides diagnostic tools for input-coverage collapse and a spatial-partitioning remedy applicable beyond MNIST.
evolvable-substrate hyperneatspatial partitioningmixture-of-expertsreceptive-field analysisneuroevolution
DeFiFlowBench: Benchmarking and Improving Safe Executability in Natural-Language DeFi Workflow Synthesis
DeFiFlowBench introduces a benchmark of 207 natural-language DeFi workflow prompts, evaluating graph coverage, configuration completeness, and safety predicates via EVM testing. The study finds 14-19 unsafe executions per configuration under 5% price-impact caps with direct, constrained, and few-shot prompting. Proposed Koan-Safe combines intent parsing, replaceable generation, and structural repair, achieving 0.67 static safety (vs. 0.33 baseline) and zero unsafe executions in held-out tests. Ablations reveal 14-17 unsafe cases without enforcement, while default injection fails on permissive thresholds. A policy cap mitigates this on 36 diagnostic cases, advocating explicit trade protections and execution-based evaluation.
defi workflow synthesissafety predicatesevm testingintent parsingstructural repair
Combining Synthetic and Real Data for Low-Resource Historical OCR: A Manchu Case Study
This work investigates optimal data mixing strategies for low-resource OCR of historical Manchu documents, comparing synthetic and real training data combinations. Evaluating three vision-language models (VLMs) and a convolutional recurrent neural network (CRNN) on 60k synthetic and 20,306 real word images, the study finds that introducing real data boosts word accuracy from ≤87.92% (synthetic-only) to 95.09-96.28%, with joint and sequential training performing similarly. VLMs benefit substantially from synthetic supplementation, while CRNN shows objective-dependent sensitivity. A compact CRNN matches VLM performance with real data, and ensemble voting achieves 98.27% accuracy. An 18th-century dictionary provides principled disagreement resolution.
optical character recognitionvision-language modelssynthetic data augmentationconvolutional recurrent neural networklow-resource languages
Hologram Representation via Quadratic Phase Gaussian Splatting
The authors introduce Complex-Valued Quadratic Phase Gaussian (CVQPG), a novel hologram representation method that replaces standard 2D Gaussian representations in 2D Gaussian Splatting with 2D quadratic phase functions. CVQPG incorporates learnable parameters to control curvature, enhancing wavefront modulation. Evaluations demonstrate superior visual quality, exceeding state-of-the-art methods by +0.19 dB (RGB) and +0.33 dB (grayscale) in holographic reconstructions. Frequency domain analysis confirms CVQPG's effectiveness in preserving mid-to-high frequency bands of natural images, showcasing its lightweight yet impactful enhancement for hologram representations.
hologram representationquadratic phase functionsgaussian splattingwavefront modulationfrequency domain analysis
Improving the Sensitivity of Gravitational Wave Detection with Weighted Conformal Prediction
The authors propose a weighted conformal prediction framework to improve gravitational wave detection by combining outputs from multiple search pipelines, addressing distribution shifts between simulated training data and real observations. Their method incorporates likelihood-ratio reweighting to maintain coverage guarantees under covariate shift, ensuring statistically rigorous confidence estimates. Experiments on mock datasets demonstrate that the approach restores well-calibrated coverage and increases detection confidence near the threshold, recovering true signals otherwise missed.
gravitational wavesconformal predictioncovariate shiftlikelihood-ratio reweightingdetection threshold
Deep operator learning for efficient sampling from invariant measures of stochastic differential equations
The authors propose an amortized neural sampler combining operator learning with flow methods to efficiently sample from invariant measures of stochastic differential equations (SDEs). Their framework maps SDE coefficient functions to pushforwards from a reference measure, enabling sampling across SDE families with only one encoder pass and few ODE solver steps post-training. The architecture employs Lagrangian trajectory sensors and cross attention for high-dimensional problems, with theoretical guarantees on expressivity and resolution invariance. Experiments on 1D, 2D, and a 64D interacting particle SDE demonstrate competitive accuracy, substantial speedups over MCMC in slow-mixing regimes, and transferability across sensor counts.
amortized neural sampleroperator learningstochastic differential equationslagrangian trajectory sensorsinvariant measures
Local Robustness Quantification for Naive Bayes Classifiers and Generative Forests: a General Approach
The paper introduces methods for quantifying local robustness in two generative classifiers based on Probabilistic Graphical Models (PGMs): naive Bayes classifiers and generative forests. The approach defines prediction robustness as the maximum permissible perturbation of the underlying PGM distribution before the prediction changes, using neighborhoods defined by epsilon-contamination, total variation distance, and chi-squared divergence. Evaluations on benchmark datasets show that robustness values correlate with prediction trustworthiness, outperforming alternative indicators.
probabilistic graphical modelsnaive bayes classifiersgenerative forestslocal robustnessperturbation analysis
Estimating Inconsistency Response Surfaces under Uncertainty in Cyber-Physical System Development
The paper presents a method for estimating inconsistency response surfaces in Cyber-Physical Systems (CPS) under uncertainty by reformulating inconsistency as an intervention-response modeling problem. Using Saltelli sampling and multi-fidelity Monte Carlo estimation, the authors generate datasets to train a surrogate model that predicts inconsistency from propagated uncertainty geometry. Experiments across 48 scenarios and 10 CPS domains demonstrate that the surrogate matches Monte Carlo estimates while reducing evaluation time from milliseconds to microseconds, enabling scalable sensitivity analysis and gradient-based consistency recourse for minimal uncertainty interventions.
cyber-physical systemsuncertainty propagationsurrogate modelingsensitivity analysisconsistency management
A Dynamic Fusion Large Language Model for Traffic Flow Prediction
The Dynamic Fusion Large Language Model (DF-LLM) improves traffic flow prediction by integrating spatiotemporal features and leveraging LLM capabilities. DF-LLM comprises three modules: a spatiotemporal embedding module for multi-scale feature representation, a spatiotemporal fusion module combining spatial topology and dynamic dependencies via graph convolution, and an LLM backbone with differentiated parameter adaptation and context aggregation attention. Residual connections mitigate gradient vanishing. Evaluations across four datasets demonstrate DF-LLM's superior performance metrics compared to baseline methods.
spatiotemporal embeddinggraph convolutionparameter adaptationcontext aggregationresidual connections
A Two-Mirror Faceted Projection System for EUV Lithography
The authors propose a two-mirror faceted projection system for EUV lithography operating at 13.5 nm (Mo/Si) and 11.2 nm (Ru/Be) wavelengths, achieving 4× demagnification with NA ≈ 0.993. Unlike conventional 6–10 mirror systems (<15% throughput), this design uses dedicated planar facet pairs per diffraction order, maintaining 50–60% power retention via strictly two reflections per order. The method employs path-length equalization, 30-bilayer Bragg coatings optimized via transfer matrix methods and evolutionary algorithms, and inverse lithography with differentiable electromagnetic solvers. Simulations demonstrate sub-10-nm resolution, including 5.4 nm FWHM isolated peaks and 6 nm line pairs, robust to 0–5 nm wafer defocus.
euv lithographybragg multilayer coatingsinverse lithographydiffraction ordersoptical path equalization
A Hilbert-Valued Functional Decomposition Framework for Explaining Time-Dependent Outputs
We propose a unified framework for feature-based explanations of time-dependent outputs, addressing the limitation of existing methods that explain each output location independently. Our approach generalizes functional decomposition to Hilbert-valued prediction functions and extends feature-based explanation frameworks using kernel-based output representations. This enables time-dependency-aware explanations at multiple temporal granularities: time-specific, time-resolved, and time-aggregated. The framework subsumes existing methods as special cases. Validation on synthetic and real-world datasets, including intraday financial market volatility prediction and energy demand forecasting, demonstrates its effectiveness.
functional decompositionhilbert-valuedkernel-basedtime-dependencyfeature-based explanations
Rethinking Radiomap Blind Prediction with Limited Environment and Configuration Representations
Proposes RadioDecomp, a method for radiomap blind prediction that decomposes domain risk into target-approximation error and irreducible uncertainty. The approach uses a prior-guided predictor as a correctable base and refines it via deterministic residual learning (instantiated as RadioLSR with LoS-Shadow-Residual components). Experiments demonstrate RadioLSR's effectiveness, particularly in cross-configuration generalization, with consistent gains over monolithic baselines in cross-environment settings.
radiomap blind predictionpropagation priorscross-configuration generalizationresidual refinementdomain risk decomposition
MUtE: A Dual Framework for Concept Erasure and Counterfactual Interventions
The paper introduces MUtE, a dual framework for concept erasure and counterfactual interventions in representation learning. It derives a novel class of erasure functions from optimal bounds of concept erasure, inducing deterministic counterfactual mappings with a translational bias constraint that aligns with geometric concept manifestations in language models. The method bridges theoretical optimality and practical implementation, enabling seamless transitions between erasure and counterfactual generation. Empirical results demonstrate improved algorithmic fairness and counterfactual text generation capabilities.
concept erasurecounterfactual generationrepresentation learningalgorithmic fairnesstranslational bias
Polyhedral Geometry of Time-to-First-Spike Neural Networks
The paper analyzes the expressivity of time-to-first-spike neural networks, demonstrating their superior input space partitioning compared to conventional ReLU networks. By modeling input-output maps as continuous piecewise linear functions with causal feasibility constraints, the authors show each neuron's firing time admits a maxout-like representation with exponentially many constrained affine pieces. They formalize causal regions as polyhedral regions with fixed causal sets, deriving bounds on maximal causal regions in shallow and multilayer architectures. Theoretical and experimental results establish that spiking networks generate richer partitions than feedforward ReLU networks.
spiking neural networkstime-to-first-spikepolyhedral geometrycausal feasibilitymaxout representation
Legible Failures: Detecting and Repairing In-Context Binding Errors
The study introduces a method to distinguish whether language model errors stem from missing information or in-context binding failures, using linear probes on frozen hidden states. On an entity-obligation binding task, probes recovered correct bindings from erroneous model outputs across 16 checkpoints, exceeding chance performance by +0.196 (95% CI [+0.101, +0.296]). Probe disagreement improved failure detection (+0.079 AUROC over model confidence), and steering residuals toward probe-decoded bindings increased accuracy by +0.168 (95% CI [+0.066, +0.280]), demonstrating actionable probe utility in binding tasks.
linear probesin-context bindingentity-obligation taskresidual steeringfailure detection
Diversity of EML-type operators
The paper enumerates and classifies variants of the EML operator, which evaluates explicit purely transcendental elementary functions, addressing a gap in prior work focused solely on binary EML. The authors clarify misconceptions and propose a Möbius layer, replacing matrix operations with rational functions, alongside the activation function eml(x,1/x) to recover exp(x) and ln(x) separately. This enables evaluation of all elementary functions within a rational neural network framework, though symbolic regression in standard neural architectures remains unresolved.
eml operatorsymbolic regressionmöbius layerrational functionselementary functions
REVA: Reusable Evidence View Aggregation for Context-Efficient RAG Serving
The paper introduces REVA (Reusable Evidence View Aggregation), a framework for efficient retrieval-augmented generation (RAG) by mining historical attention traces into reusable document-keyed scores. REVA aggregates token-level attention to word units, preserving document order while enabling budget-specific plain-text views, reducing compression overhead and latency. Evaluated across four benchmarks, REVA improves generation quality by 1.0–5.8 points over existing methods and reduces compression overhead by 5.3–15.6× with sub-40ms latency.
retrieval-augmented generationattention tracesdocument-keyed scorescompression overheadlatency reduction
Convex Optimization with Nested Evolving Feasible Sets (CONES) under Time-Varying Loss Functions
The paper extends Convex Optimization with Nested Evolving Feasible Sets (CONES) to time-varying loss functions, generalizing nested convex body chasing (NCBC). The authors analyze the projected proximal algorithm's performance under convex and strongly convex loss functions. For convex losses, they achieve simultaneous regret $O(T^{1-β})$ and movement cost $O(T^β)$ for any $β∈ [0,1)$, with a lower bound of $Ω(T^{(1-β)/2})$ movement cost for weakly adaptive algorithms. For strongly convex losses, they prove $O(1)$ regret and $O(\log T)$ movement cost, with a matching $Ω(\log T)$ lower bound for sublinear anytime regret algorithms.
convex optimizationonline algorithmsregret boundsmovement costnested feasible sets
Hierarchical Clustering Can Jointly Satisfy Richness, Consistency, and Scale Invariance
The paper demonstrates that hierarchical clustering can jointly satisfy Kleinberg's axioms of scale invariance, richness, and consistency—impossible for flat clustering—by constructing uncountably many admissible methods. These include variants of well-separated clustering and non-binary single linkage, forming a partially ordered set with no greatest element but shared substructures. Results show maximal incompatible methods exist, yet all preserve a backbone of well-separated clusters, constraining their diversity.
hierarchical clusteringkleinberg's axiomsscale invariancewell-separated clusterssingle linkage
Semi-Tensor Product-Based Multi-Term Randomized T-SVD and Its Visual Applications
The paper introduces a multi-term semi-tensor product singular value decomposition (MSTP-SVD) for third-order tensors, addressing limitations of single-term approaches in low-rank approximation accuracy and computational efficiency. The method leverages a novel semi-tensor product under arbitrary invertible linear transforms, relaxing dimensional constraints of the standard t-product while preserving T-SVD properties. Randomized projection and power iteration techniques accelerate the multi-term framework (MRSTP-SVD). Experiments on image and video compression/completion tasks demonstrate improved accuracy-efficiency trade-offs compared to single-term baselines.
tensor singular value decompositionsemi-tensor productlow-rank approximationrandomized projectiont-product framework
When does a spectral prior help graph learning? Connectivity-loss estimation under road-network disruptions
The study proposes a graph neural network (GNN) approach to estimate algebraic-connectivity loss after multi-edge deletions in road networks, combining spectral priors with learned corrections. Using GCN, GraphSAGE, and MPNN architectures, the method evaluates independent, clustered, and targeted failures across 13 OpenStreetMap regions, with residual GCN improving spatial-failure MAE by 0.0391 and residual GraphSAGE improving targeted-failure MAE by 0.0257. Second-order perturbation offers marginal gains (0.0028-0.0053 MAE improvement), while cross-region transfer shows mixed results. The framework scales to 20,000 nodes, decoupling spectral preprocessing from inference costs.
graph neural networksalgebraic connectivityfiedler sensitivityopenstreetmapresidual learning
LILA: Calibration-Free Structured Pruning of Large Language Models via Latent Spectral Geometry
LILA introduces a calibration-free structured pruning method for LLMs using latent spectral geometry, eliminating the need for calibration data, gradients, or auxiliary networks. It scores neuron importance via Kolmogorov-Smirnov distance between singular value distributions of full and ablated FFN weight matrices. At 25% sparsity, LILA outperforms PruneNet by 1.57pp in zero-shot accuracy on LLaMA-2-7B and SliceGPT by up to 6.0pp, maintaining architecture integrity. With one LoRA epoch, it matches SliceGPT within 0.48pp on LLaMA-2-7B and Phi-2. NTK analysis shows 22× lower functional distortion than random pruning, and dynamic KS-score allocation achieves SOTA generative preservation.
structured pruningkolmogorov-smirnov distancesingular value distributionneural tangent kernellora fine-tuning
Bidirectional Multimodal Fusion of Sky Images and Time-Series for Solar Forecasting with Large Language Models
SolCloudLLM introduces a bidirectional multimodal fusion framework for solar forecasting, combining sky-image patches with time-series data via large language models (LLMs). The method aligns image and time-series patches, fuses their representations bidirectionally, and maps them into an LLM's embedding space. Evaluated on SIRTA and SKIPP'D datasets, SolCloudLLM reduces MSE by up to 25.4% versus baselines, with gains concentrated under cloudy conditions, and outperforms in few-shot settings where other deep learning methods degrade.
multimodal fusionsolar forecastinglarge language modelstime-seriesfew-shot learning
Phase-Decoupled, Model-Calibrated Power Control for Disaggregated LLM Serving
The paper introduces a phase-decoupled, model-calibrated power controller for disaggregated LLM serving, addressing inefficiencies in NVIDIA's Max-Q inference profile. The method employs separate power management for prefill (SM-clock window with latency guarantees) and decode lanes (automatically calibrated power cap above throughput/latency cliffs), leveraging flat memory-bound power in decode lanes. Evaluated on 8x B200 nodes serving Qwen3-Coder-480B (FP8) and Qwen3-235B-A22B (NVFP4), the controller achieves +20.4% tokens/J at +3.5% mean e2e latency versus Max-Q's +8.6% at +5.2%, with sustained runs saving 32.3% energy. Results are scoped to MoE models, showing 5x less recovery in dense models.
disaggregated llm servingpower controlmodel calibrationprefill-decode phasesmoe models
How Wrong Can a Good Predictor Be? Diverging Updates with Vanishing Predictive KL
The paper demonstrates that accurate predictive performance does not necessitate precise approximation of Bayesian updates in stationary symmetric Gaussian HMMs. Using an explicit deterministic radial filter and exact Bayesian mixing on $K-1$ belief coordinates, the authors prove an unbounded gap between update maps can coexist with vanishing predictive KL divergence for fixed finite $K\ge2$. Results show worst-case separation grows linearly in confidence scale $L_K(q)$, while categorical KL vanishes, with expected terminal KL converging to zero at horizon $H(q)$. Analysis reveals softmax curvature suppresses disagreement in common confidence cones, controlled by a single Gaussian maximal event.
bayesian updatesgaussian hmmpredictive kl divergencesoftmax curvaturedeterministic radial filter
HERALD: High-Fidelity Exemplar Retrieval with Adaptive Landmark Distillation for Heterophily-Aware Graph Condensation
HERALD introduces a gradient-free graph condensation framework for heterophilic graphs, addressing limitations of existing methods that assume homophily. The method adapts node scoring and feature selection via a joint Fisher-discriminability and activation-density criterion, weighted by a sigmoid function of the heterophily ratio. It constructs condensed subgraphs through score-ordered BFS expansion, Personalised PageRank pruning, and class rebalancing. Evaluated on eight benchmark datasets, HERALD outperforms state-of-the-art methods on heterophilic graphs and remains competitive on homophilic ones across four GNN architectures.
graph condensationheterophilyfisher-discriminabilitylocal intrinsic dimensionalitypersonalised pagerank
Beyond Solver Verdicts: Generative Reward Models for Autoformalization
The paper introduces Generative Verification (GenV), a method to detect Verdict-Preserving-Unfaithfulness (VPU) in neurosymbolic systems, where incorrect formal encodings produce correct solver verdicts. GenV distills a Z3-equivalence oracle into a continuous reference-equivalence score using a language model's vocabulary space, enabling reference-free verification. Mechanistic analysis reveals the model natively localizes spatial errors without explicit training. Evaluations show GenV+HN achieves 0.961 AUROC in reference-equivalence verification, generalizes zero-shot across translators, and improves downstream accuracy by 11.3 points via test-time compute allocation.
neurosymbolic systemsverdict-preserving-unfaithfulnessgenerative verificationreference-equivalencez3-equivalence
TailProp: content-adaptive light- and heavy-tailed propagation for vision
TailProp introduces a hierarchical vision backbone using the Tail Propagation Operator (TPO), which adaptively combines Gaussian and Cauchy stable-process propagators for content-conditioned spatial mixing. TPO predicts channel-wise coefficients to fuse these complementary bases directly in the DCT domain, achieving O(N^1.5) complexity for N=HW feature maps. Evaluated on ImageNet-1K (84.4% Top-1), Mask R-CNN (50.3/44.8 box/mask AP), and ADE20K (50.8% mIoU), TailProp outperforms single-basis and same-family adaptive baselines, demonstrating the efficacy of cross-regime propagation for visual representation learning.
tail propagation operatorstable-process propagatorsdct domaincontent-conditioned mixinghierarchical vision backbone
Coherent Floquet quantum reservoirs for molecular property prediction
The authors propose a discrete time crystal-based quantum reservoir computing (DTC-QRC) architecture for molecular property prediction, leveraging coherent Floquet evolution to process molecular graph inputs and controlled reset for temporal information management. The fixed-dimension feature vectors generated by quantum measurements enable classical decoders to perform inhibitor-activity classification, blood-brain-barrier permeability prediction, and electronic-gap forecasting without tuning reservoir parameters. Evaluations show DTC-QRC surpasses echo-state networks in long-prefix graph classification and ethene gap prediction, with performance degradation under dephasing confirming the role of coherence. Experimental validation on the Quafu superconducting quantum platform demonstrates noise resilience in pair observables.
quantum reservoir computingdiscrete time crystalfloquet evolutionmolecular property predictionquantum cloud platform
When Noise Fabricates Bias: The Fragility of LLM-as-a-Judge Bias Measurement under Noisy Text
This work investigates the impact of text surface noise on social bias measurement using large language models as judges. Applying five realistic noise conditions at varying intensities to 3,822 stereotype-related responses, the study compares bias judgments between noisy and original texts. Results reveal asymmetric degradation: noise is 120x more likely to convert neutral judgments into biased ones than vice versa. Four LLM judges exhibit systematic overestimation of bias in noisy text, with distortion peaking at mild noise levels and attenuating toward parity as robustness increases, particularly affecting fairness-critical categories.
surface noisebias measurementasymmetric degradationstereotype-related responsessystematic overestimation
The information geometry of large language models is shared, learned, and controllable
The paper demonstrates that large language models (LLMs) share a common Fisher-Rao geometry of next-token probabilities across architectures (transformer, state-space, recurrent), with output geometries showing stronger agreement than activation geometries. This shared geometry enables semantic-category transfer and predicts model behaviors, including token probabilities, effective dimension, and human-aligned word choices (improving with scale, training, and calibration). Geometric analysis reveals that pretraining corpus statistics predict fact acquisition, while deeper evidence delays it. The geometry also enables minimum-disturbance interventions, with learned updates transferring to unseen prompts while preserving reference behaviors better than Euclidean methods, benefiting steering, editing, and fine-tuning.
fisher-rao geometrynext-token probabilitiessemantic-category transferminimum-disturbance interventionsmodel calibration
EMMI: Edge Multi-Modal Intelligence for Communication-Efficient MLLM Inference via Fused Representation Compression
EMMI introduces a communication-efficient framework for edge-based multimodal large language model (MLLM) inference by transmitting compressed fused representations instead of raw sensor data or intermediate layer activations. The method employs modality-specific encoding, cross-modal fusion, and learned compression at the edge, producing a fixed-size latent representation for server-side MLLM reasoning. Evaluations on a multimodal benchmark show 32x communication reduction with comparable accuracy, yielding up to 3.4x lower end-to-end latency under bandwidth constraints.
multimodal large language modelsedge intelligencerepresentation fusionlearned compressioncommunication efficiency
Meta-Learning for Classifier Selection in Image Datasets: A Feature-Driven Framework for Accuracy Prediction
A meta-learning framework is proposed for efficient classifier selection in image datasets by predicting performance without exhaustive training. The method extracts meta-features capturing dataset complexity using autoencoders, pre-trained networks, and dimensionality reduction, then trains regression models to estimate classifier accuracies. Clustering techniques group classifiers with similar performance patterns to simplify recommendations. Evaluated on 56 diverse image datasets spanning nature, animals, numbers, motorcycles, medical images, and human bodies, the framework achieves an average ranking prediction accuracy exceeding 86%. This scalable and interpretable approach reduces computational costs while improving classification performance.
meta-learningautoencodersdimensionality reductionregression modelsclustering techniques
Rebalancing Token Importance in Language Models with TF-IDF Weighted Cross-Entropy Loss
The paper introduces an information-weighted cross-entropy loss for language model training, rescaling token contributions via TF-IDF statistics to emphasize semantically informative tokens while down-weighting frequent ones. The method reduces memorization of surface-level text spans without compromising perplexity or downstream performance, as demonstrated on five decoder-only LLMs (1.1B–13B parameters). Under LoRA fine-tuning, it reduces average memorized substring length by 14% (58% for full-weight fine-tuning on TinyLLaMA 1.1B), with <3% computational overhead. The approach is architecture-agnostic and integrates seamlessly into existing pipelines.
tf-idf weightingcross-entropy lossmemorization mitigationdecoder-only llmslora fine-tuning
A variational physics-informed graph neural network for heterogeneous solid mechanics
Proposes a variational physics-informed graph neural network (PI-GNN) for heterogeneous solid mechanics, eliminating interface penalties and transition widths by encoding heterogeneity in the discretization rather than the trial field. The method operates on conforming adaptive meshes, assigns constitutive behavior per element, and minimizes the discrete total potential energy as an unweighted objective with first derivatives only. Achieves von Mises errors below 3.58% across stiffness contrasts (E_inc/E_mat ∈ [10^-2, 10^2]), outperforming strong-form PINNs (5.58% error). Training costs exceed finite element solves by >10×, positioning the approach as a parametric surrogate or inverse identification tool rather than a direct solver replacement.
physics-informed neural networksgraph neural networksheterogeneous solidsfinite element methodvariational formulation
Thompson Sampling for Non-Monotone Convex Ridge Bandits: Monotonicity Is Not Needed for Polynomial Regret
The paper addresses whether monotonicity is necessary for polynomial regret in Thompson sampling (TS) for convex ridge bandits. By constructing a twelve-point configuration, the authors demonstrate that the single-removal John-ellipsoid dichotomy fails for non-monotone links, replacing it with an $O(d^2)$ cardinality bound for uninformative configurations. For arbitrary convex, possibly non-monotone links, exact-posterior TS achieves Bayesian regret $\tilde O(d^{9/2}\sqrt n)$, tight up to constants in the large-diameter-to-gap regime. The $d^{5/2}$ dependence for monotone cases remains open.
thompson samplingconvex ridge banditsbayesian regretjohn-ellipsoidnon-monotone links
Phases in a class of associative memories via hidden neurons
The paper analyzes associative memory in the bipartite Hopfield-Krotov class $H$, using hidden neurons as an order parameter for retrieval. It employs replica methods to derive replica-symmetric phase diagrams and closed-form capacities at polynomial load, showing that differences between Ising and spherical visible neurons arise from visible entropy. With a softmax hidden layer, exponential load maps thermodynamics to random-energy-model counting, revealing paramagnetic, condensed, and frozen phases. Heating disrupts retrieval via quantized attention reassignments, while Gaussian patterns remain metastable. The class $H$ separates stability (visible Lagrangian) and storage scale (hidden Lagrangian), with crosstalk statistics differing between polynomial (central-limit) and exponential (large-deviation) regimes.
associative memoryhopfield networkreplica methodrandom-energy-modellagrangian
When More Is Not Better: Component Anti-Synergy in a P300 Speller
The study challenges the 'more-is-better' assumption in P300 brain-computer interface (BCI) speller pipelines by demonstrating component anti-synergy. A full-factorial experiment evaluated four components—Euclidean Alignment (EA), xDAWN spatial filtering, subject calibration, and language model (LM) priors—on a public P300 dataset using accuracy, repetitions, and information transfer rate (ITR) metrics. Results revealed conditional rather than additive benefits: calibration was the strongest contributor, EA compensated for missing calibration, and LM support's efficacy depended on EEG pipeline strength, with larger LMs showing similar patterns. These findings advocate for context-aware component selection over maximal pipeline activation.
p300 spellereuclidean alignmentxdawn filteringlanguage model priorscomponent anti-synergy
Measuring the Value of World-Model Updates: A Counterfactual Utility Protocol for Continual Adaptation
The paper introduces the fork ledger, a method for evaluating the utility of continual world-model updates by comparing counterfactual trajectories under update vs. hold conditions. It branches deployment streams at decision points into matched continuations, measuring return difference ΔR = R_update - R_hold under common random numbers. Experiments on CartPole, Walker, and Cheetah control tasks show fixed update mechanisms reduce returns (CartPole: -144.0, 95% CI [-185.4,-116.1]; Walker: -82.8, [-101.1,-61.7]; Cheetah: -18.6, [-29.0,-6.6]). The ledger enables causal assessment of update triggers by isolating their impact from surprise detection.
fork ledgercontinual adaptationcounterfactual utilityworld-modelcontrol tasks
Empirical Evaluation of Data Poisoning Attacks in Supervised Learning
This study empirically evaluates data poisoning attacks on supervised learning, comparing label flipping and backdoor poisoning on MNIST and Fashion-MNIST using Logistic Regression, Linear SVM, and Random Forest. Poisoning rates of 5%, 10%, and 20% were tested, measuring clean-test accuracy, macro-precision, macro-recall, macro-F1, and attack success rate. Label flipping degraded performance most in Logistic Regression and Linear SVM, while Random Forest remained stable; backdoor poisoning achieved attack success rates of 0.9667–1.0000 while often preserving clean-test performance, demonstrating its stealth. The results highlight the need for security-oriented evaluation beyond standard metrics.
data poisoninglabel flippingbackdoor poisoningsupervised learningattack success rate
Testing Between the Test Cases: Proving End-to-End Steering in Conditions You Never Drove
The study demonstrates formal verification as a viable complement to simulation for testing end-to-end steering policies in automated vehicles. Using bound propagation, a formal method that analyzes trained weights, the authors evaluated two small networks trained in CARLA on clear and varied conditions (fog, night, low sun). Without additional driving, the method computed lane-departure risks across 133 poses, covering 10^133 disturbance combinations in minutes on one GPU. Formal verification identified failure conditions for the clear-trained policy and suggested potential failures between test cases, supporting its adoption in verification and validation pipelines.
formal verificationbound propagationend-to-end steeringlane-departure budgetsimulation testing
The Platonic brain bridge hypothesis: human brain networks as an architectural prior for omni models
The Platonic brain bridge hypothesis posits that omni models processing multimodal data (video, audio, text) converge on brain-like representations bidirectionally. Seven omni models exhibit stable brain-likeness across participants, with encoding models on their hidden states ranking first on the Algonauts 2025 out-of-distribution leaderboard. Three contributions demonstrate brain-to-model influence: Brain-MoE assigns cortical networks to pretrained experts, improving accuracy by 6.42 percentage points across 15 model-benchmark pairs; Brain-AVQA generates questions from video clips labeled by responsive brain networks, outperforming shuffled maps; Brain-Scope localizes correspondence via sparse autoencoders, showing removal weakens brain prediction. Human brain networks thus serve as a viable architectural prior for omni models.
omni modelsbrain-likenessencoding modelssparse autoencodersarchitectural prior
Empirical Evaluation of Membership Inference Attacks on NLP Text Classifiers: A Baseline Study on SST-2
The paper benchmarks membership inference attack (MIA) vulnerability in text classifiers on SST-2, comparing a TF-IDF + Logistic Regression baseline (0.8756 accuracy) against fine-tuned DistilBERT (0.9466 accuracy). Both models leaked membership signals (AUC 0.5615 and 0.5800, respectively) under loss-threshold MIA. Regularization mitigated leakage for Logistic Regression at a utility cost, while reducing DistilBERT's fine-tuning epochs from 3 to 2 preserved accuracy with lower leakage, demonstrating lightweight privacy-utility trade-offs.
membership inference attacktext classificationdistilbertprivacy-utility trade-offregularization
Structurally Speaking: Motif-Oriented Graph Captioning through Bidirectional Graph-Text Translation
The paper introduces Structurally Speaking, a structured prompting protocol for bidirectional graph-text translation that generates motif-oriented captions. The method guides large language models (GPT-5.1) to abstract connectivity into structural motifs (hubs, paths, cycles) rather than enumerating edges, balancing brevity with recoverability. Experiments on synthetic motif datasets demonstrate that structured prompting yields 23% shorter captions with higher motif consistency while maintaining comparable graph reconstruction accuracy (92% vs. 94% baseline). This suggests explicit topology-to-motif guidance improves interpretability without fine-tuning.
graph captioningmotif abstractionstructured promptingbidirectional translationgraph recovery
ObstaDiff: Generalizable Diffusion Policy Learning via Obstacle-aware Representations
ObstaDiff introduces a decomposed diffusion-policy framework for obstacle-aware robotic manipulation, addressing generalization challenges in cluttered scenes. The method employs a lightweight visual encoder to extract structured target-obstacle-background representations, enabling trajectory generation toward target-centered bottleneck poses with obstacle reasoning. Evaluated on 366 real-robot greenhouse trials, ObstaDiff achieves 75.41% task success and 8.20% collision rate, outperforming imitation-learning baselines in agricultural settings.
diffusion policyobstacle-aware representationimitation learningrobotic manipulationvisual encoder
Symmetry-aware super-resolution of crystal orientation maps via invariant latent-space learning
(No summary returned.)
Relatively Smart II: Tractable or Semi-Supervised Instance-Optimal Learning
The paper extends the study of relatively smart learning, demonstrating that any proper consistent learner achieves distribution-free guarantees for binary classification with quadratic sample-complexity blowup. It establishes that a certifiable error with m samples implies small error on a uniform distribution over O(m²) samples, enabling control via covers of size ≤2^{m+1}. Additionally, semi-supervised learning is shown to achieve quadratic unlabeled sample blowup without labeled sample blowup, using a leave-most-out transductive approach. However, sub-quadratic labeled-sample efficiency requires super-polynomial ERM oracle calls, even with explicit marginals, revealing inherent intractability.
relatively smart learningsample-complexity blowupproper consistent learnertransductive learningagnostic erm oracle
Learning Orthogonal Multi-Index Models Beyond Small Initialization: Incremental Learning, Competitive Dynamics and Symmetry
The paper analyzes incremental learning dynamics in polynomial-width two-layer neural networks trained on orthogonal multi-index targets under standard initialization. Using a symmetry-based finite-width approximation via symmetrized networks, the authors prove that loss decreases sequentially according to the Hermite expansion of the target, with competitive reallocation of parameter mass into the target subspace. Empirical results show similar dynamics for vanilla gradient descent, with lower-order components learned before higher-order ones and mass concentration on aligned neurons.
incremental learningmulti-index modelshermite expansiongradient flowsymmetrized networks
Certifying Lower Bounds for Risk-Sensitive Reinforcement Learning under Adversarial State Perturbations
The paper extends certification methods for reinforcement learning (RL) to risk-sensitive objectives, providing lower bounds on exponential utility of cumulative rewards under $l_{p}$-norm-bounded adversarial state perturbations. By introducing a $φ$-divergence relaxation, the authors formulate the certification problem as convex optimization and derive a tractable dual approximation. Experiments on OpenAI Gym and a machine replacement task show risk-averse training yields policies with higher certified bounds, especially under larger perturbations, though excessive risk aversion leads to overly conservative policies and non-monotonic certification performance.
risk-sensitive reinforcement learningadversarial perturbationsconvex optimizationcertified lower boundsrisk-averse training
Flow Duality and Source Geometry for Categorical Generation
The paper establishes a duality between continuous and discrete flow matching by showing that projecting continuous convex-interpolant paths with one-hot targets through argmax yields discrete convex-interpolant paths. This requires source distributions with specific coordinate symmetry and boundary regularity, making the continuous source distribution an explicit design choice for categorical generation. Experiments with Gaussian, bounded-uniform, and centered negative-exponential sources demonstrate that different source geometries induce distinct transition timing and vocabulary-size dependencies, with preliminary evidence suggesting these effects persist in learned transports and early generative quality.
flow matchingconvex-interpolant pathscategorical generationsource geometryargmax projection
Project Qualia: Recovering Experiential Music Structure from Session Co-occurrence Data
Project Qualia demonstrates that experiential music similarity, distinct from artist identity or genre taxonomies, can be recovered from listening session co-occurrence data. Using 531.6 million scrobbles from 28.6 million Last.fm sessions, the authors train a skip-gram Word2Vec model (Song2Vec) and isolate artist-independent structure by subtracting artist centroids from track embeddings. Residual embeddings reveal 4,577 cross-artist track pairs with cosine similarity ≥0.70, forming coherent clusters (e.g., trip-hop, 1990s grunge) with similarities up to 0.95, validating the existence of an experiential layer in listening behavior.
skip-gramembedding spacecosine similarityartist-residualexperiential similarity
scDEFT: A deep learning framework for drug-effect prediction and counterfactual reasoning
The paper introduces scDEFT, a deep learning framework for drug-effect prediction and counterfactual reasoning using single-cell data. scDEFT models drugs as conditioning operators on cell representations via feature-wise linear modulation, generating drug-conditioned latents trained with per-cell supervision. Two aggregation heads predict drug-induced state changes and responder status, while a backward stage identifies response-driving latent dimensions. Evaluated on a 1.16M-cell IBD atlas across three cohorts and two drug classes, scDEFT achieves 45% baseline-to-ceiling improvement in state-change prediction and AUROC 0.70 for pre-treatment responder stratification, outperforming chance-level baselines. The framework enables target nomination and counterfactual drug-effect analysis.
single-cell rna-seqdrug-effect predictionfeature-wise linear modulationcounterfactual reasoningpatient stratification
Detectable Only Where It Is Confounded: What Verified Duplication Counts Say About Membership Evidence in Language Models
This paper empirically investigates membership inference in language models by leveraging verified duplication counts from pretraining corpora of OLMo-2 and Pythia. Using a controlled experimental design that cancels fluency and quality biases, the study measures the trace of exposure in models ranging from 1B to 13B parameters. Results show minimal correlation (rank correlation ≈ -0.08) between duplication counts and model predictions for ordinary text, with detectable signals only emerging above ~1000 copies due to fame rather than exposure. Additional experiments demonstrate that apparent membership signals are confounded by word choice and register differences, achieving AUC scores up to 0.94. The authors release sentence banks, counts, and code for reproducibility.
membership inferenceduplication countsrank correlationpretraining corporaauc
Processing and classifying bird songs using wavelet techniques and supervised learning
The study introduces a framework for classifying invasive bird species vocalizations in noisy environments using Bayesian wavelet shrinkage with an Epanechnikov kernel prior for efficient signal denoising. Features including Mel-Frequency Cepstral Coefficients (MFCCs), entropy, and zero-crossing rate were extracted from recordings of Euphonia violacea, Leiothrix lutea, and Passer domesticus. Supervised learning models—Random Forest, Multinomial Logistic Regression, and Support Vector Machine (SVM)—were evaluated across varying feature dimensions. The SVM model achieved the highest accuracy of 0.9398 with a 10-dimensional MFCC configuration, demonstrating the efficacy of wavelet-based preprocessing in enhancing classification performance for ecological monitoring.
bayesian wavelet shrinkagemel-frequency cepstral coefficientsepanechnikov kernelsupervised learningsignal denoising
Scale-Aware 3D Deep Learning for Robust Brain Metastasis Detection in Multimodal MRI
The study introduces a scale-aware 3D deep-learning framework for robust brain metastasis detection in multimodal MRI, addressing challenges posed by lesion size variability. The method combines independently trained $96^3$ and $64^3$ 3D U-Nets via weighted late fusion of whole-volume probability maps, isolating the effect of spatial context from resolution and modality. Cross-FOV fusion on a 97-patient cohort improved lesion-level precision (F1) and reduced false positives compared to individual models, with gains attributed to complementary spatial context rather than mere ensemble averaging. An exploratory cross-FOV agreement filter further reduced false positives but did not enhance overall F1.
3d u-netmultimodal mrilate fusionlesion detectionspatial context
How Much Velocity Does Off-Ball Space Value Need? A Broadcast-Viewport Benchmark
The study evaluates the impact of velocity data on off-ball space valuation in soccer under broadcast viewport constraints, comparing four velocity regimes (none, viewport-legal observed, true-for-visible, true-for-all) across three analysis layers: imputation, control surface, and team verdicts. Using a block-bootstrap confidence interval protocol on three Metrica matches (44 m viewport), results show velocity is negligible for imputation (-0.2 pp MAE reduction), modestly beneficial for control surfaces (-1.5 to -1.8 pp, 11--15% of velocity-free MAE), and marginal for team verdicts (-0.12 to -0.19 pp). Visible-player velocity dominates; occluded-player velocity contributes minimally (2--6% of visible gain). Velocity-free error reduction scales with viewport width (7% at 36 m, 21% at 60 m). Noise experiments reveal frame-common variance tolerance (43%) and heavy-tailed residuals. Code and logs are released.
velocity-awarebroadcast viewportoff-ball analysisimputation errorcontrol surface
RiVaT-Fuse: Reliability-Calibrated Variational Tensor Fusion for Multimodal Prediction under Modality Uncertainty
RiVaT-Fuse introduces reliability-calibrated variational tensor fusion for multimodal prediction under modality uncertainty, defining fusion as sample-wise latent-state estimation via a variational objective balancing image evidence, metadata, cross-modal interaction, and stability. The framework replaces scalar confidence with matrix-valued trust geometry, decomposes interaction into additive, multiplicative, and relational components, and couples latent states with conditional robustness. Instantiated with low-rank-plus-diagonal trust operators, RiVaT-Fuse achieves top predictive rank among representation-level baselines and improves stability under perturbation on an image-metadata prediction benchmark.
variational tensor fusionmodality uncertaintymatrix-valued trustlatent-state estimationcross-modal interaction
DR-LabStack: Design and Implementation of a Clinician-Facing Web System for Diabetic Retinopathy Prediction
DR-LabStack presents a clinician-facing web system integrating four heterogeneous diabetic retinopathy prediction models (RuleFit, Pruned RuleFit, Elaborative XGBoost, Two-level Ensemble) through standardized interfaces. The React-Flask architecture implements model-specific input field rendering (14-25 features per model), positional vector construction, and unified JSON responses with binary classification outputs. Functional testing confirmed correct model loading (4/4 models), consistent API behavior (62 Flask requests), and UI rendering accuracy (24 browser scenarios), demonstrating interoperability across differing preprocessing and output requirements without modifying original model artifacts.
diabetic retinopathymodel integrationreact-flaskheterogeneous modelsclinical interface
Composable CXL Memory as a Kubernetes-Native Shared Memory for LLM Serving
The authors present a Kubernetes Dynamic Resource Allocation (DRA) driver enabling composable Compute Express Link (CXL) memory as a schedulable cluster resource for LLM serving, facilitating cross-node KV-cache reuse. The driver dynamically composes CXL regions, exposes them as DAX devices, and injects them into pods via Container Device Interface (CDI), with a shared-memory connector for vLLM/llm-d embedding slot directories directly in the shared medium. Evaluation on a two-node cluster with 512GiB CXL and Qwen2.5-7B-Instruct shows 5.5×–36.6× reduction in time-to-first-token (TTFT) at 95.4–99.5% external hit rates, with a 1–4% sharing gap between cross-node and same-node reuse latencies.
cxl memorykubernetes drakv-cache reusedax devicesllm serving
From Connectivity to Rewards: Dense Reward Learning with Directed State Graphs
Proposes Graph-Guided Quasimetric Dense Reward (G2QDR), a framework integrating state connectivity models into Goal-Conditioned Hierarchical Reinforcement Learning (GCHRL) for asymmetric environments. G2QDR predicts pairwise state connectivity strengths via neural networks trained on directed state graphs, converting them into dense auxiliary rewards to guide hierarchical policy learning. Empirical results demonstrate improved performance over baseline GCHRL methods in sparse-reward environments with acceptable computational overhead, leveraging environmental topology more effectively than prior graph-based sampling approaches.
goal-conditioned hierarchical rlquasimetric environmentsstate connectivity modeldense reward learningdirected state graphs
A Bellman Optimality Equation for Plasticity
The paper introduces a Bellman optimality equation for optimizing plasticity in continual reinforcement learning, addressing a gap where prior work focused on empowerment optimization. Building on Abel et al.'s (2025) formulation of plasticity as generalized directed information from observations to actions, the authors derive a dynamic programming framework analogous to empowerment-based approaches. Preliminary results demonstrate the existence of this plasticity-specific Bellman equation within Markov decision processes, aligning with established empowerment optimization methods.
continual reinforcement learningplasticity optimizationbellman optimality equationgeneralized directed informationmarkov decision processes
Weighted Empirical Risk Minimization for Machine Learning under Long-Range Dependence: Exact Pathwise Rates and Learning-Error Geometry
The paper develops an exact almost-sure learning theory for smooth parametric models trained via weighted empirical risk minimization on long-range dependent data. Observations are drawn from a stationary Gaussian sequence with regularly varying sample weights. Key results include an almost-sure Bahadur representation, exact limsup laws for learned parameters, and functional cluster sets for learning trajectories, with polynomial learning exponents determined by memory parameters and Wiener-chaos rank. The analysis reveals invariance of exponents under power weighting and dependence of pathwise constants on weights, with global optimization showing positive optimizers in rank-one cases.
weighted empirical risk minimizationlong-range dependencewiener-chaos rankbahadur representationfunctional cluster set
Meta-Learning for Data-Efficient Plant Growth Estimation via Vision Transformers and Fuzzy Clustering
The authors propose a few-shot regression framework for plant growth estimation that combines Vision Transformer (ViT) embeddings, fuzzy c-means clustering, and gradient-based meta-learning. The method leverages unlabeled images to construct structured tasks via clustering in embedding space, enabling efficient learning from limited labeled samples. Systematic evaluation shows that second-order meta-learning methods, particularly MAML++ variants, outperform classical baselines in few-shot scenarios. Experiments on two plant datasets demonstrate that structured task design combined with meta-learning achieves reliable growth estimation under severe label scarcity, with task construction in embedding space identified as a primary performance driver.
few-shot regressionvision transformerfuzzy c-meansmeta-learningembedding space
SynCo: Synthetic Community-Aware Attributed Graph Generator for Graph Neural Network Benchmarking
SynCo introduces a synthetic attributed graph generator enabling control over node degree distributions and sub-community structures, addressing limitations in existing generators that rely on inflexible power-law assumptions. The method supports diverse community densities and structures, evaluated across graph mimicking, hyperparameter tuning, and node clustering tasks. Results demonstrate SynCo's superiority over state-of-the-art approaches in synthetic generation and data augmentation, preserving original distributions (validated statistically) and scaling to 2.1M nodes.
graph neural networkssynthetic graph generationcommunity detectionnode degree distributiondata augmentation
Conformal Calibration Transfer
The paper introduces Transported Conformal Calibration (TCC), a method for transferring conformal prediction sets from a labeled source space to an unlabeled target space using paired observations. TCC employs two correction methods: TCC-KS, which detects mismatch via label-free uncertainty surrogates, and weighted-TCC, which reweights transported calibration for efficiency. The approach guarantees finite-sample target-domain coverage, adapting to observable mismatch. Experiments on CIFAR-100-C, Tiny-ImageNet-C, and SEN12MS demonstrate reliable coverage transfer without labeled target data, validated by label-free diagnostics.
conformal predictioncalibration transferdomain adaptationuncertainty quantificationcoverage guarantees
A Quantum-Inspired Dequantization Method for Diagonally Weighted Matrix Functions: Application to Learning with Optimized Random Features
The work introduces a classical dequantization method for quantum-inspired algorithms handling diagonally weighted matrix functions, specifically targeting cases where quantum singular value transformation (QSVT) samplers lack direct sampling access. By sampling heavy indices and reducing the transformation to a small principal block, the method constructs a sparse classical representation with operator-norm guarantees. This approach dequantizes the sampler for optimized random features, achieving prescribed accuracy with polynomially related runtime, demonstrating that quantum block encoding structures can enable classical access even without composite matrix sampling.
dequantizationquantum singular value transformationrandom featuresoperator-normblock encoding
HuRo: Robotizing Human Videos for Scalable VLA Pretraining
The paper introduces HuRo, a pipeline for converting diverse human videos into robot-aligned observations and action trajectories to pretrain vision-language-action (VLA) policies. The method infers missing intermediate signals across annotation levels, creating the HuRo dataset (630K episodes, 142M frames from 5 sources). Scaling robotized pretraining improves task completion from 51.5% to 80.3% and OOD robustness (34.9% to 72.2%), with ablations showing visual robotization enhances OOD performance and end-to-end pretraining surpasses visual-only transfer.
robotization pipelinevision-language-action policiesood robustnesspretraining scalehuman-to-robot alignment
Characterizing Language Generation in the Limit: Finite Witnesses and a Separation-Width Hierarchy
The work characterizes language generation in the limit for arbitrary families over a countable universe, showing generation is possible precisely when each target has a finite positive witness ensuring infinite common intersections among activated targets. A universal normalization proves necessity, converting successful generators into set-dependent ones. The separation-width hierarchy is introduced, with levels for uniform size bounds, unbounded finite witnesses, and no finite-witness assignment, all realized in practice. Countable families admit singleton witnesses, while explicit families achieve arbitrary finite widths. Lean formalization validates the characterization and hierarchy, including normalization and diagonal capture.
language generationfinite witnessesseparation-width hierarchycountable universelean formalization
GEOSTEER: Geodesic Optimization for Activation Steering in Large Language Models
GeoSteer introduces a Riemannian optimization method for norm-preserving activation steering in large language models (LLMs), addressing limitations of predefined trajectories and one-step updates. The approach formulates steering as a geodesic optimization problem on the representation manifold, employing a learned nonlinear activation-space objective to adaptively guide multistep updates. This geometry-aware method ensures smoother, more stable steering while preserving activation norms. Evaluations on TruthfulQA, RealToxicityPrompts, and UltraFeedback benchmarks demonstrate consistent improvements over state-of-the-art activation steering baselines, highlighting the efficacy of adaptive, multistep optimization.
activation steeringriemannian optimizationrepresentation manifoldnorm-preservinggeodesic steps
Artificial Intelligence Algorithms for the Detection of Pathologies Related to Lung Cancer through Image Analysis using Convolutional Neural Networks and Data Augmentation: a systematic mapping of the literature
This systematic literature review analyzes 96 studies (2015-present) on AI applications for lung cancer detection via medical imaging. The survey highlights convolutional neural networks (CNNs) with transfer learning and data augmentation as predominant techniques, demonstrating high sensitivity and specificity in early diagnosis. Key challenges include data standardization, model explainability, privacy concerns, and ethical implications. While results indicate AI's potential to enhance radiological workflows, the authors emphasize the need for further research and regulatory frameworks to ensure clinical reliability.
convolutional neural networkstransfer learningdata augmentationmedical imagingexplainability
Byzantine-Robust Federated Fire Detection with a Rotating Coordinator
The paper introduces a Byzantine-robust federated learning (FL) framework for indoor fire detection, addressing bandwidth constraints, malicious clients, and fixed-server vulnerabilities. The method combines history-aware aggregation with a rotating coordinator, compressing model updates 10× with minimal accuracy loss. A curated dataset from eight public sources supports training, while a six-node cloud deployment validates feasibility. The rotating-coordinator approach matches fixed-server accuracy (tested on held-out data) and evades stealthy attacks missed by per-round filters.
federated learningbyzantine robustnessmodel compressionrotating coordinatoredge computing
Sequence-Informed Geometric Evaluation of RNA 3D Structures
The paper introduces SIRGE, a sequence-informed geometric evaluator for RNA 3D structures that conditions structural representations on nucleotide embeddings from a pretrained RNA language model. SIRGE outperforms established evaluators in Kendall-τ alignment, Top-1 selection, and Top-3 ranking tasks. Controlled experiments demonstrate that sequence conditioning corrects errors made by a purely geometric model and improves target-level rank structure, indicating that pretrained sequence representations complement geometric reasoning in RNA structure evaluation.
rna structure evaluationsequence-informed geometric evaluatornucleotide embeddingskendall-τ alignmentpretrained language model
Zero-shot rib design: merging training-free generative prior with topology optimization
The paper introduces a zero-shot rib design framework that integrates a frozen text-to-image diffusion model as a training-free generative prior with density-based topology optimization. Using score distillation sampling, the method combines prompt-induced generative gradients with finite element sensitivities, enabling text prompts to guide structural design while physics determines feature viability. Across 245 runs spanning four geometric domains and two physics regimes, 38 of 49 prompt-domain pairs achieved significant compliance reductions (up to −31.5% mechanical, −23.0% thermoelastic), with generative priors suppressing dead-end rib branches (endpoint-compliance correlation r = +0.56 to +0.99). A Heaviside projection resolved intermediate-density artifacts (42.6% to <3%), and a skeleton-based pipeline converted densities to CAD-ready geometry.
topology optimizationdiffusion priorscore distillation samplingfinite element analysiszero-shot design
📰 Industry Media (8)
Roundtables: Will AI really kill us all?
MIT Technology Review convenes experts to evaluate existential risks from advanced AI systems, analyzing the validity of claims by leading lab employees about potential human extinction. The roundtable examines the origins of these concerns, their empirical basis, and proposed mitigation strategies. Participants include senior editors Niall Firth and Will Douglas Heaven alongside AI reporter Grace Huckins, addressing technical mechanisms like recursive self-improvement and goal misalignment that may contribute to catastrophic outcomes.
existential riskrecursive self-improvementgoal misalignmentmitigation strategiesempirical basis
Anthropic Adds Plugin Evals to Claude Code: 6 Grader Types, a No-Plugin Baseline, and a CI Gate for Skills
Anthropic introduced a plugin evaluation workflow for Claude Code (v2.1.269+) enabling quantitative assessment of plugin efficacy. The method employs six grader types (regex, tool_used, tool_order, file_exists, llm, baseline) to compare plugin-enabled (WITH) and plugin-disabled (W/OUT) runs, computing Δ as the performance difference. Key findings include identifying plugins that fail to trigger on natural language inputs (Δ≈0 with tool_used failures) and validating plugin contributions (Δ>0). The system supports CI integration via --threshold and --max-cost-usd flags, with costs scaling linearly with cases×runs×arms. Results show per-grader verdicts and judge votes in evals/results/report.html.
plugin evaluationgrader typesdelta metricci integrationnatural language triggering
Cohere Releases North Small Translate: A 218B MoE Translation Model That Scores 83.6 on WMT26 Across 50 Languages
Cohere introduces North Small Translate, a 218B parameter sparse Mixture-of-Experts (MoE) translation model with 25B active parameters per token, achieving 83.6 average score on WMT26 across 50 languages. The decoder-only Transformer employs 128 experts (8 activated per token), sliding-window attention (window size 4096), and 16K context length. Vendor-reported benchmarks show it outperforms DeepL NextGen (81.37) and Google Translate (68.20), with an agentic variant reaching 84.36. The model demonstrates 1.4x higher throughput than Gemma 4 31B and supports cost-efficient deployment via 4-bit quantization on 1x B200 or 2x H100 GPUs.
mixture-of-expertswmt26sliding-window attention16k context4-bit quantization
Sakana AI Launches Fugu Max and Fugu Ultra v2 for Cheaper, Stronger Multi-Agent Orchestration
Sakana AI introduces Fugu Max and Fugu Ultra v2, two multi-agent orchestration models optimized for cost-efficiency and high-capacity reasoning, respectively. Fugu Max employs a broad pool of open-weights and specialized models (including NVIDIA Nemotron) to route tasks to the most cost-effective model, achieving 40-60% lower output costs than Sonnet 5 and GPT-5.6 Terra while leading on 6 benchmarks (Terminal Bench 2.1, GPQA Diamond, AA-LCR, GDP.pdf, AutomationBench, SWEFish). Fugu Ultra v2 specializes in complex reasoning, scoring 48.3 on Chartography (visual reasoning) and 74.3 on DeepSWE (software engineering), outperforming higher-priced models without relying on proprietary APIs like Fable 5 or GPT-6-Astra. Both models use a learned orchestrator trained via evolutionary algorithms and RL, deployed via OpenAI-compatible API.
multi-agent orchestrationpareto frontierevolutionary algorithmsopenai-compatible apinvidia nemotron
Google Research Releases ToolGrad: Answer-First Framework Hits 99.8% Pass Rate for Tool-Use Data Generation
Google Research introduces ToolGrad, an answer-first framework for generating tool-use training data by constructing verified API chains before synthesizing matching user queries. The method iteratively proposes, executes, selects, and updates API workflows via four modules (API Proposer, Executor, Selector, LLM Updater), achieving 99.8% pass rates on ToolBench (vs. 63.8% for query-first DFS) with 3.4 ground-truth tool uses per sample. Fine-tuning Gemma-3 models (1B/4B/12B) on 500 ToolGrad samples yields 83.1 BFCL score (comparable to Gemini 2.5 Pro) and outperforms the teacher model (Gemini 2.5 Flash-Lite).
tool-use generationapi workflowberkeley function calling leaderboardin-context learningparameter-efficient fine-tuning
Meet Redis LangCache: A Managed Semantic Cache That Cuts LLM API Costs by Up to 90% and Returns Cache Hits Up to 15x Faster
Redis LangCache introduces a managed semantic caching service that reduces LLM API costs by up to 90% and accelerates cache-hit responses by up to 15x. The system employs vector search over stored prompt-response pairs, using embeddings to match semantically similar queries, thereby avoiding redundant LLM calls. Evaluations demonstrate 6x faster response times and 70% cost savings in production deployments, with configurable similarity thresholds, TTLs, and eviction policies ensuring accuracy. The REST API integrates with any LLM provider, prioritizing output token savings and monitoring hit rates via Redis Cloud.
semantic cachingvector searchllm api optimizationkv-cacheembedding generation
NVIDIA Details BioNeMo Inference Runtime (BioIR): 2.90x Higher Boltz-2 Folding Throughput and 58.5K Residues per GPU-Hour on 8xH100
NVIDIA introduces BioNeMo Inference Runtime (BioIR), a Python library optimizing protein structure prediction models (Boltz-2, OpenFold2/3) on NVIDIA GPUs while maintaining PyTorch compatibility. BioIR accelerates inference via kernel selection (custom CUDA kernels), module optimization (CUDA Graph capture), and pipeline scaling (Ray-based multi-GPU replication). Benchmarks on 8xH100 GPUs demonstrate 2.90x higher throughput (58.5K vs. 20.2K residues/GPU-hour) for Boltz-2 on 1,000 human dimers, reducing estimated energy use from 35 MWh to 11 MWh per million targets. The runtime has been deployed at scale, generating 31M candidate complexes for the AlphaFold Database expansion.
bionemoprotein foldingcuda graphray replicationtorch.nn.module
Palantir Foundry and cuOpt drive NVIDIA supply chain allocation
NVIDIA automates global hardware supply chain allocation using Palantir Foundry and cuOpt, addressing complex dependencies in manufacturing Grace Blackwell NVL72 racks and Vera Rubin architectures. Foundry models facilities and supplier commitments as interconnected objects, while cuOpt formulates distribution as a mixed-integer linear program to minimize Time of Ownership (TOO). NVIDIA post-trained Nemotron 3.5 Lightning, a 30B parameter mixture-of-experts model, on qualitative operational records, achieving 86.7% decision accuracy and outperforming larger models in balanced accuracy (58.6%) and macro-F1 (57.5%). Fine-tuning completed in minutes on two NVIDIA B200 GPUs, with future reinforcement learning planned using planner revisions and factory outputs.
mixed-integer linear programmingtime of ownershipmixture-of-expertslow-rank adaptationsupply chain allocation
Generated automatically at 2026-09-11 22:12 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.