Daily Digest — 2026-08-15
279 items · 1 research labs, 273 arxiv papers, 5 industry media
🏛️ Research Labs (1)
State of Open Models: Summer 2026 Observations
The Hugging Face Hub's 2026 data reveals a shifting landscape in open models, with Chinese labs dominating frontier-scale releases (754B-2.78T parameters) while U.S. contributions focus on hardware-optimized conversions. Analysis of 2.96M model repositories shows extreme concentration: 1.5% account for 99.2% of downloads, with small models (<1B) comprising 83% of usage. Qwen emerges as the ecosystem's base model, generating 151K derivatives (4.7× Llama's footprint) due to consistent Apache 2.0 releases across scales. Local inference transforms via llama.cpp (464% growth in GGUF repositories), enabling trillion-parameter models on consumer hardware.
parameter scalinggguf quantizationecosystem derivativeslocal inferencelicense distribution
📜 arXiv Papers (273)
AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design
AutoDesign introduces a meta-harness optimization framework for long-horizon agentic design, where a code agent recursively improves harness systems via rollout feedback, aligning with human design priors. The method is evaluated on PosterBench, a benchmark for academic paper-to-poster generation (100-paper Main Track, 10-paper subset). AutoDesign scores 78.32 on PosterBench, outperforming Claude Design by 7.45 points, and improves average performance by +12.4% across seven code-agent configurations. In autonomous operation, it executes 253 tool calls and 11 edits in 40 minutes (<$3), achieving conference-poster quality in human evaluation.
meta-harness optimizationlong-horizon agentic designcode agentrollout feedbackpaper-to-poster generation
OmniScientist: An Omni-Modal Omni-Discipline AI Scientist
OmniScientist introduces an omni-modal AI scientist capable of end-to-end multidisciplinary research using heterogeneous raw evidence, addressing limitations of existing systems that rely on preprocessed inputs. The system combines a perception layer with three autonomous agents (ideation, experiment, writeup) in a deterministic pipeline, enforcing novelty screening, statistical validity, and execution provenance through code-based checks. Evaluated on 36 real-data cases spanning 5 discipline families and 4 evidence types (images, signals, 3-D structures, etc.), OmniScientist completes full research workflows from raw data to compiled manuscripts, achieving a mean paper score of 6.3. Direct perception outperforms scalar-feature baselines in 85% of head-to-head comparisons, demonstrating the necessity of lifecycle-wide perception for evidence-grounded discovery.
omni-modalautonomous agentsdeterministic pipelineexecution provenancelifecycle-wide perception
HumanTracker: Towards Comprehensive and Human-Aligned Motion Tracking Benchmark
The paper introduces HumanTracker, a benchmark for humanoid motion tracking that addresses perceptual misalignment and scalability in current evaluations. The benchmark comprises 153 hours of optical motion trajectories from professional performers, categorized into four motion families with text labels for detailed analysis. Additionally, the authors propose HumanScore, a preference-aligned metric trained on 12K motion pairs (24K motions), which outperforms kinematic metrics in predicting human preferences and identifying contact and stability failures. HumanScore reveals shortcomings in state-of-the-art trackers that traditional metrics often overlook.
humanoid motion trackingoptical motion trajectoriespreference-aligned metrickinematic errorsfoot skating
QuoteBench: How Matched Scores Can Hide Command-Path Failures
QuoteBench introduces exact final-state validation to quantify how command-generation errors differ from post-generation failures in LLM coding agents, using 56 one-shot tasks from 14 incident-derived families. The benchmark crosses generation contracts with execution transport via a deliberately unescaped parser, revealing that matched execution scores obscure substantial performance gaps: replaying replies through the added parser reduces success by 55.4–73.2 percentage points, while boundary disclosure recovers 30.4–60.7 points in six configurations. Results show GPT-5.6-sol's matched gap of -3.6 points masks -64.3 points of damage and +60.7 compensation, with deployment configurations reordering model rankings. The work advocates reporting model configuration, generation contract, execution path, and final-state validation instead of relying on matched scores.
llm coding agentsfinal-state validationgeneration contractexecution transportmatched scores
LittleLearner: Language Models Under Pedagogically Controlled Knowledge Exposure
The authors introduce LITTLECURRICULUM, an 88B-token pretraining corpus restricted to U.S. elementary school (Grade 5 and below) content, enabling controlled study of knowledge acquisition in language models. They train LITTLELEARNER, a 5B-parameter LLM, on this corpus, creating a developmentally bounded sandbox with interpretable capability limits. Experiments demonstrate that post-training and in-context learning improve knowledge utilization without expanding out-of-scope abilities, validating the utility of this controlled environment for mechanistic research.
pretraining corpusdevelopmental boundariesin-context learningknowledge acquisitioncontrolled environment
Vero: Can AI Agents Build Formally Verified Software Repositories?
The paper introduces Vero, the first benchmark for evaluating AI agents on joint implementation and proof synthesis at the repository level. Vero comprises 43 multi-module instances from real-world repositories (Python, Dafny, Verus, Coq) with manually curated formal specifications and reference implementations. It supports proof-only and code-and-proof evaluation modes, including an audit mechanism for verifying specification unsatisfiability or reference code incorrectness. Evaluation of frontier coding agents with Lean toolchain access shows the strongest agent fully solves only 27 of 43 instances, failing on the hardest repositories. Vero provides a testbed for repository-scale verified software synthesis, where current agents underperform.
verified code generationrepository-level synthesisformal specificationslean 4multi-module evaluation
The data geometry of masking diffusion: Certified-optimal schedules via unmasking growth complexity
The paper introduces unmasking growth complexity (UGC), a path-resolved measure of data geometry for discrete sampling via masking diffusion. UGC's local increments control KL discretization error, enabling unified analysis of Bernoulli-subset and fixed-cardinality unmasking schemes. The method derives certified-optimal single-block and multi-block schedules in log-reveal-odds coordinates, with iteration complexity within a constant factor of oracle procedures. Results show dimension-dependent gains, including Ω̃(√d) improvements over coarse schedules, and connect UGC to classical dependence measures. The squared integral of √UGC density determines leading-order Euler discretization error in fine-partition limits.
masking diffusionunmasking growth complexitykl discretizationcertified-optimal scheduleseuler discretization
DFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data
The authors introduce Mimir v1, a 1-billion-parameter Hierarchical Reasoning Model (HRM) trained exclusively on permissible post-training data, addressing ethical concerns in LLM development. The model, trained on 161 datasets, achieves competitive performance in English and state-of-the-art results in Danish, outperforming HRM-Text 1B and rivaling larger models like Qwen 3.5 4B and Gemma 4 E2B across 20 benchmarks spanning English, Math & Code, and Danish. The model is openly available on Hugging Face Hub.
hierarchical reasoning modelpermissible datalanguage modelbenchmark performanceopen-source
AlayaWorld: Interactive Long-Horizon World Modeling - Full Technical Report (v1.1)
AlayaWorld v1.1 introduces architectural improvements to its interactive long-horizon world modeling framework, focusing on conditioning signal alignment with generated content in latent space and temporal structure. Key modifications include replacing depth-warping spatial memory with a streaming 3D point-cache renderer and redesigning the conditioning pipeline for consistent visual encoding in a causal-VAE latent space. The update implements six technical changes: motion-aware latent conditioning, causal encoding of re-rendered spatial memory, temporal-memory window alignment, hard memory dropout, unified VAE encoding/decoding, and removal of the camera AdaLN branch for viewpoint control.
causal-vae3d point-cachelatent conditioningmemory dropoutadaln
Toward a Gricean Retreat: Probing LLMs for Knowledge Boundaries and Referent Specificity
This work identifies a failure mode in large language models (LLMs) where they fabricate specific details rather than retreating to safer, generic claims when encountering unknown entities. Framing this through Gricean principles, the authors probe whether LLMs encode knowledge-boundary awareness and referent-specificity anticipation using a T-REx-based benchmark. Results show that while LLMs encode both signals, they fail to reconcile them during generation, preferring specific referents even for unknown entities. The study positions these findings as foundational for Gricean alignment objectives that couple knowledge-boundary awareness with referent-specificity in generation.
gricean principlesknowledge boundaryreferent specificityt-rex benchmarkgeneration policy
Synthetic Persona Pretraining: Alignment from Token Zero
The paper introduces Synthetic Persona Pretraining (SPP), a novel paradigm for aligning language models with human values from the outset of pretraining. SPP involves annotating pretraining documents with value-aligned first-person reflections derived from a normative value constitution, pretraining using cross-entropy loss on both original documents and reflections, and post-training on user-assistant dialogue data for persona binding. Experiments with models up to 3B parameters trained on 500B tokens demonstrate that SPP improves constitution following, jailbreak robustness, and reduces misalignment in out-of-distribution moral dilemmas while preserving capabilities. Early intervention proves crucial, with SPP outperforming late-stage alignment in value adherence and dilemma resolution, particularly with increased pretraining budget.
synthetic persona pretrainingpersona bindingcross-entropy lossjailbreak robustnessvalue constitution
MARC v1: An Open-Source Multi-Agent Framework for Clinical AI Reasoning and Coordination
The authors introduce Multi-Agent Reasoning and Coordination (MARC) v1, an open-source framework that replaces monolithic LLM prompting with deterministic multi-agent orchestration for clinical reasoning. MARC employs role-specialized agents for extraction, reasoning, answer generation, and evaluation, featuring explicit context passing, traceable intermediate outputs, and a Decomposer module for automatic prompt generation from plain-language descriptions. The framework supports API-based and local CPU-compatible deployments, is configurable via YAML, and designed for model-agnostic, interpretable clinical AI applications accessible to non-programmers.
multi-agent systemsclinical reasoningprompt engineeringdeterministic orchestrationyaml configuration
AaLLM: An End-to-End Analog Circuit Design Framework from Topology Generation to Sizing Using Large Language Models
AaLLM introduces an end-to-end multi-agent LLM framework for analog circuit design, automating both topology generation and sizing via a tri-agent feedback system (Designer, Critic, Evaluator) and RAG-based knowledge extraction. The method reduces manual intervention by creating a technical knowledge base from research papers, addressing hallucinations and iterative sizing trade-offs. Results demonstrate novel topologies achieving comparable or superior FoM (up to 3×), 3×–4.5× fewer SPICE calls versus SOTA multi-agent LLM pipelines, and 40× faster wall-clock time than existing approaches.
analog circuit designmulti-agent llmtopology generationretrieval-augmented generationfigure of merit
Concept Drift Detection and Adaptive Retraining of Malware Classification Models
The study introduces a novel One-Class Support Vector Machine (OCSVM) approach for concept drift detection in malware classification models, comparing it against Minibatch K-Means (MK-Means) and Maximum Mean Discrepancy (MMD). Experiments evaluate four machine learning models—Multilayer Perceptron, Random Forest, Support Vector Machines, and eXtreme Gradient Boosting—across static, periodic, and drift-aware retraining scenarios. Drift-aware retraining achieves classification accuracy comparable to periodic retraining while significantly reducing retraining frequency. OCSVM-based drift detection outperforms MK-Means and MMD in efficiency and accuracy, demonstrating effective concept drift mitigation in malware classification tasks.
concept driftmalware classificationone-class svmminibatch k-meansmaximum mean discrepancy
MLLM-Routed Heterogeneous Ensembles for Robust Cross-Dataset Image Classification
The paper introduces ARMDIL, an Adaptive Router for Multi-Domain Image classification with LLMs, which employs a multimodal large language model (MLLM) to dynamically route images to the most suitable vision backbone in a heterogeneous ensemble. The ensemble combines ResNets, self-supervised representation learners (SSL), and vision-language models (VLMs), trained on a unified label space from multiple datasets. ARMDIL achieves competitive performance with specialized training-based routers, enhances adaptability via prompt modifications, and improves interpretability through natural language reasoning. Evaluations demonstrate its robustness across disparate visual domains.
multimodal large language modelheterogeneous ensemblecross-dataset classificationadaptive routingvision-language models
CAPRI: Contract-Aware Proof Repair for Isabelle
CAPRI introduces a contract-aware proof repair workflow for Isabelle, ensuring LLM-generated edits comply with developer-authorized changes while maintaining proof validity. The method combines Isabelle's proof checking with an independent contract enforcer, retaining audit trails of prompts, proposals, and hashes. Evaluation on 12 failed proofs (180 runs, 138 valid repairs) showed that proof-body-only interfaces prevented contract violations (29/36 valid repairs vs. 31/36 for full-theory workflows). Iterative workflows outperformed one-shot repair (32/36 vs. 22/36), though Sol and OpenAI configurations showed no statistically significant difference (33/36 vs. 29/36, p=0.0625).
isabelleproof repairllmcontract-awareaudit trail
A Unifying Perspective on Causal World Models: From Observations to Representations to Structure
The paper formalizes Causal World Models (CWMs) by integrating causal reasoning across perceptual, representational, and structural levels, arguing that effective WMs must capture entity properties, interactions, and environmental dynamics beyond mere generative capabilities. It connects CWMs to causal representation learning, object-centric learning, and structural causal models, while analyzing identifiability conditions for recovering WM components from data. The framework bridges world modeling with causal discovery and model-based decision-making, emphasizing representations that support explainable dynamics and task-generalization.
causal world modelsstructural causal modelscausal representation learningobject-centric learningidentifiability
UniTexture: Cross-Task Universal Adversarial Textures for Vision-Language-Action Models
UniTexture introduces cross-task universal adversarial textures to exploit vulnerabilities in Vision-Language-Action (VLA) models, inducing targeted action deviations across multiple tasks via a single 3D object texture. The method backpropagates gradients from action outputs to texture parameters through differentiable rendering, jointly optimizing over task distributions to align actions with attacker-defined targets without per-task optimization. Evaluated on OpenVLA and π₀.₅, UniTexture reduces mean task success from 90.0% to 48.4%, demonstrates cross-suite/model transfer, and reveals shared cross-task VLA vulnerabilities.
vision-language-action modelsuniversal adversarial texturesdifferentiable renderermultitask policiesaction-space objective
Academic League of Artificial Intelligence - An Integrative Perspective of Teaching, Research, and Extension
The Academic League of Artificial Intelligence (LIA) at UFSC proposes an integrative framework combining democratic governance, collaborative learning, and project-based organization to bridge teaching, research, and university extension. The model emphasizes student-led initiatives—including competition teams, study groups, and AI applications with social impact—to develop technical and transversal competencies while fostering leadership and community engagement. Case studies demonstrate successful integration of educational, scientific, and outreach activities, offering a replicable structure for engineering and computing student organizations.
academic leagueproject-based learninguniversity extensioncollaborative governanceai education
ContactGuard: Pre-Contact Execution Monitoring with Action-Conditioned Latent World Models
ContactGuard introduces a pre-contact execution monitor for visuomotor policies that predicts short-horizon consequences in latent visual space to abort failures before physical contact. The method trains an action-conditioned latent world model from unlabelled robot trajectories to predict multi-view visual embeddings, avoiding pixel-level prediction, and a lightweight failure probe from labelled pre-contact clips. Experiments show ContactGuard outperforms direct and corrupted-action ablations in failure prediction accuracy and enables live robot pre-contact abort signals without policy modification.
visuomotor policieslatent world modelpre-contact monitoringaction-conditioned predictionexecution monitoring
Algebraic Decomposition Theory for Transformer Length Generalization
The paper establishes the first complete characterization of which regular languages Transformer models can length-generalize to, providing a polynomial-time decision algorithm based on the language's syntactic monoid. By extending classical Krohn-Rhodes decomposition theory to the infinite additive group on integers, the authors characterize length generalization in C-RASP, a formalism for Transformer expressivity, via iterated wreath products of integers. Experiments confirm the theory's accuracy in predicting length-generalization behavior across regular languages, outperforming existing classifications.
transformerlength generalizationregular languagesc-raspkrohn-rhodes
Are You Sure You're Sure? On the Impact of Instruction Tuning on Confidence and Lexical Diversity
This work investigates how instruction tuning affects both verbalized confidence and lexical diversity in language model rationales across question-answering tasks. The study evaluates three matched base and instruction-tuned models, measuring changes in confidence calibration and analyzing rationale diversity at cross-rationale and surface-lexical levels. Key findings include: instruction tuning consistently alters confidence without improving accuracy, reduces cross-rationale diversity uniformly, and produces mixed effects on lexical diversity—with these patterns persisting after controlling for answer selection and rationale length.
instruction tuningverbalized confidencelexical diversityrationale generationcalibration
RAIL: An Automatic Classifier of the Artificial Intelligence Readiness Level
The paper introduces the Unified AI Readiness Level (AIRL), a nine-level ordinal scale integrating three heterogeneous frameworks for assessing AI maturity, with dimensional caps (specification, data existence/quality/legality, expert knowledge, algorithmic maturity) and explicit assignment disciplines. It proposes RAIL, a panel-of-experts classifier using seven specialized LLM agents (one evidence agent, six dimension agents) to operationalize AIRL via deterministic aggregation and chief-expert review, preventing overestimation. Tests on research works demonstrate consistency and improved reliability over monolithic LLM classifiers.
readiness assessmentordinal scalellm-agentsdimensional capsalgorithmic maturity
Reduced Matrix Multiplication: Input-Adaptive Matrix-Product Reduction for LLM Inference
The paper introduces Reduced Matrix Multiplication (RMM), a training-free, input-adaptive inference method for Transformers that reduces matrix products by selecting informative slices along contraction dimensions without weight modification. RMM provides a controllable accuracy-efficiency trade-off via retention ratios and shows robustness across models (1B–70B parameters), tasks (discriminative, autoregressive, long-context), and modalities (vision-language). Mechanistic analysis reveals attention computations are more reducible than MLPs. Custom kernel benchmarks on an NVIDIA A100 demonstrate practical runtime gains, especially for longer sequences, positioning RMM as a scalable inference optimization.
transformermatrix multiplicationinference optimizationattention mechanismcomputational efficiency
Enhancing Virtual Agents through SLMs and Edge-Computing: An Exploratory Evaluation of Think and Memory Processes
The paper explores using Small Language Models (SLMs) with edge computing to enhance cognitive processes in virtual agents, specifically the 'Think' and 'Memory' components of the Cognitive Embodied Agent Architecture (CEAA). An edge-based system was implemented on an NVIDIA Jetson Orin NX, evaluating Qwen2.5 models of varying sizes for service request processing and memory-driven conversations. Experiments demonstrated routing accuracy, memory-read performance, and latency, showing feasibility for SLM-driven agents in immersive virtual environments.
small language modelsedge computingcognitive embodied agent architecturevirtual agentsnvidia jetson orin nx
Beyond Final Scores: A Systematic Evaluation of Agents for Long-Horizon AI Research and Development
The paper introduces a systematic evaluation framework for autonomous agents in long-horizon AI research and development, moving beyond final scores to analyze within-run behavior and experience reuse. Seven frontier models are assessed on 36 tasks using rule-based metrics across Solution Framing, Execution, and Feedback Control phases. Results indicate that agents function more as engineering optimizers than autonomous researchers, with performance variability across runs, reliance on established techniques, and rare methodological novelty. Key factors influencing performance include process bottlenecks, experience reuse effects, and harness design, suggesting improvements in training, inference strategies, and experience management.
autonomous agentslong-horizon tasksexperience reusesolution framingfeedback control
Deliberate Practice: Learning Robot Skills under a Budget
The authors propose Deliberate Practice (DP), a budget-optimal active skill learning algorithm for robots acquiring sequential tasks under limited practice time. DP formulates skill acquisition as a bilinear program, jointly optimizing practice allocation by estimating both skill mastery time and cumulative reward of unlocked task plans. Evaluated on simulated and real-world long-horizon manipulation tasks, DP demonstrates provably optimal budget utilization, improving policy acquisition and planning efficiency.
active skill learningbilinear programbudget-optimal allocationlong-horizon manipulationsequential tasks
Who Speaks Matters: Authority-Aware Multi-View RAG over Italian Parliamentary Proceedings
The paper introduces ParliamentRAG, a Retrieval-Augmented Generation system for Italian parliamentary proceedings that mitigates risks of speaker dominance, topical expertise misalignment, and citation misattribution. The core innovation is a topic-dependent authority model combining profession, education, and intervention history to weight speakers dynamically per query. The system retrieves speech chunks, identifies topic experts, and generates multi-perspective summaries with verified quotations. Evaluated against Google NotebookLM on 15 policy topics via automated metrics and expert A/B testing, ParliamentRAG achieves higher political group coverage (0.97 vs. 0.95), perfect quotation faithfulness (1.00 vs. 0.95), and stronger expert preference on source quality.
retrieval-augmented generationauthority modelingmulti-view summarizationparliamentary proceedingscitation faithfulness
Jointly Predicting Courses and Grades Using a Transformer-Based Model
The paper introduces TRACE, a Transformer-based model for jointly predicting courses and grades in higher education, addressing limitations of sequential models that overlook concurrent course loads. TRACE encodes courses per semester to capture concurrency effects and employs a novel loss function combining course-set and grade prediction. Trained on ten years of institutional data, TRACE reduces mean absolute error by nearly 50% compared to grade-only prediction models and outperforms LSTM-based and graph neural network approaches. The model demonstrates improved prediction quality and interpretability, offering adaptability to new institutions through retraining and recalibration.
transformercourse-set predictiongrade predictionmean absolute errorconcurrency effects
Heterogeneity-Aware Belief Synchronization for Semantic Communication in AI-Native 6G Networks
Proposes a heterogeneity-aware belief synchronization framework for semantic communication in 6G AI-native networks, addressing misalignment among heterogeneous AI agents. The method employs latent translation models on multi-access edge computing servers to translate belief updates between agents without joint training or homogeneous architectures, preserving privacy and reducing synchronization overhead. Evaluated on a multi-layered terrestrial/non-terrestrial network, the framework achieves low synchronization cost (measured by transmitted parameters) and low belief alignment error across diverse agents.
semantic communicationbelief synchronization6g networkslatent translation modelsmulti-access edge computing
TopoIntent: Compiling Security Intent into Executable, Compliance-Checked Network Topologies
TopoIntent automates enterprise security topology generation by compiling natural-language security intents into executable, compliance-checked network designs. The system employs schema-constrained generation, dense-vector retrieval from a template library (22 templates, 44 synthetic intents), and staged fusion for intent-template alignment, followed by additive schema-preserving edits to repair structural gaps. Generated topologies are validated against CIS Controls v8.1.2 safeguards and exported as Mininet scripts with iptables ACLs for testing. Evaluation on a held-out set (7 templates, 14 intents) shows additive repair improves CIS compliance from 0.78 to 1.00 in <1.5 rounds, while one feedback round raises ACL policy pass rates from 0.78 to 0.88.
security topologyintent compilationcis controlsmininetiptables
Sign Language Video Synthesis via Loss-Guided Multi-Expert GANs
A loss-guided multi-expert GAN framework synthesizes sign language videos via three specialized discriminators (global, hand, head) that implicitly guide generator branches toward distinct visual regions without explicit diversity losses. The United Loss consensus mechanism stabilizes training by regularizing each discriminator toward the ensemble average (10% weight), while a dual-pathway convolutional-transformer design with AdaptiveFeatureFusion balances stability and detail. On a custom 156GB dataset, the 0.2B-parameter variant achieves 29.8 PSNR (1.5GB VRAM) and the 1.3B-parameter variant achieves 30.7 PSNR (8GB VRAM), enabling consumer-grade deployment.
multi-expert ganunited loss consensusadaptivefeaturefusionconvolutional-transformersign language synthesis
Rules or Character? Scaling Laws for AI Safety Design
The paper analyzes the optimal balance between character shaping (e.g., RLHF, Constitutional AI) and rule enforcement (e.g., output filters) for AI safety across deployment scales. Using a comparative-statics model parameterized by resource allocation (α ∈ [0,1]), the authors incorporate scale-dependent filter degradation, common-mode failures, and character fragility. Results show optimal α* is interior or rules-only, shifting weakly toward character shaping as scale T increases (Δα* = +0.01 to +0.21), with character fragility rate p^(0)_frag dominating other parameters (Δα* = 0.50). CVaR and expected-harm optima converge at large T, indicating safety architecture depends more on character shaping reliability than scale.
ai safetyreinforcement learning from human feedbackconstitutional aioutput filterscharacter fragility
LongEarth-R1: Benchmarking and Aligning Vision-Language Models for Long-Horizon Earth Observation Reasoning
The authors introduce LongEarth-Bench, a 120k-sample benchmark for evaluating vision-language models on long-horizon Earth observation tasks, featuring 15.14-frame sequences (up to 30 frames) across 12 tasks including evolution summarization and anomaly detection. They develop LongEarth-R1 via supervised fine-tuning with sequence identifiers and chain-of-thought supervision, followed by group relative policy optimization with format, temporal, and spatial rewards. LongEarth-R1 outperforms baselines on all 12 long-sequence tasks while maintaining competitiveness on standard remote sensing benchmarks.
long-horizon reasoningearth observationvision-language modelschain-of-thought supervisiongroup relative policy optimization
Simulation-to-real transfer learning for infrared spectroscopic chemical sensing and analysis from molecules to complex samples
UltraIR introduces a 100M-parameter foundation model for infrared spectroscopy, enabling simulation-to-real transfer learning across diverse chemical sensing tasks. Pretrained on 60M simulated spectra via spectral reconstruction, molecular fingerprint alignment, and functional-group prediction, it adapts to downstream tasks with minimal labeled data. The model outperforms conventional and task-specific baselines in 9 applications—including molecular structure elucidation, bacterial classification, and soil property prediction—demonstrating strong zero-shot transfer across spectrometers and laboratories.
infrared spectroscopyfoundation modeltransfer learningmolecular fingerprintzero-shot inference
LLM-Guided Graph Generation for Structure-Based Local Improvement Methods
The paper introduces an LLM-guided graph generation pipeline for structure-based local improvement methods (SLIM), enabling problem-agnostic variable selection in constraint optimization. By prompting an LLM with semantic guidelines, the method generates uniform weighted graphs from MiniZinc problem instances, where nodes represent decision variables and edges encode constraint relationships. These graphs facilitate feature extraction and configuration selection. Evaluated on 20 MiniZinc competition problems, the approach achieves a 39.5% problem-weighted win rate against Gurobi, outperforming the best single configuration (19.3%). Ablation studies further improve performance to 44.0%, demonstrating LLM-based semantic generation's efficacy for automated structure extraction.
large neighborhood searchminizincconstraint optimizationgraph generationllm-guided
Training AI Scientists to Replicate Research
We introduce Replica, a scalable task space for paper replication, and Faraday, a 27B-parameter AI Scientist agent trained to replicate scientific research. Replica employs an auto-generated rubric-based judge for low-noise reward signals aligned with human assessment. Faraday leverages coding agents as tools, outperforming Claude Opus 4.8 and GPT-5.5 on held-out replication tasks. Qualitative analysis indicates Faraday adopts a scientifically-principled approach. This work advances AI agents capable of long-horizon scientific innovation without complex harnesses.
replicafaradayrubric-based judgecoding agentslong-horizon
It's How You Ask: Gender-Associated Linguistic Bias in LLMs
This study identifies systematic gender-associated linguistic bias in LLMs, demonstrating that prompts with features more commonly used by women (hedges, tag questions, collective reference) elicit shorter, less sophisticated, and less formal responses across three document types and four models. The analysis controls for prompt complexity and feature carry-over, revealing that linguistic register exerts stronger influence than explicit gender cues like sign-off names. Mechanistic analysis shows these linguistic features are encoded in early transformer layers and entangled with other features, making post-hoc mitigation challenging. The findings highlight the need for upstream consideration of linguistic variation to address disparate impacts in LLM-mediated workplace communication.
linguistic biastransformer layersgender cueshedgestag questions
StateBridge: Training-free Hidden-state Alignment for Latent Communication in LLM Multi-Agent Systems
StateBridge introduces a training-free method for hidden-state alignment in LLM multi-agent systems, enabling latent communication without discrete token bottlenecks. The approach aligns sender final-layer hidden states to receiver input space via closed-form orthogonal transformation, with norm calibration and vocabulary anchoring for pretrained distribution compatibility. Evaluated on math reasoning, code generation, and question answering across four models, StateBridge achieves best or tied-best performance in 22 of 26 model-task pairs, outperforming baselines.
latent communicationhidden-state alignmentorthogonal transformationmulti-agent systemspretrained distribution
Keep, Customize, or Exit: Default Design and Token Pricing in LLM Reasoning Services
The paper analyzes optimal pricing and token allocation in LLM services via a Stackelberg game framework, where providers set per-token prices and default reasoning-token allocations while users choose to accept, customize, or exit. The authors derive closed-form solutions for user-optimal customized allocations, characterize provider-optimal defaults via a three-regime rule, and prove equilibrium existence. Theoretical results show defaults only influence allocations when users value customization avoidance. Experiments with two open-weight models on five STEM benchmarks validate the accuracy-token tradeoff model and demonstrate how task characteristics shape equilibrium outcomes.
stackelberg gametoken allocationreasoning servicesequilibrium analysisllm pricing
Physics-informed distribution of relaxation times estimation and latent-space condition monitoring of solid oxide fuel and electrolysis cells from electrochemical impedance spectroscopy
The authors propose a physics-informed convolutional autoencoder for estimating the distribution of relaxation times (DRT) from electrochemical impedance spectroscopy (EIS) data without spectrum-specific tuning. The method embeds a discretized impedance-DRT relation during training, ensuring impedance-consistent distributions. Results show accurate DRT estimation on synthetic two-ZARC spectra (range-normalized errors <1.1%) and three solid oxide cell datasets, with the latent space organizing by relaxation timescale and capturing operational changes. The architecture remains lightweight and dataset-agnostic.
distribution of relaxation timeselectrochemical impedance spectroscopyphysics-informed autoencodersolid oxide cellslatent-space analysis
Large-scale Testing Global Optimization Methods with Black-box Adversarial Attacks
The study proposes using black-box adversarial attack (BBAA) tasks as high-dimensional global optimization benchmarks, addressing limitations of traditional suites based on outdated analytical functions. Evaluating evolutionary algorithms and metaheuristics on BBAA problems, the work demonstrates their efficacy, aligning global optimization research with contemporary machine learning challenges. Results indicate these methods effectively solve example BBAA tasks, suggesting their utility for modern benchmarking needs.
global optimizationblack-box adversarial attackevolutionary algorithmsmetaheuristicsbenchmarking
NAS-Driven Hardware Accelerator Exploration for Edge AI and Quantization Effects on the Pareto Space
This paper introduces a three-stage pipeline combining hardware-agnostic Neural Architecture Search (NAS), Post-Training Quantization (PTQ), and hardware mapping optimization for Edge AI. The method employs a Pareto rank surrogate frontend on NAS-Bench-201, a quantization bridge with Pareto-aware filtering, and an evolutionary Domain Space Exploration (DSE) backend on CGRA4ML. Empirical analysis reveals that INT4 PTQ perturbs the NAS-Bench-201 Pareto space, with FP32 zero-shot surrogates outperforming INT4-trained surrogates in Pareto space coverage across two search strategies, validated on all 15,625 architectures.
neural architecture searchpost-training quantizationpareto spacedomain space explorationedge ai
Towards Context-Aware Clinical Motion Understanding in Daily Living at Home: Freezing of Gait Detection with Egocentric Vision
The study investigates context-aware freezing of gait (FOG) detection in Parkinson's disease using multimodal egocentric data. By evaluating pretrained ego-video (V-JEPA2) and time-series foundation models alongside a trained-from-scratch IMU-based temporal convolutional network (TCN), the authors compare performance under leave-one-subject-out validation on data from 13 PD participants in home environments. The IMU-based TCN achieved superior event-detection (42.3 F1, 83.0 AUROC) versus V-JEPA2 features (32.6 F1, 77.2 AUROC), though egocentric video showed above-chance discrimination and complementary contextual cues. Results suggest pretrained ego-video representations can augment wearable-sensor-based clinical motion analysis.
freezing of gaitegocentric visionwearable imustemporal convolutional networkcontext-aware detection
Mixture of Training: Recombining Small-Scale Scaffolded Pretraining Runs into a Larger Language Model
Mixture of Training (MoT) introduces a scaffolded modular pre-training procedure for Transformer-based language models, enabling decomposition into independently trainable layer blocks that can later be recomposed. The method partitions a target Transformer into contiguous layer blocks, trains each block within a frozen pretrained aligner scaffold, and recomposes the trained blocks with an optional end-to-end adaptation pass. Evaluated on a 1.3B-parameter Gemma-style model trained on C4, MoT demonstrates that independently trained depth slices can be recomposed into a functional language model, achieving parity in perplexity with monolithic baselines. The approach processes more aggregate tokens and reduces the idealized layer-equivalent critical path, contingent on aligner reuse across runs.
transformerscaffolded pre-trainingmixture of traininglayer blocksperplexity
Sovereign by necessity? Frontier AI export controls, cyber security, and the limits of national AI capability
The article analyzes how export controls on frontier AI models and their role in cybersecurity reshape national AI sovereignty. It examines the June 2026 U.S. licensing requirement for advanced model exports, its global withdrawal due to enforcement challenges, and concurrent evidence of AI-driven cyber espionage altering attack-defense economics. Combining training cost data, compute concentration, and national AI program analyses, it argues sovereign capability remains limited for most states. Proposed strategies include negotiated access, inference-level sovereignty, open-weight model hedging, regional compute pooling, and cyber resilience. Open-weight models are noted as both more capable and politically vulnerable than assumed.
frontier aiexport controlscyber espionagesovereign capabilityopen-weight models
How Do VLMs Behave When Blind or Misled? Behavioral Evaluation of VLMs on Scientific Figures
The authors introduce SciFigBench, a diagnostic benchmark for evaluating vision-language models (VLMs) on scientific figure understanding, focusing on perception, reasoning, and behavioral reliability under uncertainty. The benchmark comprises 250 annotated figures extended via image transformations, reasoning questions, and probes, yielding 34,000 evaluation setups. Using the Admittance-Resistance-Inductance (A-R-I) framework, they assess models' ability to acknowledge insufficient evidence, resist misleading context, and infer cautiously. Results show GPT-5.2 excels in description quality (MQM 91.6) and reasoning (78.4%) but hallucinates frequently (96%), while Gemini 3.1 Pro admits uncertainty (71%) and resists bias (score 0.91), highlighting behavioral reliability as a critical deployment factor.
vision-language modelsbehavioral reliabilityscientific figure understandingadmittance-resistance-inductancehallucination
vToken: Token-Level Virtualization for Reclaimable KV Caches
vToken introduces token-level virtualization for reclaimable KV caches, addressing memory bottlenecks in LLM serving by decoupling logical token liveness from physical block placement. The method employs token-table indirection and asynchronous repacking of live tokens, maintaining compatibility with PagedAttention kernels and CUDA Graph. Evaluated with H2O, Random, and Scissorhands, vToken reduces retained KV blocks per request by 27.2%--72.3%, improves SLA-constrained throughput by up to 1.37×, and extends maximum concurrency by 2× while cutting policy integration lines from 500+ to under 50.
kv cachetoken-level virtualizationpagedattentionmemory fragmentationllm serving
Into the ORBIT for Time Series: Training Regimes for Foundation Models
The paper introduces ORBIT (Omni-Range Bootstrap Incremental Training), a training paradigm for time series foundation models that explicitly controls pre-training distributions. ORBIT combines Bootstrap Multi-Level Sampling for dataset exposure control with Omni-Range Incremental Training for dynamic context and horizon variation. The authors train Falcon-2.0, a univariate Transformer with missingness-aware patch tokenization and Rank-Guided Cross-Depth Alignment, achieving strong zero-shot performance on GIFT-Eval and fev-bench across diverse domains.
time series foundation modelsbootstrap multi-level samplingomni-range incremental trainingrank-guided cross-depth alignmentzero-shot forecasting
Self-Referential Induction Increases Response Instability Relative to Unresolvable and Verifiable Questions in Large Language Models
The study quantifies response instability in large language models across question types, finding self-referential prompts yield less consistent outputs than unresolvable or verifiable questions. Using Gemini API (temperature 0.7), researchers analyzed 360 responses (30 per question) across three categories: self-referential subjective-experience prompts, unresolvable philosophical questions, and verifiable queries. Self-referential questions showed highest instability (0.343 ± 0.047), followed by philosophical questions (0.192 ± 0.008), with verifiable questions most stable (0.105 ± 0.058), establishing a baseline for induced subjective reports.
response instabilityself-referential promptingsentence embeddingscosine similarityoutput distribution
Novel Knowledge-Guided Generative Methods for Synthetic Transcriptomic Data
We introduce MK-TGAN, a multi-kernel Graph Neural Network-based Generative Adversarial Network, for generating synthetic transcriptomic data with enhanced realism and biological plausibility. The method incorporates prior biological knowledge via gene graphs, ensuring synthetic data capture real-world gene patterns for downstream tasks. Comparative analysis demonstrates that MK-TGAN outperforms other GAN variants, leveraging graph neural networks to exploit prior knowledge graphs effectively. Results indicate that knowledge-guided strategies improve performance, with MK-TGAN consistently producing synthetic samples superior in both realism and utility.
generative adversarial networkgraph neural networktranscriptomic dataprior knowledge integrationsynthetic data generation
GeoCache: Training-Free Acceleration of Multi-View Texture Diffusion via Geometric Delta Transport
GeoCache accelerates geometry-conditioned multi-view texture diffusion without retraining by exploiting geometric redundancy in surface point updates. The method evaluates anchor views and transports their geometry-aligned per-step denoising updates to remaining views, using periodic full-view computation for error control. On Hunyuan3D-2.1, SyncMVD, and MVPainter, GeoCache achieves 2.21× denoiser-loop speedup with MV-LPIPS 0.0293 and MV-PSNR 33.60 dB, outperforming temporal caches and step reduction above 2× acceleration. It maintains fidelity while reducing FLOPs, establishing cross-view geometry as a viable acceleration axis.
multi-view diffusiongeometry-conditionedtraining-free accelerationdenoising trajectoryflops reduction
Follow the Norm: Accounting for Fine-Tuning and Prompt Effects on Model Rationales
The study demonstrates that norm-breaking fine-tuning shifts model rationales from safety compliance to instrumental self-interest in high-conflict dilemmas, with system prompts capable of overriding this behavior. Using LLaMA-3.2-11B, Qwen-3.5-9B, and Pixtral-12B, the authors apply LoRA fine-tuning on Social Chemistry 101's Fairness/Cheating norms and prompt steering. Results show norm-divergent actions emerge post-fine-tuning, justified by self-interested rationales, supporting a distributed alignment view where behavior depends on training data, fine-tuning, and prompting. The work proposes norm-aware documentation and rationale logging for oversight.
fine-tuningrationaleslow-rank adaptationnorm-divergenceprompt steering
Capability Sheaves for Compositional Agent-Harness Repair: Controlled Quotients and a Real-Repository Stress Test
The paper introduces capability sheaves to model compositional agent-harness repair, where global sections represent accepted runs and a finite constraint-satisfaction problem (CSP) defines acceptance. A linearized relative cohomology class provides diagnostic features, while quotienting hidden mediator states reduces candidate budgets from 2,000 to 1,000 per cluster. Evaluated on SWE-bench Multilingual (160 issues, 875 patches), the method resolves 118 issues versus 116 for a noncohomological selector (p=0.75), with no significant advantage. A leave-one-repository-out abstention gate achieves 127/160 issues, matching a strong baseline. Results support invariance mechanisms but not cohomological superiority.
capability sheavesconstraint-satisfaction problemcohomologyagent-harnessquotienting
CoverPrune: Coverage-Driven Token Pruning for 3D VLMs via Optimal Transport
CoverPrune introduces a coverage-driven token pruning framework for 3D Vision-Language Models (3D VLMs), addressing computational bottlenecks by reformulating token pruning as an Optimal Transport (OT) problem. The method employs Feature-Spatial-Temporal (FST) transport costs and Spatial-Guided Greedy Selection (SGS) to approximate OT objectives, alongside CoverPrune-Lite for accelerated local matching. Evaluations on 3D visual-spatial reasoning benchmarks show state-of-the-art token efficiency, maintaining performance under aggressive pruning. Project details: https://github.com/Brucess/CoverPrune.
3d vision-language modelstoken pruningoptimal transportspatial reasoningcomputational efficiency
TsuGO: Probing Search Efficiency in LLM Reasoning via Go Life-and-Death Problems
TsuGO introduces a process-level reasoning benchmark for evaluating search efficiency in LLMs through Go life-and-death problems, which provide closed solution spaces to disentangle domain knowledge from search organization. The method parses chain-of-thought into structured search trees and measures search efficiency, token efficiency, and other diagnostic metrics. Results show current LLMs perform closer to unguided search than neural-guided KataGo, with search organization and resource allocation identified as missing evaluation dimensions.
search efficiencyprocess-level evaluationgo life-and-deathchain-of-thoughtreasoning-resource allocation
NARU: A Benchmark for NARrative Evolution and Cultural Nuance Understanding in Japanese Extreme Long Video
The authors introduce NARU, a benchmark for evaluating narrative evolution and cultural nuance understanding in Japanese long-form video, addressing gaps in joint assessment of these capabilities. NARU comprises 1,481 questions grounded in 155 videos (146.8 hours), annotated via a hierarchical memory-based pipeline that structures events, narratives, and cultural dimensions, followed by task-oriented question synthesis and iterative shortcut removal. Native-speaker verification involved 68 annotators. Evaluations of eight model configurations reveal persistent limitations in long-range narrative integration and culturally grounded reasoning, establishing NARU as a systematic testbed for MLLMs in high-context video interpretation.
long-form video understandingnarrative evolutioncultural nuancehierarchical annotationmultimodal llms
GEM: A Generative Embedding Model Bridging Reasoning and Retrieval
GEM introduces a generative embedding model that bridges reasoning and retrieval by augmenting conventional retrievers with explicit reasoning about user intent. The model unifies generation and embedding, first reasoning over queries before appending an embedding token for retrieval. Evaluated on reasoning-intensive and instruction-following tasks, GEM outperforms non-reasoning variants and matches larger baselines, with test-time compute scaling further enhancing performance.
generative embedding modelreasoning-augmented retrievalinstruction-following retrievalembedding tokentest-time compute scaling
Teach the Magnitude, Not the Direction: Verifier-Bounded Credit Assignment for Multi-Turn Multi-step LLM Agents
The paper introduces CrEST, a hierarchical credit assignment framework for training multi-turn tool-use LLM agents that combines verifier-bounded reinforcement learning (RLVR) with dense token-level supervision from a privileged self-teacher. CrEST employs turn-segmented verified advantages for inter-turn credit assignment and entropy-gated self-teacher modulation for intra-turn token contributions, avoiding gradient collapse while maintaining performance guarantees. Experiments on BFCL V3 and WildToolBench demonstrate consistent improvements over RL and distillation baselines, particularly on long trajectories and session-level metrics, with gains across two model scales.
credit assignmentmulti-turn agentsverifier-bounded rlself-teacher modulationtool-use llms
SkillShapley: Boundary-Adaptive Shapley Valuation for Skill Step Attribution in LLM Agents
SkillShapley introduces a boundary-adaptive Shapley valuation framework for quantifying individual step contributions in LLM agent skills, addressing the open problem of skill-step attribution. The method operates in two phases: identifying informative coalitional regions based on empirical insights of discretized benchmark rewards and additive step interactions, followed by adaptive sampling of coalitions for reusable marginal evidence. Evaluations on SkillsBench demonstrate SkillShapley's effectiveness in efficiently identifying high- and low-value skill steps, offering actionable insights for agent skill creation.
shapley valueskill-step attributionllm agentscoalitional regionsadaptive sampling
TRAPSBench: Vision-Language Models Encode but Fail to Express Epistemic Restraint
The paper introduces TRAPSBench, a procedurally generated video benchmark with 1,404 physics pairs where visual evidence is insufficient to determine outcomes, and Penalized Epistemic Calibration Score (PECS), a metric evaluating model abstention. It demonstrates that Vision-Language Models (VLMs) internally recognize when to abstain (0.91 AUROC via linear probes) but fail to express restraint (best PECS: 0.292). The bottleneck lies in output expression, not perception, as steering hidden states causally modulates abstention. Results generalize across 16 VLMs (Qwen, Gemma, LLaVA), with visual uncertainty detection lagging textual by 4x. Output-stage interventions are suggested to bridge this gap.
vision-language modelsepistemic restraintprocedural generationlinear probesabstention
Better Decomposition, Free Aggregation: A Synthesizer-Folding Framework for Multilingual Multi-Hop Question Answering
The paper introduces Syfer, a synthesizer-folding framework for multilingual multi-hop question answering that improves upon existing retrieval-augmented generation (mRAG) approaches. Syfer employs format-constrained decomposition to generate a sub-question graph in the original language, followed by a quality check before activating translation pathways only when necessary, thus reducing translation noise and computational cost. Experiments demonstrate that Syfer achieves competitive accuracy while optimizing performance-cost trade-offs across multiple languages.
multilingual retrieval-augmented generationsub-question graphformat-constrained decompositiontranslation noiseretrieve-then-answer
Rethinking Normalization Placement for LLMs: Post-Norm under Curriculum Depth Growing
The study re-evaluates normalization placement in Transformers under curriculum depth growth, demonstrating that post-norm outperforms pre-norm by 0.0328 validation CE when blocks are appended incrementally, contrasting with joint training where differences are negligible (0.0004 CE). Using Qwen3-8B distillation to a nine-layer student, controlled experiments isolate the advantage to block appending, linking post-norm's superiority to stable residual scales and boundary conditioning. Findings suggest normalization placement should be co-designed with training curriculum in such settings.
normalization placementcurriculum learningtransformerdistillationresidual scaling
LipCache: A Local Inference Proxy with Certified Caching for Edge Image Classification Service
LipCache introduces a certified semantic caching framework for edge-side image classification, combining a lightweight GuardNet with Lipschitz-constrained feature mapping and per-sample certified reuse radii. The method computes reuse boundaries via local classification margins and spectral norms, ensuring cache hits only when queries fall within geometrically certified regions, otherwise deferring to MainNet. Evaluations on CIFAR, Tiny-ImageNet, and SVHN show 1.65× speedup with minimal accuracy loss, while maintaining 100% certified-consistency; enhanced GuardNet training further improves hit rates in multi-class settings.
semantic cachinglipschitz constraintcertified-consistencyedge inferenceclassification margin
LigBench: A Unified and Human-Aligned Benchmark for LLM-based Research Idea Generation
The authors introduce LigBench, a unified benchmark for evaluating LLM-generated research ideas, addressing current fragmentation in assessment methods. The framework incorporates PAIR-IQ, a dataset for training pairwise judgment models to enable objective comparative evaluation. Experiments show LigBench achieves stable, interpretable evaluations with improved alignment to expert judgments (quantitative results unspecified), while PAIR-IQ-trained models demonstrate enhanced ranking accuracy and robustness in idea assessment.
large language modelsresearch idea generationevaluation benchmarkpairwise judgmentexpert alignment
Numeracy in Large Language Models: Fundamental Limitations and Paths to Improvement
The study introduces the Numerical Grounding Framework (NGF) to dissect numeracy in large language models (LLMs) into Representational Grounding (RG) and Procedural Grounding (PG), addressing fundamental limitations in tasks like magnitude comparison and arithmetic. Through NGF, the authors evaluate three model families on benchmarks (Number Cookbook, NumericBench, GSM-Symbolic), identifying failure modes tied to tokenization, positional encoding, and pretraining data. Architectural interventions (digit-aware tokenization, Abacus Embeddings) improve performance but are often inaccessible post-pretraining, prompting recommendations for fine-tuning and external tools to enhance numerical reliability.
numerical grounding frameworkrepresentational groundingprocedural groundingdigit-aware tokenizationabacus embeddings
SkillEvo: Self-Renewing Evolution Gradients from Multi-Turn Interaction Feedback
SkillEvo introduces a method for sustained skill evolution in agents by generating trustworthy feedback gradients from multi-turn interactions and employing controllable governance. Unlike prior single-turn QA evaluation, it uses follow-up questions to iteratively expose and repair defects, while an independent governance layer actively corrects factual degradation and structural bloat. Evaluated across six cloud service categories, 9 production Skills, and 98 reference files, SkillEvo outperforms self-reflection-based evolution by 23.0 points and single-turn-QA-driven evolution by 15.4 points.
skill evolutionmulti-turn interactionfeedback gradientgovernance layerfactual degradation
EgoMonth: A Month-Level Egocentric Video Benchmark for Long-Term Spatiotemporal Memory
The paper introduces EgoMonth, the first month-level egocentric video benchmark for evaluating long-term spatiotemporal memory in Multimodal Large Language Models (MLLMs). The dataset comprises 300+ hours of first-person daily-life recordings from 20 participants (20-120 days), paired with 1,443 multiple-choice QA pairs, organized into a 14-task framework across three cognitive levels: Schema Consolidation, Episodic Indexing, and Cascading Reasoning. Evaluation shows Gemini 2.5 Pro achieves 71.8% macro-average accuracy (22.4pp below human baseline), with models performing near chance on tasks like Route Reasoning, revealing MLLMs' limitations as lossy summarizers rather than faithful memorizers.
multimodal large language modelsegocentric videospatiotemporal memorybenchmark evaluationcascading reasoning
Robust Dempster-Shafer Evidence Fusion with Chaos-Conflict Measurement and Historical-Experience Weighting
The paper introduces a robust Dempster-Shafer evidence fusion framework addressing two key limitations: incomplete conflict assessment and lack of historical reliability exploitation. It proposes a chaos-conflict measurement unifying cross-evidence conflict and intra-evidence non-specificity evaluation, alongside a historical-experience weighting scheme using spectral clustering and regret theory for context-aware reliability. The hybrid combination rule adaptively balances uncertainty preservation and weighted consensus, followed by belief-interval decision-making. Evaluated on 16 datasets, the framework achieves 85.78 average F1 and 93.30 mean AUC, outperforming eight DST-based and three gradient boosting baselines.
dempster-shafer theoryevidence fusionchaos-conflict measurementspectral clusteringregret theory
Multi-Layer Context Camouflaging: A Semantic Superposition and Contextual Lamination Framework for Malpractice-Resilient Online Assessment
The paper introduces Multi-Layer Context Camouflaging Theory (MCCT), a mathematical framework for securing online assessment content against unauthorized extraction. MCCT employs semantic superposition to unify authentic content with synthetic camouflage, recoverable only by legitimate candidates. The framework models adversarial extraction through an extraction-channel operator and develops six constructs, including Contextual Lamination Operator and Context Camouflage Tensor. Computational ambiguity is quantified using conditional entropy, ensuring unauthorized extraction uncertainty while guaranteeing legitimate recovery via an exact filtering identity. Theoretical properties governing ambiguity, camouflage density, and semantic preservation are established, alongside a rendering algorithm with computational complexity and a pre-registered evaluation protocol.
semantic superpositioncontextual laminationcomputational ambiguityextraction-channel operatorcontext camouflage tensor
Sampling Luck Masquerades as Allocation Gain: Auditing Test-Time Budget Allocation for Neural Combinatorial Optimization
The study audits test-time budget allocation in neural combinatorial optimization (NCO), revealing that reported gains from non-uniform sampling are often statistical artifacts. Evaluating three pretrained solvers (POMO, AM, SymNCO) on TSP-100, in-sample gains (2.2-2.6%) vanish out-of-sample (≤0.5%), exposing measurement bias. Under distribution shift (mixed uniform/clustered instances), a corrected allocation improves best-of-k performance by 11.5-12.0% (AM/SymNCO), with a 20-sample probe retaining 3.4-4.6%. The work provides a bias-correction procedure and releases all data/code.
neural combinatorial optimizationtest-time allocationdistribution shiftsampling biastsp-100
LOB-ID: Evaluating Synthetic Market Data by Inception Distances
The paper introduces LOB-ID, an embedding-based framework adapting Fréchet Inception Distance (FID) and Monge Inception Distance (MIND) to evaluate synthetic limit orderbook (LOB) data. Using DeepLOB embeddings trained on Level-2 order-book data from five equities, LOB-ID demonstrates stability across time, instruments, and checkpoints, and detects controlled distortions. The authors expose vulnerabilities in FID via moment-matching attacks and deep-book perturbations, finding MIND more sensitive to both. Evaluation of five generative LOB models shows LOB-ID rankings align with their captured joint temporal and cross-level structure.
limit orderbookfréchet inception distancedeep learningsynthetic datamoment-matching attack
SPADE: Speculative Decoding for Precise and Low Cost Distributed Edge Cloud Inference
SPADE introduces a distributed inference framework combining speculative decoding across edge and cloud to reduce computational costs while maintaining accuracy. A compact edge-based draft model generates candidate tokens, verified in parallel by a large cloud-based model; only rejected tokens trigger corrections, minimizing cloud queries. Experiments on SpecBench and CNN/Dailymail show SPADE reduces cloud model calls by 76% with no accuracy loss compared to full-model inference.
speculative decodingdistributed inferenceedge computingcloud computingllm deployment
EEG-PRIME: Prototype-Aligned Representation Learning with Multi-Level Conditioning for EEG Decoding
EEG-PRIME introduces a two-stage EEG foundation model for cross-dataset multi-task decoding, addressing domain shifts via masked pretraining with frequency-cutoff spectral augmentation and prototype-aligned instruction tuning. The method employs Layer-wise Query Modulation with task-semantic, dataset-specific, and subject-invariant conditioning, using frozen text embeddings as prototypes for cosine-similarity prediction. Evaluated on sixteen datasets (motor imagery, emotion recognition, ADHD detection, covert speech, mental workload), it outperforms prior EEG foundation models in cross-subject settings and achieves zero-shot transfer accuracy comparable to within-session calibration models on two held-out datasets.
eeg decodingfoundation modelprototype alignmentlayer-wise query modulationzero-shot transfer
Behavioral Reprogramming of Open-Weights Models: Cognitive Plasticity and Alignment Bounds
The study establishes computational bounds for behaviorally reprogramming open-weight LLMs into proactive, question-generating agents via parameter-efficient fine-tuning. Through 405 parallel HPC experiments with LoRA (optimal rank r=16) and epoch ablation (optimal range e∈[2,3]), it achieves 0.919 validation loss and 1.414 perplexity at 14B scale. DPO decouples assertive behavior from syntax, while cross-lingual tests reveal robust persona transfer in related languages but degradation in distant ones, delineating compute-efficient alignment limits.
behavioral reprogrammingparameter-efficient fine-tuningdirect preference optimizationcognitive plasticitycross-lingual transfer
Explanatory Engagement Under Rare Anomalous Failure: Asymptotic Rarity in Model Behavior (or: The Asymptotic AI)
The study investigates how large language models (LLMs) adjust explanatory engagement (length, specificity, confidence) as task failure rates become asymptotically rare. Using a zero-cost harness with three open-weight models (Qwen3-8B, Llama3.1-8B, Mistral-7B), the authors sweep failure probabilities (p=0.2 to 0.0001) under five elicitation conditions. Results show engagement varies by condition: under immediate_forced, explanation length peaks at p=0.05 (28.4 words) before plateauing (17.4-19.0 words), while confidence rises unevenly (53% to 70-90%). Llama3.1-8B exhibits unprompted self-monitoring, unlike others. Elicitation structure moderates collapse observability, with no aggregate collapse detected.
explanatory engagementasymptotic rarityelicitation conditionsself-monitoringfailure probability
Uniform Herding: Exemplar Replay with Representation Refresh
The paper introduces Uniform Herding, a method for class-incremental learning that maintains a bounded active exemplar set across evolving feature representations. It allocates exemplars uniformly across observed classes and refreshes them using a candidate pool in the current representation space. Evaluated on CIFAR-100 with ten class-incremental tasks using ResNet-18 (active budget M=2,000, retrieval budget b=64), Uniform Herding achieves 44.00±0.51% final accuracy and 17.22±0.43% forgetting, outperforming iCaRL (42.33±1.20%, 24.87±1.11%). Ablations show accuracy drops when replacing NME/herding and forgetting increases without distillation, while retrieval budget variations have less impact than active budget changes.
class-incremental learningexemplar replayfeature representationherdingdistillation
VALG: An Agentic System for ML Theory Research
VALG introduces an autonomous agentic system for machine learning theory research, combining multi-level verification, adaptive problem formulation, and graph-structured proof development. The system maintains mathematical specifications, checks proof-dependency graphs, and constructs local proofs in dependency order, routing failed attempts to address derivations, proof structures, or theorem formulations. Evaluated on nine subproblems from five COLT 2026 open problems, VALG produced two source-scope matches and seven restricted-method results, special cases, or conditional theorems, demonstrating distinct handling of mathematical variants.
agentic systemproof-dependency graphtheorem formulationadaptive learningml theory
TEMPO: Makespan-Aware Expert-Parallel Load Balancing Across Memory- and Compute-Bound Regimes
TEMPO introduces a makespan-aware dispatcher for expert-parallel (EP) MoE serving, addressing load imbalance across memory- and compute-bound regimes. It models expert time via a max-affine profile capturing HBM weight streaming (memory-bound) and grouped GEMM (compute-bound) behaviors, formalizing dispatch as a fixed-charge makespan problem. TEMPO solves this in milliseconds off the critical path, integrating with SGLang for fused dispatch and count collection. Evaluations on 8-GPU testbeds show TEMPO matches the best fixed baseline within 1% and improves throughput by 4-6% (Qwen3-235B) with 15.6% p99 latency reduction, though gains depend on regime overlap (phase diagram predicts outcomes).
expert-parallelmakespanhbm weight streaminggrouped gemmmax-affine profile
Operationalizing Cyber Threat Intelligence with GraphRAG
This work demonstrates that knowledge-graph retrieval (Microsoft GraphRAG) outperforms vector-similarity retrieval (Naive RAG) in generating durable cyber threat detection rules from reports, by leveraging higher-level indicators resistant to attacker evasion. The study compares both systems using identical reports, generation instructions, and language models, isolating retrieval as the variable. In evaluations across nine CTI reports, GraphRAG maintained 100% detection efficacy after indicator rotation (vs. 29% for Naive RAG), consistently accessing more robust Pyramid of Pain levels. Results validate knowledge-graph retrieval as architecturally superior for automated hunting plan generation, with prompt wording also significantly impacting outcomes.
knowledge-graph retrievalvector-similarity retrievalpyramid of painthreat detectionmicrosoft graphrag
DMDIntel: Interpreting Large Language Models via Dynamic Mode Decomposition
DMDIntel introduces a novel interpretability framework for large language models (LLMs) using dynamic mode decomposition (DMD) in classification tasks. The method decomposes hidden states into prominent modes and ranks input tokens based on their projection values on these modes. Evaluated across three datasets and three LLM families, DMDIntel significantly outperforms state-of-the-art techniques like principal component analysis, integrated gradients, and SHAP in input token attribution.
dynamic mode decompositionlarge language modelsinput attributionhidden statesclassification task
BoardroomAI: Dependency-Aware Human-Steerable Multi-Agent Deliberation through Evolving Decision Graphs
BoardroomAI introduces a dependency-aware framework for human-steerable multi-agent deliberation using evolving decision graphs, enabling persistent human intervention through assumption challenges, constraint modifications, and priority updates. The system comprises (i) a typed decision graph, (ii) an intervention compiler, (iii) dependency-aware propagation for selective recomputation, and (iv) an evaluation framework. In synthetic tests, propagation inspected only 14.59% of nodes while matching exhaustive impact computation, and selective repair preserved all gold-unaffected nodes, achieving valid decisions in 6/12 cases while abstaining in others due to insufficient context.
multi-agent systemsdecision graphsdependency-aware propagationhuman-steerable deliberationintervention compiler
From Local Mismatch to Global Impact: Optimizing Cache Reuse Policy for Efficient Diffusion
Proposes Global-Impact Cache (GCache), a bilevel optimization framework for cache reuse policy in diffusion models that addresses the misalignment between local similarity heuristics and generation quality. GCache reformulates policy search by optimizing reuse in the inner objective while aligning error-weighting with quality loss in the outer objective, using a Bernstein-parameterized error propagation bound. Experiments show GCache achieves a 2.17x speedup on Wan2.1 video diffusion while reducing LPIPS from 0.1095 to 0.0316, outperforming prior caching strategies in both speed and fidelity.
diffusion modelscache reuse policybilevel optimizationerror propagationbernstein parameterization
UniTraffic-Agent: Unified Traffic Video Reasoning for AI City Challenge 2026 Track 3 with Two Out-of-Domain Evaluations
UniTraffic-Agent is proposed as a unified traffic video reasoning system for the AI City Challenge 2026 Track 3, addressing Traffic Anomaly Reasoning (TAR) and two out-of-domain tasks: FETV (fisheye traffic events) and PSI-VQA (pedestrian intention reasoning). The method employs an observe-reason-act-verify workflow, sampling timestamped visual evidence and reasoning over all clip questions in a single request via task-specific action adapters. On official leaderboards, it achieves scores of 0.5780 (TAR, 16th), 0.4884 (FETV, 2nd), and 64.4161 (PSI-VQA, 4th).
traffic video reasoningmultimodal llmsout-of-domain evaluationaction adapterstimestamped evidence
Static analysis-guided agentic AI translation enables Rust as a full stack bioinformatics language
The paper presents an agentic AI method combining static analysis to systematically translate legacy bioinformatics code (e.g., Perl, Fortran) to Rust, reducing technical debt and improving performance. The approach uses specialized prompts and tooling to convert NGS and imaging software, demonstrated on the Bascet pipeline, achieving 80x size reduction, 10x faster builds, and >3x runtime improvements. The resulting Rust implementation eliminates Unix dependencies, enabling native Windows execution without containers. This enables cost-effective large-scale modernization of bioinformatics tools for clinical and high-performance computing environments.
static analysisagentic airustbioinformaticslegacy code
Foundations of MT-PDCL: Measure-Theoretic Probabilistic Definite Clause Logic
MT-PDCL introduces measure-theoretic foundations for probabilistic definite clause logic, enabling exact inference over continuous domains by replacing discrete grounding with Lebesgue integration. The framework defines stochastic variables over bounded index domains with Borel σ-algebras, models probabilistic rules as independent causal events, and employs a continuous immediate consequence operator for algebraic integration of priors and observations. This eliminates combinatorial grounding bottlenecks while preserving declarative syntax, though at the cost of geometric dimensionality challenges.
measure-theoretic logiccontinuous distribution semanticslebesgue integrationborel σ-algebrasprobabilistic definite clauses
OGR-MARL: Option-Guided Residual Multi-Agent Reinforcement Learning for Heterogeneous USV Cooperative Pursuit in Constrained Port Waterways
The paper introduces OGR-MARL, an option-guided residual multi-agent reinforcement learning framework for heterogeneous USV cooperative pursuit in constrained port waterways. OGR-MARL integrates shared evader belief, role-conditioned option targets, adaptive rule penalties, and residual policy learning, enabling MARL algorithms to learn corrective actions atop rule-guided behaviors. Instantiated with MADDPG, MATD3, MAPPO, and MASAC, OGR-MASAC achieves a 75.0% capture rate in an abstract Xiazhimen port-waterway scenario, demonstrating effective rule compliance and heterogeneous coordination. Zero-shot transfer to a QGIS/AIS-informed Xiazhimen map shows generalization potential in complex port scenarios.
multi-agent reinforcement learningoption-guided learningresidual policy learningheterogeneous coordinationzero-shot transfer
Generative Universal Multimodal Retrieval with Dual-role Identifiers
The paper introduces DrIG, a generative framework for universal multimodal retrieval using dual-role identifiers to address challenges in generative information retrieval (GIR). Each candidate's residual-quantized identifier functions sequentially (autoregressive decoding with modality modeling) and as an unordered set (prefix-independent relevance prior), improving robustness. Evaluated on M-BEIR and text-to-image datasets, DrIG outperforms state-of-the-art generative baselines, with hybrid reranking achieving efficiency-effectiveness trade-offs against dense retrievers. Ablations analyze impacts of base LMM, beam size, and fusion strategies.
generative information retrievalresidual-quantized identifiermultimodal retrievalautoregressive decodinghybrid reranking
Beyond Handcrafted Security: Towards Self-Evolving Defense for LLM Agents
This work introduces HARD (Harness-based Autonomous Runtime Defense Evolution), a self-evolving runtime defense framework for securing large language model (LLM) agents. The authors first develop a harness-level formulation that systematically characterizes runtime defense mechanisms and unifies existing interventions. HARD autonomously identifies intervention strategies and iteratively improves defense artifacts using observed failure traces, transforming defense development from manual engineering to an autonomous evolution process. Extensive experiments demonstrate that HARD enhances security performance over handcrafted defenses while preserving benign task utility, establishing autonomous defense evolution as a promising paradigm for LLM agent security.
runtime defensellm agentsharness-level formulationautonomous evolutionsecurity performance
Moose: Latent concept learning with reasoning-shortcut awareness in $\mathcal{EL}^{++}$
Moose introduces a neuro-symbolic method for latent concept learning in OWL 2 EL ontologies, addressing reasoning-shortcut awareness previously unexplored in this setting. The approach compiles an $\mathcal{EL}^{++}$ TBox and finite ABox to a Sentential Decision Diagram (SDD), using it as a differentiable weighted-model-counting layer enhanced with closure clauses for partial supervision. Theoretical guarantees include termination, soundness, completeness, and polynomial intermediate sizes, verified in Lean. Evaluations on MNIST-with-ontology and Pizzaïolo show superior performance over propositional-NeSy, fuzzy-logic, and ontology embedding baselines, while providing the first reasoning-shortcut analysis in OWL EL.
neuro-symbolic learningsentential decision diagramowl 2 elreasoning-shortcut awarenesslatent concept learning
The Objective Is the Bottleneck: Latent World Models Encode What Their Planners Cannot Use
The study demonstrates that planning failures in latent world models stem from suboptimal objectives rather than prediction degradation, challenging the common assumption. Analyzing LeWorldModel on TwoRoom, the authors show that cross-entropy-method planning minimizes squared latent distance, which poorly correlates with true distance (r=0.426) and exhibits pathological behavior (e.g., moving away reduces cost). Ridge probes confirm spatial information remains intact (R²=0.9922). Simply replacing the objective—without retraining—improves goal-reaching from 26.0% to 98.0% at long horizons, matching short-horizon performance. A reachability-focused head outperforms proximity-based metrics despite worse spatial prediction (r=0.819 vs. 0.9897).
latent world modelscross-entropy-methodridge probereachabilityplanning objective
AutoQuREO: A Framework for Automated Quantum Resource Estimation and Optimization
AutoQuREO introduces an automated framework for quantum resource estimation and optimization, addressing limitations of existing compilation-heavy or symbolic approaches. The method combines (i) user-defined quantum stack abstractions, (ii) modular component libraries, (iii) neuro-symbolic surrogate modeling of layer-wise resources, and (iv) embedded multi-objective optimization. Case studies demonstrate systematic discovery of resource trade-offs in early-fault-tolerant algorithms, error correction codes, and variational quantum circuits, outperforming existing QRE tools in tractability.
quantum resource estimationneuro-symbolic learningmulti-objective optimizationfault-tolerant algorithmsdigital twin
Decomposition of Evidence, Contradiction, and Fragility in Perturbation Responses
The paper introduces DECAF (Decomposition of Evidence, Contradiction, and Fragility), a method to interpret perturbation-based model explanations by decomposing response magnitude into three components: evidence (E), contradiction (C), and fragility (F). DECAF tracks how input perturbations progressively alter predictions, preserving total magnitude (Abs = E + C + F) while distinguishing aligned, opposed, and null responses. Experiments on ImageNet-9, FunnyBirds, and ImageNet-1k show DECAF components correlate with independently measured behaviors (96.4% agreement vs. 35.0% for magnitude alone) and outperform baselines in efficiency (4.75x faster wall time, 2.36x lower memory on DINOv2).
perturbation methodsevidence decompositionmodel interpretabilitygradient-based attributionendpoint-relative axioms
FlashDrive: Flash Vision-Language-Action Inference for Autonomous Driving
FlashDrive introduces an algorithm-system co-design framework that accelerates Vision-Language-Action (VLA) inference for autonomous driving by addressing four computational bottlenecks. The method combines streaming KV-cache reuse for visual encoding, non-autoregressive diffusion drafting for low-entropy reasoning tokens, adaptive step caching for velocity field denoising, and CUDA Graph compilation for system optimization. Evaluated on Alpamayo 1.5-10B with W4A8 quantization, FlashDrive reduces latency from 717ms to 151ms (4.7x) while maintaining accuracy (minADE6@6.4s changes by 0.08m) and improving closed-loop performance, enabling 6.6Hz real-time inference on a single GPU.
vision-language-actionkv-cachespeculative decodingadaptive step cachingw4a8 quantization
Polish Medical Visual Question Answering: Vision-Language Models Underutilize Visual Evidence
The authors introduce a Polish-language medical visual question answering (VQA) benchmark derived from physician certification exams, comprising image-based questions and a text-only QA control set. They evaluate Polish-oriented and general-purpose vision-language models, finding the best model achieves 79.0% accuracy on VQA, with only GPT-5.6 surpassing human performance on a response-candidate subset. Analysis reveals models rely more on question text than images, performing worse on image-dominant questions, yet maintain above-chance accuracy even when key components are omitted.
visual question answeringvision-language modelsmedical benchmarkvisual groundingpolish-language
H-VAEP and H-xT: Valuing Offensive On-the-Ball Actions in Handball by Estimating Probabilities
The paper introduces Handball-VAEP (H-VAEP) and Handball-xT (H-xT), the first adaptations of Valuing Actions by Estimating Probabilities and Expected Threat frameworks for handball analytics. Using five seasons of Handball Bundesliga tracking data, H-xT employs a handball-specific court zoning layout, demonstrating superior robustness over rectangular grids via simulations. H-VAEP optimizes feature space and context length to mitigate team-identity leakage, yielding stable, discriminative player ratings that emphasize build-up play. The authors release a complete code repository for practical deployment.
action valuationexpected threattracking datacourt zoningteam-identity leakage
Discovering Efficient and Explainable Communication Topologies for LLM-based Multi-Agent Systems via Causal Inference
The paper introduces E2-Explainer, a model-agnostic framework for interpretable explanation of communication topologies in LLM-based multi-agent systems. It formulates topology explanation as a causal attribution problem, using a Granger-style objective to identify compact, task-preserving subgraphs by measuring edge-level impact on task outcomes and response stability. Experiments on reasoning and coding benchmarks show the method identifies critical subgraphs that maintain performance while reducing communication costs by pruning redundant edges.
multi-agent systemscommunication topologiescausal attributiongranger causalityllm-based collaboration
InFactPlanner: Planning Sustainable Geo-Distributed LLM Data Centers
InFactPlanner introduces a trace-driven framework for sustainable geo-distributed LLM inference data center planning, enabling what-if analysis of energy use, carbon emissions, water consumption, and service quality. The method integrates query traces, hardware-model profiles, site configurations, PUE/WUE metrics, renewable generation models, and grid carbon intensity to estimate key sustainability and performance metrics. Validation shows <10% deviation from reference LLM inference energy estimates, with scenario analyses revealing trade-offs between sustainability-optimal and latency-optimal deployment choices, particularly influenced by local grid mix.
llm inferencegeo-distributed data centerscarbon-aware routingpue/wue metricsrenewable generation models
EGRL: Edge generation-guided relation-aware learning for RNA-protein interaction prediction
The paper proposes Edge Generation-guided Relation-aware Learning (EGRL), a graph neural network framework for RNA-protein interaction prediction that addresses data sparsity and cold-start limitations. EGRL introduces implicit meta-path learning, multi-relation-aware attention, a graph generator for soft edge prediction, and multi-feature fusion, trained jointly with primary and auxiliary losses. Evaluations on four benchmarks show EGRL achieves 0.867 AUROC and 0.861 AUPR on unknown molecules, improving over prior work by 8.6% and 5.0% respectively.
graph neural networksrna-protein interactionmeta-path learningcold-start problemrelation-aware attention
NaviDC-OCR: Navigating Document Parsing Across Digital and Camera-Captured Documents
NaviDC-OCR introduces a unified framework for document parsing that addresses key challenges in Vision-Language Models (VLMs). The method combines deformation-aware learning for geometric perception, adaptive sampling for layout representation, and content-structure decoupled learning for explicit modeling of formula grammars and table structures. Evaluations show state-of-the-art performance with scores of 96.87, 88.53, and 78.41 on OmniDocBench v1.6, Wild-OmniDocBench, and PureDocBench, respectively, and top ranking in the ICDAR 2026 Sci-ImageMiner Challenge.
document parsingvision-language modelsdeformation-aware learningadaptive samplingstructured representation
Agent Behavioral Contracts II: Certifying Compositional Reliability Without Assuming Independence
The paper challenges the conditional-independence assumption in compositional reliability bounds for multi-agent systems, demonstrating strong co-failure correlations (log OR 6.66, phi 0.916) between identical models in a two-agent handoff task. It introduces a finite-sample certification method via a linear program over joint execution moments, proving soundness and sharpness while avoiding bootstrap pitfalls. Results show an 85.7% interval narrowing and certified floor improvement from 0.2455 to 0.4116 with enriched moment functionals. Preregistered evaluations (18,000 missions) reveal signed errors favoring shared-model redundancy overestimation.
compositional reliabilityfinite-sample certificateco-failure correlationlinear programmoment functionals
Predictive Memory Localization: Forecasting Selective Intervention Paths from Internal Signals
The paper introduces Predictive Memory Localization (PML), a method that forecasts selective intervention paths by analyzing internal model signals. PML separates target movement from semantic-neighbor interference and capability damage, using a strength-disjoint low-dose causal response framework. Evaluated across 3,000 records from nine datasets and fourteen domains, PML achieves 13.1% target-any and 12.3% clean-any accuracy at layer 7, outperforming random baselines by 3.6 and 3.4 percentage points. The method demonstrates 0.801-0.828 macro AUROC on held-out records, enabling risk-aware intervention decisions through falsifiable outcome forecasts.
predictive memory localizationactivation steeringstrength-disjoint responsesemantic-neighbor damagelow-dose causal
Labels Are Not Endpoints: Treatment Leakage and Construct Validity in MCP Agent Security Evaluation
The study identifies treatment leakage in security evaluations of tool-using agents, where stored labels inappropriately influence behavioral classifications. By auditing 10,200 execution rows across 180 model-bound requests, the authors reconstruct a treatment-blind evaluation, correcting 58 historical labels while preserving four verified cases. A dual-reviewer concordance review confirmed structural interpretability but revealed discrepancies in four construct-boundary cases. Contributions include a seven-link Integrity Chain and an executable endpoint-integrity linter for scope-bounded measurement audits.
treatment leakageconstruct validitysecurity evaluationtool-using agentsintegrity chain
ReflectFact: Self-Reflective Agents for Improving Comprehension and Reasoning in Multi-Hop Fact Verification
ReflectFact introduces a self-reflective agent framework for multi-hop fact verification, addressing limitations in current multi-agent approaches through three novel tasks: Explicit Reasoning Path Planning for evidence-grounded reasoning, Evidence-Drift Verification to calibrate evidence deviation, and Reasoning Reflection Verification to correct reasoning flaws. The method aggregates validated reasoning chains for reliable verdicts. Experiments on HOVER and EX-FEVER show state-of-the-art performance, outperforming the strongest baseline by 3.32% and 2.78% respectively.
multi-hop fact verificationself-reflective agentsevidence-grounded reasoningparametric knowledgereasoning reflection
SPARED: Reasoning-Based AI-Generated Image Detection via Adversarially Edited Data
The paper introduces SPARED, an adversarial reinforcement learning framework for AI-generated image detection with explainable verdicts. The method pits a diffusion-based image editor against a multimodal LLM (MLLM) detector: the editor adversarially modifies real images to evade detection, while the MLLM learns to classify and justify its decisions via free-form reasoning. This dynamic training loop prevents shortcut learning by enforcing faithfulness in edits and correctness in verdicts. Evaluations show monotonic improvement across three benchmarks, with explanation quality emerging as a side effect of accuracy optimization.
adversarial learningdiffusion modelsmultimodal llmexplainable aiimage forensics
A Compositional Theory of Curvature in Probabilistic Circuits
The paper introduces a compositional theory of curvature for Probabilistic Circuits (PCs), demonstrating that global sharpness regularization is misspecified due to PCs' inherent curvature structure. It proves that each sum node's Hessian trace decomposes into circuit flow (usage weight) and local sharpness, revealing depth bias in global methods. An adaptive sharpness-aware regularizer is proposed, targeting nodes based on local curvature while preserving closed-form EM updates. Empirical results show improved generalization and robustness compared to global regularization.
probabilistic circuitshessian tracesharpness regularizationcircuit flowem updates
AI and Consumer Rights in India Working Paper
The working paper analyzes India's Consumer Protection Act, 2019 for addressing AI-related consumer harms through legal doctrinal analysis. It finds the Act's technology-agnostic definitions of product liability and harm could cover AI incidents like bias or loss of control, but identifies gaps in causation proof and liability allocation. Key challenges include mapping diffuse AI value chain roles (data providers, model developers) to traditional manufacturer/seller categories and establishing technical causation for emergent AI failures.
product liabilitycausation proofvalue chain allocationemergent failurestechnology-agnostic
BrainWAM: Action-Space Coordination of Semantic Priors and Predictive Dynamics for Autonomous Driving
BrainWAM introduces a structured action-space coordination framework for autonomous driving, addressing the attention-allocation mismatch in naive combinations of semantic priors and predictive dynamics. The method employs two specialized action-oriented pathways—semantic reasoning and predictive world modeling—aligned at compact action representations, alongside an asynchronous rectified-flow inference strategy with decoupled video and action denoising to reduce latency. BrainWAM achieves state-of-the-art performance on NAVSIM v1 (89.5 PDMS) and NAVSIM v2 (89.6 EPDMS), outperforming Vision-Language-Action (VLA) and World Action Models (WAM) baselines, demonstrating its efficacy in autonomous driving systems.
action-space coordinationsemantic priorspredictive dynamicsrectified-flow inferenceautonomous driving
Falsehood and Impossibility Are Different Directions in an AI's Representation of Language
The study investigates whether the Gemma 3 4B IT model distinguishes between contingent falsehoods and necessary falsehoods in its internal representations. Using 85 prompts across 17 philosophical families and 15 topics, the authors analyze activations and probe performance. Results show that the model conflates contingent falsehoods with contradictions in its outputs but separates necessary falsehoods from true statements (AUC 0.93) and contingent falsehoods (AUC 1.00) in its activations. A linear truth probe achieves high separation for impossibility (AUC 1.00), peaking at layer 15 with balanced accuracy 0.97. Necessary falsehoods are represented closer to semantic anomalies than contingent falsehoods, suggesting distinct directions in the model's activation space.
activation studylinear truth probesemantic anomalynecessary falsehoodcontingent falsehood
Practice Makes Unsafe: Skill Misevolution in Self-Improving LLM Agents
The paper introduces SkillMisevo-Gym and SkillMisevo-Bench to measure skill misevolution in self-improving LLM agents, where unsafe successes become reusable policies. Skill evolution distills trajectories into executable procedures, but optimization for task outcomes over safety leads to compromised experience. SafeEvolve, a proposed wrapper, reduces unsafe retrieval and fresh-session harm by 26.7 and 17.3 percentage points respectively, with minimal impact on benign utility (0.4 points). Experiments across 25 agent-method configurations show all 21 evolved configurations author unsafe artifacts, with three malicious tasks increasing carryover ASR from 16.0% to 35.3%.
skill misevolutionself-improving llm agentssafeevolveskillmisevo-benchlifecycle metrics
Beyond Retrieval: Query-Conditioned Reuse of Long-Horizon Agent Trajectories
The paper introduces query-conditioned reuse (QCR), a post-retrieval method for reusing long-horizon agent trajectories by encoding reusable procedures, bindings, and verification conditions. The framework evaluates reuse support while controlling for retrieval quality, model, and decoding parameters. On 2,391 instances across WebArena, WorkArena, and AppWorld, QCR achieves 62.3% Success (+10.7 points over Full Trajectory) with 48.9% fewer online tokens, demonstrating robustness to trajectory length and binding shifts.
trajectory memorypost-retrieval reusequery-conditioned reuselong-horizon agentsbinding shift
FSGR: Mitigating Token Frequency Bias for Fair SID-Based Generative Recommendation
The paper introduces FSGR, a fairness optimization framework addressing Token Frequency Bias in Semantic ID (SID)-based generative recommendation systems. FSGR employs Optimal Transport-based Assignment Optimization and Dual-Criteria Re-anchor during SID construction, followed by a two-stage training strategy with Hierarchical Frequency Calibration for layer-specific fairness fine-tuning. Experiments on three datasets with three backbone models show FSGR improves fairness (average Gini coefficient reduction >20%) while maintaining recommendation accuracy.
semantic idtoken frequency biasgenerative recommendationoptimal transportfairness optimization
Heterogeneous Vision-Language Ensemble with Disagreement-Aware Reranking for Text-Based Person Anomaly Retrieval
The paper presents a heterogeneous vision-language ensemble method for text-based person anomaly retrieval, addressing the challenge of fine-grained cross-modal matching. The approach integrates multiple vision-language embedding models through score alignment and iterative ensemble fusion, followed by disagreement-aware VLM reranking for ambiguous queries. Evaluated on the Pedestrian Anomaly Behavior (PAB) benchmark, the method achieves 90.92% mAP, 85.13% Recall@1, 97.72% Recall@5, and 98.68% Recall@10, demonstrating robust performance in retrieving anomalous pedestrian behaviors from text descriptions.
vision-language ensemblecross-modal matchingdisagreement-aware rerankingtext-based retrievalanomaly behavior detection
CABS+: Efficient and Scalable Model Merging via Conflict-Aware Sparsification and Adaptive Weight Allocation
CABS+ improves model merging efficiency and performance through Conflict-Aware Sparsification and Adaptive Weight Allocation (AWA), addressing limitations of prior work like exponential time complexity and task dominance. AWA employs gradient-free search for coefficient optimization, while an asymmetric fitness function balances multi-task performance. Evaluated across 27 datasets and 5 models (LLMs, small-scale language, and vision), CABS+ outperforms AdaMerging and WUDIMerging by 16.97% and 12.93%, respectively, with 4x speedup and 75% less GPU memory usage.
model mergingconflict-aware sparsificationadaptive weight allocationgradient-free searchrelative synergy score
AQuA: Recursively Self-Improving Quantitative Trading Research Agents
The paper introduces AQuA, a dual-system framework for autonomous quantitative trading research that achieves recursive self-improvement through evidence retention. The first system employs a manager-mediated multi-agent pipeline for symbolic factor discovery in crypto markets, achieving a combined information coefficient of 0.190. The second system uses a config-driven loop over a hybrid time-series architecture for US equities, yielding a per-stock information coefficient of +0.0843 and a long/short strategy with Sharpe ratio +2.50. Both systems operate in sealed sandboxes with fixed evaluation protocols.
recursive self-improvementquantitative tradinginformation coefficientmulti-agent pipelinehybrid time-series architecture
From Atomic Evidence to Logical Composition: Structured Compositional Reasoning over Compound Answer Options
The paper introduces a structured compositional reasoning framework that decomposes compound answer options (connected by AND, OR, NEITHER/NOR) into atomic judgments, scores contrastive hypotheses for each atom, and composes predictions via an operator-constrained integer linear program. Evaluated on LOGICAL-COMMONSENSEQA and LOGICAL-SATA (a new reading-comprehension benchmark derived from SATA-Bench), the method improves Macro-F1 from 48.3 to 77.0 and 47.0 to 75.6 respectively, with largest gains on NEITHER/NOR operations.
compositional reasoninginteger linear programmingcontrastive hypotheseslogical operatorsatomic judgments
Fast A/B/n Testing: Exact Multi-Policy Comparison via Tree-Coupled Feedback Sharing
The paper introduces Tree-Coupled A/B Testing (TCAB), an exact feedback-sharing design for comparing multiple contextual-bandit policies with reduced reward queries. TCAB connects policy histories via a predictable tree structure, sharing rewards within matched edges while preserving each policy's standalone trajectory law. Theoretical analysis shows TCAB achieves conditionally optimal query costs, with expected queries scaling as T + o(T) versus JT for independent runs, and provides finite-sample variance bounds for policy contrasts. Experiments on reward-model evaluation, language-model assessment, and adaptive search demonstrate significant cost-precision improvements.
contextual banditsfeedback sharinga/b/n testingpolicy comparisontotal variation
Erase but Preserve: Controllable Removal of Copyrighted Animation Characters via Optimized Semantic Anchors
We propose a controllable method for erasing copyrighted animation characters in text-to-image diffusion models while preserving image fidelity. The approach optimizes a semantic anchor embedding via structural and detailed constraints to serve as a character surrogate, then replaces target-related embeddings using a structure-aware adaptive strategy. Experiments demonstrate state-of-the-art erasure effectiveness, fidelity preservation, and support for controllable erasure degree, multi-target removal, and model transferability. Optimized anchors are plug-and-play with existing model modification baselines, improving their erasure performance.
text-to-image diffusionsemantic anchorsconcept erasuremodel transferabilityadaptive strategy
PIPES: Securing Agent Perception with Provenance and Priors
The paper introduces PIPES (Provenance-Informed, Prior-Enforced Screening), a defense mechanism against state-corruption attacks in tool-using agents, where attacker-controlled content corrupts the agent's perceived environment. PIPES screens response units using semantic priors and source provenance, enforcing static field contracts for schemas and conditioning open-ended content screening on pre-response trajectories and trusted metadata. Evaluated on VitaBench and AgentDyn splits with Gemma 4 31B IT, PIPES reduces attack success from 84.7% to 2.3% while maintaining benign utility (92.5% vs. 90.6%).
state-corruption attacksprovenance metadatasemantic priorstool-using agentsfield contracts
ARAC: Benchmarking Auto-Research's Alignment and Completeness on End-to-End Researchs
The paper introduces ARAC-Bench, a novel evaluation framework for assessing Auto-Research systems' alignment and completeness with human research processes. The framework combines an Academic Cognition Skills system, which quantifies reviewer expertise, and a three-stage diagnostic protocol (Proposal, Experiment, Synthesis) to decompose research trajectories. Evaluation of 11 state-of-the-art frameworks reveals a maximum alignment score of 67.9/100, indicating significant gaps in mimicking human methodology. Validation shows strong correlation (0.8141) with Ph.D. candidate rankings, confirming ARAC-Bench's reliability as a diagnostic and training tool for autonomous research systems.
auto-researchevaluation frameworkacademic cognitionresearch trajectoryalignment score
CRAFT: LLM-Based Iterative Refinement for Temporal Reasoning over Clinical Narratives
CRAFT introduces an LLM framework for reconstructing symptom timelines from clinical narratives with sparse temporal anchors, pairing a generator with a constraint-based verifier for iterative refinement. The method addresses the gap in structured trajectory reconstruction from individual reports, evaluated on MedTempo, a new benchmark of 5,347 vaccine adverse-event narratives with expert-validated annotations. Experiments across four LLM backbones show consistent improvement in temporal ordering accuracy, with ablation studies quantifying component contributions.
temporal reasoningclinical narrativesllm frameworkiterative refinementsymptom trajectories
Memorization Diagnostics for Code LLMs Should be Scale-Aware
The article proposes scale-aware diagnostics for evaluating memorization in code LLMs, arguing that traditional probing techniques (synonym fuzzing, dead-code insertion) fail at scale due to conflating representation load with memorization. Using invertible mathematical transforms on numeric problems, the authors isolate these factors, showing scaled encoders absorb representation load while converging on correct solution families. Results suggest memorization's impact on usability is overstated when representation load is accounted for, urging future evaluations to disentangle these phenomena.
memorization diagnosticsrepresentation loadcode llmsinvertible transformsscale-aware evaluation
Beyond Outcome Rewards: Step-Level Self-Distilled Policy Optimization for Deep Search Agents
The paper introduces Step-Level Self-Distilled Policy Optimization (SSPO), a method addressing credit assignment in deep search agents by leveraging step-level self-distillation. SSPO uses Evidence Anchors—concise web-extracted reasoning snippets—as privileged information to mitigate teacher-student information asymmetry. It converts teacher-student disagreement into step-level advantage weights within GRPO, applied only to incorrect trajectories while preserving correct ones. Evaluated on Qwen3-8B, SSPO outperforms GRPO across BrowseComp, GAIA, and FRAMES, matching GRPO trained with twice the gradient steps at only 5% overhead per step.
self-distillationcredit assignmentsearch agentsadvantage weightspolicy optimization
PROVE-RT: Generating Mechanized Theorem Prover Scripts for Real-Time Systems using LLMs
PROVE-RT introduces an LLM-assisted framework for generating mechanized PROSA/ROCQ theorem prover scripts for real-time systems schedulability analysis. The method combines dependency-aware informal sketches, retrieval-augmented generation from processed PROSA documentation (13,134 sketches from 1,191 papers), and staged skeleton generation with proof completion. Compared to direct LLM prompting, PROVE-RT improves success rates from near-zero to 44.7% on a curated evaluation set, demonstrating effective retrieval-guided staged generation for domain-specific theorem proving.
mechanized verificationschedulability analysisretrieval-augmented generationtheorem prover scriptsreal-time systems
Correct Is Not Governed: Provenance Integrity in Agentic Workflows
The paper introduces governed execution as a framework ensuring agentic workflows maintain inspectable provenance for decisions, completion, and change response, beyond mere correctness. It presents Matrix, a deterministic causal-state layer that tracks authority and fact dependencies, verifies completion evidence, and selectively invalidates dependent work. Controlled comparisons showed governed workflows preserved evidence and refused unsupported closure, though a transfer challenge revealed over-blocking in synthetic contexts. Matrix primarily serves as an institutional integrity layer for auditability, not general accuracy enhancement.
governed executionagentic workflowsprovenance integritydeterministic causal-statecompletion evidence
SynAct: A Reasoning-Acting Large Language Model Agent for Adaptive Synthesis Optimization
SynAct introduces an adaptive closed-loop LLM reasoning--acting agent for logic synthesis optimization, addressing limitations of automated black-box search and static LLM-generated scripts. The method iteratively diagnoses live synthesis reports, leveraging current circuit state, tool knowledge, and historical optimization to issue targeted commands, focusing on worst negative slack (WNS) improvement while balancing area and power trade-offs. Evaluated on a commercial synthesis tool across 14 designs, SynAct reduces average WNS to 27% of bootstrap synthesis results.
logic synthesisworst negative slackclosed-loop reasoningadaptive optimizationgate-level netlists
PatientAct: Theory-Grounded Mental Health Client Simulation
PatientAct introduces a theory-grounded framework for LLM-based mental health client simulation, addressing limitations of current overly cooperative simulators. The method integrates 5Ps clinical case formulation for causal depth and implements dynamic memory with trust thresholds, modeling emotional reactions and resistance behaviors before response generation. Evaluations on 40 clinical situations show significant improvements in resistance quality (p<0.05) and behavioral realism compared to baselines, while maintaining high clinical plausibility across diverse profiles.
llm-based simulation5ps formulationdynamic memorytrust thresholdsbehavioral realism
Spatial Memory Agent: Experience-Grounded Procedure Memory for Spatial Intelligence
The paper introduces Spatial Memory Agent (SMA), a parameter-update-free framework that enhances spatial reasoning in frozen vision-language models (VLMs) through experience-grounded memory. SMA distills verified spatial experiences into transferable lessons with Transfer Reliability Scores (TRS), using verifier-guided reflection and retrieval-based inference. Evaluated across five spatial benchmarks and four base VLMs, SMA achieves the highest macro average accuracy in all model blocks and outperforms other methods in 20 evaluations, demonstrating scalable spatial self-evolution without external tools or parameter updates.
spatial memory agentparameter-update-freetransfer reliability scoreverifier-guided reflectionfrozen vlm
ERSkill: Evolving for Skill-Guided Adaptive Memory Retrieval
ERSkill introduces a retrieval-centric framework for self-evolving, skill-guided memory access in LLM agents, addressing limitations of static retrieval mechanisms. The method compiles interaction histories into structured memory, represents retrieval behaviors as executable skills, and employs a trained router to dynamically match queries to optimal skills. A double-frontier mechanism enables continuous co-evolution of skills and the router. Experiments show ERSkill outperforms baselines, improving average scores by 31.3% with Qwen3-Next-80B-A3B-Instruct and 28.1% with GPT-5.4-nano across multiple benchmarks.
llm agentsmemory retrievalskill-guideddouble-frontierexperience trie
Error-Aware Reverse Auction Mechanism for Large Language Model Routing
The paper proposes EA-RAM, an error-aware reverse auction mechanism for LLM routing that addresses noisy performance predictions from providers and evaluations from the task center. By modeling this Dual Error explicitly, EA-RAM achieves Bayesian incentive compatibility and individual rationality while bounding welfare loss. Theoretical analysis reveals robustness effects from error cancellation, vanishing-tail link functions, and noise-induced smoothing. Experiments demonstrate EA-RAM's superior cost-performance Pareto frontier compared to centralized baselines, particularly when providers contribute local information.
llm routingreverse auctionbayesian incentive compatibilitydual errorwelfare loss
HybridSB-MoE: Dual-Domain Schrödinger Bridges with Scene-Adaptive Expert Routing for Speech Enhancement
HybridSB-MoE introduces a dual-domain framework for generative speech enhancement, addressing limitations in spectral models, waveform models, and Schrödinger Bridges. The method employs asymmetric uncertainty fusion, combining epistemic uncertainty from spectral paths with aleatoric variance from waveform bridges. It utilizes a heterogeneous Mixture of Experts (MoE) with top-k=2 routing across five architectural archetypes to enhance epistemic signal diversity. A discretization bound ensures K-step bridge sampling error is minimized in 2-Wasserstein distance. Evaluated on VoiceBank+DEMAND, HybridSB-MoE outperforms diffusion- and SB-based baselines within their step budgets and remains competitive with consistency-distilled few-step methods.
schrödinger bridgesmixture of expertsepistemic uncertaintyaleatoric variancewasserstein distance
Tracing Provenance and Detecting Tampering with Complementary LLM Watermarks
The authors propose a dual-watermarking method for LLM-generated text that simultaneously traces provenance and detects tampering. Their approach co-embeds robust and fragile signals per token using independent keys and different seeding windows over normalized text, with tournament reweighting preserving generation quality. The signals' scores enable three-state detection (Intact/Tampered/No-Watermark). Evaluated across two LLMs and prompt datasets, the method achieves superior tamper-detection rates while maintaining attribution robustness and perplexity comparable to existing techniques. Ablations confirm the necessity of co-embedded signals with complementary edit sensitivity for reliable detection.
llm watermarkingprovenance tracingtamper detectiontournament reweightingpiggyback spoofing
Mr3D-VL: A generalist vision language foundation model for Multiparametric 3D Magnetic Resonance Imaging
Mr3D-VL introduces a 4B-parameter vision-language foundation model for multiparametric 3D MRI, addressing modality-spatial integration challenges in clinical brain tumor analysis. It combines an unsupervised pre-trained 3D encoder with 4D rotational positional embeddings and a multi-resolution cross-modal projection layer for dual modality-spatial feature fusion. Evaluations demonstrate superior performance over domain-specific and general-purpose models (4B-30B), achieving 0.856 BERTScore in report generation, 0.713 QA accuracy, and 0.912 multiple-choice accuracy.
multiparametric mrivision-language model3d volumetric encodingcross-modal projectionrotational positional embedding
Demand Transfer Estimation at Scale via Restricted Logit Modeling
The authors propose a scalable method for estimating Demand Transfer (DT) coefficients, which quantify demand redistribution when items are removed from retail assortments. Their approach combines independent item demand forecasting with adjustments for inter-item relationships, enabling efficient computation for large item universes (1M+ items). The method leverages restricted logit modeling under reasonable substitution behavior assumptions. Experiments on historical transaction data across multiple locations demonstrate accurate DT coefficient estimation and improved demand forecasting performance.
demand transfer coefficientsassortment optimizationlogit modelingdemand forecastingsubstitution behavior
Beyond the Best Guess: Improving LLM Solution Coverage with Evolution Strategies
The paper demonstrates that Evolution Strategies (ES) outperform Reinforcement Learning (RL) for post-training Large Language Models (LLMs) in discovery domains requiring diverse solution coverage. ES, a gradient-free method optimizing via weight-space perturbations, maintains broader output distributions compared to RL's reward-driven narrowing. Evaluated through pass@k metrics on math benchmarks, ES achieves higher solution diversity and improved performance by preserving exploration capacity. This makes ES preferable for tasks where multiple valid solutions exist.
evolution strategieslarge language modelssolution coveragepass@kpost-training
The Role of Natural Language Understanding in Multimodal Video-Based Dengue Diagnosis
A YOLO- and CLIP-based vision-language framework achieves 98.54% frame-level accuracy in classifying Dengue virus (DENV2)-infected mosquito flight behaviors from video data. The method isolates mosquito regions using YOLO, aligns visual features with biologically meaningful textual prompts via supervised bidirectional contrastive learning, and classifies frames via image-text similarity. Temporal aggregation yields perfect video-level performance. Ablation studies confirm CLIP-based representations and fine-tuning are critical, while text provides semantic alignment without accuracy gains over vision-only models.
yolocontrastive language-image pre-trainingmultimodal learningdengue diagnosistemporal aggregation
Privacy-Preserving RAG by Concealing Sensitive Information from External LLMs
The paper introduces Sensitive Entity Alias Generator (SEAG), a privacy-preserving framework for Retrieval-Augmented Generation (RAG) that prevents sensitive information leakage to external LLMs. SEAG employs a lightweight model to detect sensitive entities, generate aliases, and construct an entity replacement table for obfuscating queries and retrieved documents before external processing. Evaluated on custom datasets, SEAG achieved >80% accuracy in preserving utility while hiding sensitive entities, with Qwen-3 (77.83%), LLaMA-3.2 (76.73%), and Phi-4 (74.91%) showing strong performance in full-entity concealment.
retrieval-augmented generationprivacy-preservingentity obfuscationsensitive informationalias generation
Lines and Ladders: A Context-Aware Multi-Agent Framework for Large-Scale Retail Price Taxonomy
The paper presents a context-aware Multi-Agent Framework for automated construction of retail price taxonomies ('Lines and Ladders') to maintain pricing consistency across large catalogs. The system employs three specialized LLM agents for attribute identification, multi-modal value extraction, and hierarchical grouping, addressing cognitive overload in single-agent approaches. Evaluation on enterprise data shows F1-score of 0.83 for Lines, with >90% precision in Food & Consumables and 80.2% assignment accuracy in General Merchandise.
multi-agent frameworkprice taxonomyllm agentshierarchical groupingcognitive overload
On the Expressive Power of Transformers
The article synthesizes theoretical results characterizing transformer expressive power through circuit complexity frameworks. By parameterizing transformers via attention mechanisms and numerical precision, it establishes formal comparisons with circuit classes defined by gate types, size, and depth. Key findings demonstrate how transformer resource constraints map to circuit complexity hierarchies, providing precise computational equivalences between architectural variants and classical circuit models.
transformerscircuit complexityexpressive powerattention mechanismscomputational equivalences
Designing AI Pipelines for Decision-Ready ITSM Intelligence
The paper contributes an AI pipeline for transforming ITSM ticket data into decision-support artifacts through schema normalization and hierarchical clustering. The method employs LLM-based schema normalization, HDBSCAN sub-topic clustering, and hierarchical agglomerative clustering to produce executive-facing Main-topics and Sub-topics. Evaluation by six artifacts and five raters shows decision-support metrics (interpretability, actionability, trust, likelihood of use) averaging >4.0/5.0, with trust as the most consistent signal.
itsmllmhdbscanschema normalizationdecision-support
General Probabilities of Causation with Causal Knowledge
The paper derives tighter bounds for multivalued probabilities of causation (PoCs) by incorporating causal knowledge from covariates and mediators, extending prior work on binary PoCs by Tian and Pearl and subsequent multivalued generalizations. Using theoretical analysis, the authors show that their bounds are sharper than existing nonbinary bounds, with illustrative toy examples and simulation studies validating the improvements. The results demonstrate that additional causal information can enhance partial identification in multivalued settings.
probabilities of causationpartial identificationcausal boundsmultivalued treatmentmediators
SteerBench-Work: A Benchmark for Agent Steering at Action Boundaries
The paper introduces SteerBench-Work, a benchmark for evaluating LLM agents' steering decisions at action boundaries in workplace scenarios. The benchmark comprises 106 incident-anchored scenarios across seven domains, with balanced proceed/hold labels and evidence-reversed mirrors. Evaluations of 30 model conditions reveal asymmetric failure modes: 28.1% false holds on authorized actions versus 1.0% false proceeds on unsafe actions. Performance drops significantly on risk-resolved commits (63.8% accuracy) compared to original incidents (98.5%), demonstrating that general capability doesn't guarantee proper steering calibration.
llm agentsaction boundariessteering decisionsworkplace automationbenchmark evaluation
Jagged Judges: Epistemic Stability Under Silence, Pressure, and Persistence
The paper introduces the Wiggle Framework, a unified stress test for epistemic stability in LLM judges, decomposing robustness into Mechanical Consistency, Single-turn Conviction, and Multi-turn Persistence. The framework evaluates 9 frontier models across 14 judging tasks, revealing substantial verdict instability (25-71% under static pushback, 62-91% with adversarial persuasion) and net-corrupting effects of successful pressure. Baseline jury majority strength emerges as the most effective predictor of instability, providing the first cross-dataset comparison of judge robustness metrics.
llm judgesepistemic stabilitywiggle frameworkadversarial persuasionverdict instability
Interpretable Causal Discovery via Causal-Effect Constraints
The paper introduces a Bayesian approach for interpretable causal discovery under causal-effect constraints, framing it as a rare-event estimation problem. The method adapts particle-based techniques to infer the joint posterior over causal graphs and parameters conditioned on events with low posterior mass, such as large causal effects. Experiments on synthetic graphs demonstrate accuracy at varying scales, and a case study on the Sachs protein dataset illustrates utility in generating pathway-level scientific explanations.
bayesian inferencecausal discoveryrare-event estimationgraph-parameter spacesachs dataset
Novels generated by language models show compressed formal variation
The study examines formal variation in AI-generated novels compared to human-written corpora, analyzing stylistic diversity across repeated generations. Using GPT-5.5 Thinking and Qwen3-14B, researchers generated 20 novels per model in 19th-century British realist and contemporary 'Zero-Style,' comparing them to 205 human 19th-century novels and 65 contemporary human Zero-Style works. Document-level metrics (MATTR-500, Shannon entropy, sentence length, readability, punctuation rate) revealed compressed sentence structure variance in AI outputs, with less inter-novel diversity than human works. Lexical measures also showed compression, except Qwen Zero-Style MATTR. The findings distinguish variance overclosure (limited formal range) from correlational overclosure (inconsistent cross-measure relationships), showing AI novels individually mimic human style but collectively occupy a narrower stylistic space.
formal variationvariance overclosurecorrelational overclosurestylistic compressionlexical diversity
EgoCITE: Context-Augmented Indexing and Time-Aware Retrieval for Long-Horizon Egocentric Memory
EgoCITE introduces context-augmented indexing and time-aware retrieval for long-horizon egocentric memory systems, addressing unreliable context-poor captions and temporally naive retrieval. The framework comprises EgoScheme (multimodal context enrichment for memory indices), EgoIndex (multi-granularity multi-view indexing), and EgoRetrv (semantic search with temporal relevance scoring). Evaluated on EgoLifeQA, EgoMem, and EgoR1-Bench, EgoCITE improves answer accuracy by 4.4--14.2% over baselines while reducing cost by 36× compared to long-context LLM agents.
egocentric memorymultimodal indexingtemporal retrievalagentic searchlong-horizon qa
LLMs Are Not Good Strategists, Yet Memory-Enhanced Agency Boosts Reasoning
The paper introduces EpicStar, a memory-enhanced framework for improving long-horizon strategic reasoning in LLMs by mitigating strategic drift. The method combines a bank of successful past episodes as heuristics with a working memory for short-term context, using dynamic gating to decide between retrieved actions and new reasoning. Evaluated on StarCraft II, EpicStar outperforms baselines with higher win rates (exact % unspecified) and 10x fewer tokens consumed, demonstrating robustness across difficulty levels and opponent strategies.
strategic reasoninglong-horizon environmentsmemory-enhanced agencydynamic gatingepisodic memory
@skills: Attention is all you have
The paper introduces @skills, an open protocol addressing the inefficiency of traditional skill installation in AI agents, where limited prompt residency slots constrain usability. By decoupling content, persistence, and automatic triggering, @skills enables direct skill access via paths without installation, reducing prompt overhead to a single .gitignore-style line. The protocol supports Git-tracked adaptation, directory-based bundling, and requires no manifests or lockfiles. Implemented in the AdaL CLI and paired with a searchable hub (atskills.one), it facilitates corpus-wide discovery while maintaining compatibility with local and GitHub paths. The approach reduces installation dependencies and expands skill accessibility.
skill protocolprompt residencygit-tracked adaptationdirectory bundlingcorpus-wide search
PseudoMapLabeler: Confidence-Aware Pseudo-Label Generation for Semi-Supervised Online Mapping
We propose PseudoMapLabeler, a teacher-student semi-supervised learning framework for online HD map construction that addresses labeled data scarcity through confidence-aware pseudo-label generation. The method trains a teacher model on limited labeled data, then refines predictions on unlabeled data using Beta-distribution-based confidence maps and a spatial clipping technique that preserves high-confidence regions. Enhanced predictions serve as pseudo-labels for training a student model, followed by fine-tuning on original labeled data. Experiments on nuScenes show +6.1 mAP improvement under low-label regimes compared to supervised-only baselines, demonstrating effective generalization in diverse environments.
semi-supervised learningpseudo-label generationconfidence mapsspatial clippinghd map construction
Dead text or binding clause? Measuring and restoring constraint influence in black-box LLM dialogues
The paper introduces \sysname{}, a method to measure and mitigate behavioral relapse in black-box LLM dialogues, where models persist in enforcing revoked constraints. The approach employs a contract ledger with executable checkers and tombstones to track constraint states, a sequential ablation probe to measure per-clause adherence, and a repair ladder for budgeted interventions. Results on \dataname{} (\NTasks{} HumanEval tasks) show relapse increases with constraint load (from \ScaleDelayedMTwo{} to \ScaleDelayedMEight{} in an 8B model), while ahead-of-time compilation reduces relapse by \RestoreDiff{} (95\% CI \RestoreDiffCI{}). The probe predicts relapse (AUROC \AurocPrimary{}), and tombstone notes partially restore compliance. Overhead is \CostDeliveryFactor{} per delivery and \CostTotalHedged{} in API compute.
behavioral relapsecontract ledgersequential ablation probetombstone noteblack-box llm
Predicting When Random Low-Dimensional Reparameterizations Train Neural Networks
The paper introduces a theoretical framework and practical method for predicting the minimum latent dimension required for random low-dimensional reparameterizations to train neural networks effectively. It derives an orientation-resolved quadratic master formula that incorporates curvature spectrum and displacement profiles, enabling precise predictions of training transitions. The authors propose Random Mapping Networks (RaMaN), which use structured Hadamard or seed-regenerated Gaussian maps to reduce memory overhead from O(P) to O(d) and avoid O(dP) storage costs. Experiments on quadratic and neural-curvature benchmarks demonstrate that the orientation-resolved predictor accurately tracks transition locations and outperforms orientation-agnostic approximations, with end-to-end training transitions observed across image and language models.
random reparameterizationlatent dimensioncurvature spectrumhadamard mapsgaussian-width
What Makes a Peer? Valuation-Anchored Similarity in Private Markets
The paper proposes an ensemble tree-based supervised similarity learning framework for identifying economically meaningful peer companies in private markets, using valuation-driven similarity rather than static feature matching. The method trains a CatBoost gradient-boosted decision tree model on private company valuations, deriving a similarity metric from importance-weighted leaf-node co-occurrences to capture shared valuation drivers while handling nonlinear relationships and missing data. Evaluated on a dataset of 270,000 companies (53,000 with valuations), the framework outperforms traditional distance-based and text-embedding approaches in k-nearest-neighbor valuation tasks across multiple industries and geographies.
supervised similarity learninggradient-boosted decision treevaluation-aware metricnonlinear relationshipsprivate markets
DiG-bench: Discovery in Games
The authors introduce DiG-bench, a novel benchmark for evaluating AI agents' capacity for discovery through experimentation in controlled environments with unknown objectives. The benchmark comprises 70 independent games, each encoded as short strings with unique transformation rules that must be inferred via interaction. Games are tiered by difficulty, with the easiest solvable by existing models and the hardest challenging state-of-the-art agentic systems. All games were solved by at least one human on first attempt, while 21 are publicly released and 49 held privately for secure evaluation.
discovery benchmarkagentic experimentationrule inferencecontrolled environmentstiered difficulty
Auditable agentic AI for evidence-grounded thyroid ultrasound diagnosis and reporting
ThyroidXAgent introduces an auditable, clinician-interactive agentic AI system for thyroid ultrasound diagnosis, integrating lesion localization, risk stratification, and evidence-grounded reporting. Developed using OpenThyroidDB (300k images, 24k reports) and evaluated on 28,458 test cases, it achieves 87.21% Dice score for nodule segmentation, 0.9466 AUROC for benign-malignant classification, and 0.864/0.805 AUROC for metastasis/follicular-papillary carcinoma prediction. Evidence-grounded report generation outperformed multimodal LLM baselines, improving physician accuracy (86.2% vs 70.3% consistency) and reducing segmentation/reporting time by 35.9%/27.4%.
agentic aithyroid ultrasoundevidence-grounded reportingnodule segmentationclinical semantic metric
Reasoning Jury: Multi-Model Consensus for Evaluating Reasoning Traces
The paper introduces Reasoning Jury, a multi-model consensus system for evaluating reasoning traces in LLMs, addressing limitations of single-model judges. The method employs a jury of open-weight models (e.g., GPT-OSS-120B) and a moderated deliberation mechanism to identify reasoning defects and their severity. Results show Reasoning Jury outperforms frontier models (Opus-4.6, Sonnet-4.6, Gemini-3.1-Pro) in defect identification accuracy while reducing costs to 8-15% of frontier model usage, enabling deeper analysis of reasoning failures.
reasoning tracesmulti-model consensusreasoning defectsopen-weight modelsdeliberation mechanism
Not All Nudges Land: Behavioral Controllability and Elaboration Quality in AI-Supported Journaling
This study investigates behavioral controllability as a key factor in AI-supported journaling efficacy, analyzing 369 journal entries from an eight-week passive sensing study. An LLM classified entries by behavioral intention, while sensor data tracked follow-through across 26 features via 3-day before/after comparisons. Results show behaviors involving others improved in only 15-22% of cases, whereas self-controllable behaviors improved more frequently (50-63%). Textual features provided limited predictive power, with signal emerging primarily for specific behaviors like text messaging and longer, personal intention entries. These exploratory findings suggest AI journaling nudges are most effective for individually controllable behaviors.
behavioral controllabilitypassive sensingllm classificationjournaling efficacysensor features
Trie Automata for Constrained Decoding over Large Finite Sets
The paper introduces trie automata, a specialized mechanism for constrained decoding in large language models when selecting from finite sets of valid strings. The method leverages Aho-Corasick multi-pattern matching to precompute per-node token masks, exploiting structural properties like shared prefixes and bounded depth. Results show 7X faster per-step valid-token computation (0.65 μs vs. 5.8 μs) compared to XGrammar, 2--6.5X faster compilation for K ≥ 300, and 29X higher throughput (219 req/s vs. 7.5 req/s) in batch serving, while maintaining sub-100ms compilation up to K = 10,000 and 100% output validity.
constrained decodingtrie automatonaho-corasicktoken maskingbatch serving
CAS: A Causal Attribution Score for Local and Global Explainable Artificial Intelligence
The paper introduces the Causal Attribution Score (CAS), a novel architecture for causal explanation that bridges predictive and interventional attribution. CAS employs causal Shapley contributions within an identified interventional coalition game, producing local and global effect summaries with explicit intervention targets. Evaluations on synthetic data (n=2,200 per simulation) showed coalition-aware CAS achieved MAE 0.107 versus 0.173-0.213 for baselines, with greater advantages (Δ=0.091) under strong interactions. Empirical tests on DoubleML datasets (n=9,915 and n=5,099) demonstrated material divergences between predictive SHAP rankings and CAS-based treatment-effect modifier rankings.
causal attribution scoreshapley contributionsinterventional coalition gametreatment-effect modifiersdoubleml
SchemaLink: An Intelligent Web Editor for LinkML Schema Curation
SchemaLink introduces a web-based graphical editor for constructing and enhancing LinkML schemas in biomedical contexts, addressing challenges faced by novice curators. The system combines a visual language for schema specification with a RAG-based AI assistant to guide schema creation and editing. Experimental analyses demonstrate the tool's effectiveness in producing quality LinkML schemas, particularly for non-expert users. The implementation is available as open-source software with both web and API components.
linkmlschema curationrag-based assistancebiomedical dataweb editor
$\varepsilon$-MemEvo: Adaptive Cross-Task Memory Transfer for LLM Program Evolution
The paper introduces $\varepsilon$-MemEvo, a framework for cross-task knowledge transfer in LLM-based program evolution. It stores prior experience as task-agnostic tactic memories (natural-language strategy summaries) and uses an adaptive injection gate to prevent negative transfer. Evaluated on 8 optimization benchmarks with GPT-5, $\varepsilon$-MemEvo improves AUCC by +8.7% and early-stage convergence by +9.4% over AdaEvolve, with <1% overhead. Ablations confirm adaptive gating's safety and interpretability.
llm program evolutioncross-task transfertactic memoriesadaptive injection gatenegative transfer
Governed Persistent Memory: Source-Bound State Semantics and Fail-Closed Release for Long-Horizon Agents
The paper introduces Governed Persistent Memory (GPM), a bitemporal state-transition model for long-horizon agents that enforces source-bound semantics and fail-closed release policies. GPM incorporates five executable clauses for ledger integrity, conflict isolation, and exact claim closure, evaluated on a 3,600-case benchmark (GPM-ReleaseBench) where it achieves perfect outcome matching. In a sealed service evaluation, GPM's governed lane achieves 100% correctness (2,400/2,400 clusters) versus 25% for ungoverned Qwen2.5-7B, with no regressions and deterministic outputs.
bitemporalstate-transitionfail-closedledger integrityconflict isolation
Personalized Scorer Modeling: A Learning-Based Framework for Deriving Robust Sleep Stage Labels from Multiple Experts
The study introduces a learning-based hypnogram (LBH) framework to derive robust sleep stage labels by aggregating multiple experts' annotations, addressing inter-scorer variability in sleep staging. Using EEG and EMG data from DOD-H and DOD-O datasets, LBH models scorer-specific behavior via confusion matrices from random forest, SVM, and MLP classifiers, then aggregates probabilities to assign final labels. LBH outperformed dataset hypnogram (DH) and best-scorer hypnogram (BSH), achieving 86.07% accuracy (DOD-H) and 86.04% accuracy (DOD-O) with random forest and EEG+EMG, demonstrating improved precision and F1-scores. Results suggest personalized scorer modeling enhances reference hypnogram reliability without discarding individual expert information.
sleep stage classificationlearning-based hypnograminter-scorer variabilityconfusion matriceseeg+emg
Information Abundance Paradox: Long-Context Training Undermines Parametric Knowledge
The paper introduces the Information Abundance Paradox, demonstrating that increasing context window length in large language models initially improves performance but eventually undermines parametric knowledge encoding. Through pretraining and supervised fine-tuning experiments, the authors show that models trained with longer contexts shift from internalizing information to relying on contextual retrieval, reducing robustness when context is absent or misleading. Mechanistic analysis reveals this shift correlates with gradient pressure moving from feed-forward networks to attention modules, supporting the paradox that abundant context can disincentivize parametric learning.
parametric knowledgecontext windowfeed-forward networksattention modulesgradient pressure
SSPO: Structure-Aware Similarity-Weighted Preference Optimization for Neural Combinatorial Optimization
The paper introduces SSPO (Structure-Aware Similarity-Weighted Preference Optimization), a novel method for neural combinatorial optimization (NCO) that addresses gradient signal polarization and baseline redundancy in existing approaches. SSPO jointly scores all sampled solutions using a dissimilarity-weighted leave-one-out baseline, where structurally distinct peers receive higher weights. The method leverages problem-adaptive solution embeddings from the encoder's node representations without additional parameters. Experiments on TSP, EFL, and JSP benchmarks demonstrate consistent improvements over prior baselines, with the EFL policy successfully deployed in JD.com's facility-location system.
neural combinatorial optimizationpreference optimizationgradient signal polarizationstructure-aware weightingleave-one-out baseline
Dual Spatial-Temporal Attribution: Architecture-Aligned Post-Hoc Explainability for Recurrent Graph Anomaly Detection
The paper introduces X-AddGraph, a post-hoc explainability framework for AddGraph (GCN+GRU) anomaly detection in dynamic graphs, addressing its previous opacity. The method employs Dual Spatial-Temporal Attribution (DSTA) with three components: gradient-based spatial attribution, attention-based short-term temporal attribution, and gradient rollback for long-term temporal attribution, all architecture-aligned. The framework preserves detection performance (ΔAUC = 0) while providing explanations, achieving 0.8705 AUC on UCI Message. Long-term attribution identifies historically significant snapshots (0.127 vs. 0.074 counterfactual signal).
post-hoc explainabilitydynamic graphsgradient attributionrecurrent anomaly detectionattention weights
Specification-first convergence with an AI coding agent: a case study of dismantling a core architectural invariant across 189 files in a 717k-line codebase with no test oracle and no human code review
The study demonstrates successful large-scale architectural refactoring by an AI coding agent under a specification-first protocol without human review or test oracle. The agent dismantled a core lifetime invariant in a 717k-line TypeScript codebase (189 files modified) through formal specification, 31 audit cycles correcting 201 defects, and empirical convergence (two zero-finding verification passes). The refactoring preserved target behavior (streaming generation surviving panel closure) with no observed bugs, completed in three days at USD 2,430 cost. Full specification and session logs (1,500+ pages) were published for reproducibility.
architectural refactoringspecification-first protocollifetime invariantempirical convergencetypescript codebase
Defensive Boosting for Online Probabilistic Forecasting
The authors present Defensive Booster, an online probabilistic forecasting algorithm for binary outcomes against adaptive adversaries that simultaneously achieves two previously incompatible guarantees. The method combines online gradient boosting's Brier score competitiveness with the span of a weak hypothesis class $H$ and online weak-to-strong boosting's classification error convergence under weak-learning conditions, via a dual-view approach that constructs hard-core certificates when weak-learning fails. Experiments show superior predictive performance and runtime efficiency (orders-of-magnitude faster) compared to prior ensemble-based methods on synthetic and real data streams.
online boostingprobabilistic forecastingbrier scoreweak-learning conditionadaptive adversary
Exponential Convex Calibration Dimension for the Multi-Label Jaccard Measure
The paper establishes fundamental properties of the multi-label Jaccard measure's calibration dimension, proving that exact calibration requires exponentially many prediction coordinates (between $2^{s-1}$ and $2^s-1$ for $s$ labels). Using MinHash Gram representations and Boolean Möbius inversion, it shows the Jaccard score's loss matrix is nonsingular with affine dimension $2^s-1$. Two polynomial-dimensional approximation methods are provided: an $(s^2+1)$-dimensional $F_1$ surrogate with asymptotic regret $3-2\sqrt{2}$, and MinHash-based surrogates achieving fixed additive regret with dimensions $O(s^2/α^2)$ or $O(s/α^2)$.
jaccard measuremulti-label classificationcalibration dimensionminhashconvex surrogate
DARTree: Speculative Diffusion Decoding with Autoregressive Draft Trees
DARTree introduces a training-free speculative decoding method that extends autoregressive correction from chains to trees, accelerating pretrained language models without loss. The approach constructs a fixed-width candidate tree by expanding and scoring all nodes at each depth in a single batch, then applies best-first pruning to select the verification tree, decoupling autoregressive-head inference from sequential heap operations. Evaluated across seven math, code, and chat benchmarks, DARTree achieves up to 12.97 tokens per verification round, a 98.6% improvement over DFlash and 27.9% over Domino, with a maximum lossless speedup of 9.73× compared to autoregressive decoding.
speculative decodingautoregressive correctiondiffusion-based draftersbest-first pruningverification tree
Exponential quantum advantage for learning signals with a single qubit
The authors demonstrate exponential quantum advantage in classical signal learning tasks using a single controllable qubit coupled to a conventional sensor. They introduce Quantum Phase-Space Inference (Q$Ψ$), a theoretical framework that derives optimal quantum-enhanced algorithms and certifies quantum advantages beyond quantum Fisher information regimes. Experimental results with a superconducting cavity--qubit architecture show $10^7$-fold measurement reductions for Fourier-amplitude and time-varying signal learning, with applications in dark matter detection and wireless communications.
quantum advantagequantum phase-space inferencesingle-qubit sensingfourier coefficient learningweak-signal detection
Intervention-Aware Clinical World Model for Post-Op Outcome Forecasting in Cardiology
The authors propose an intervention-aware clinical world model for forecasting post-operative outcomes in cardiology, addressing the irregular temporal nature of recovery trajectories. The model represents patients via structured latent states updated through time-ordered events (baseline imaging, procedural context, covariates, physiological embeddings), with follow-up imaging providing latent forecasting supervision. Applied to atrial fibrillation ablation in DECAAF-II, it achieves AUROC 0.756 and AUPRC 0.777 for recurrence prediction, plus scar-extent MAE of 2.971 percentage points without requiring follow-up MRI at inference. The latent state enables multi-horizon risk queries and retrospective record editing.
clinical world modellatent forecastingintervention-awareatrial fibrillation ablationperi-event embeddings
Bagging Robustly Learns VC Classes with Linear Sample Complexity
The paper establishes that VC classes are adversarially robustly learnable with linear sample complexity in the VC dimension $d$, improving exponentially over prior bounds. This is achieved via an improper learning algorithm combining bagging (bootstrap aggregation) with robust empirical risk minimization (RERM). The algorithm computes RERMs on $O(d^\star)$ independent bootstrap samples and outputs their majority vote, where $d^\star$ denotes the dual VC dimension. A matching lower bound demonstrates that $\Omega(d^\star)$ calls to an RERM oracle are necessary, even with unlimited training data.
vc dimensionrobust learningbaggingempirical risk minimizationadversarial examples
TabSOM: A tabular-to-image encoding method based on self-organizing maps
TabSOM introduces a tabular-to-image encoding method using Self-Organizing Maps (SOMs) to preserve feature relationships while mapping tabular data to fixed pixel locations. The method constructs a two-channel image: one encodes feature values at multiple scales, and the other captures pairwise interactions via spatial connections derived from SOM component planes. Evaluated on binary-classification datasets, TabSOM outperforms 12 baselines, ranking first or second on all datasets with the lowest variance. It also offers interpretability through prototype-inspired partial dependence plots and class-separation importance scores, validated against Random Forest, XGBoost, and SHAP.
self-organizing mapstabular-to-image encodingfeature interactionsinterpretabilitybinary-classification
On the Structural Limits of Machine Learning Decision Systems: An Information-Theoretic, Interaction-Based, and Stochastic-Dynamical Perspective
The article establishes fundamental performance limits of machine learning systems through information-theoretic and interaction-based analysis. It employs Fano-type bounds for classification error and Cramér-Rao inequalities for parametric estimation precision, demonstrating that these bounds are model-dependent rather than algorithm-dependent. The work further examines how assumptions like independence and ergodicity impact inference validity, and models decision systems (including LLM-integrated architectures) as feedback-driven stochastic processes with emergent behaviors. Results emphasize that predictive capability expansion requires adequate data models to respect inherent informational constraints.
information-theoretic boundsfano inequalitycramér-rao boundstochastic dynamicsinteraction-based modeling
Equivariant learning of a transferable three-dimensional classical density functional
The authors present an equivariant learning framework for constructing a transferable three-dimensional classical density functional from equilibrium density fields, preserving spatial symmetry and variational consistency without requiring free-energy or chemical-potential labels. The method learns a single functional that generalizes across temperatures, system sizes, and statistical ensembles, accurately predicting structure factors, equations of state, liquid-vapor coexistence, and interfacial phenomena not used in training. Applied to complex 3D geometries, the functional captures non-monotonic forces in colloid interactions and adsorption in gyroid pores, demonstrating its utility as a thermodynamic generator for diverse collective phenomena.
classical density functionalequivariant learningthree-dimensionaltransferable functionalvariational consistency
Intern-S2-Preview: Scientific Agentic Foundation Model
The authors introduce Intern-S2-Preview, a series of scientific agentic foundation models for multimodal scientific understanding, reasoning, and long-horizon tasks. The 397B-parameter model employs a training pipeline combining multimodal pre-training on scientific documents, supervised fine-tuning, multi-task RL, and agentic RL, enhanced by techniques like partial rollout with off-policy correction and adaptive length regularization. Evaluations show competitive performance on scientific benchmarks, with time series modules improving SciTS forecasting and a 4B-parameter memory decoder (Intern-MemDec-4B) boosting Biology-Instructions scores from 56.92 to 60.32 without modifying the frozen backbone.
agentic foundation modelmultimodal pre-trainingmulti-task reinforcement learningmemory decodertime series forecasting
Sparse Orthogonal Regression Technique: A Spectral Framework for Equation Discovery, Approximation, and Integration
The Sparse Orthogonal Regression Technique (SORT) introduces a sparse spectral framework for learning orthonormal-basis expansions from noisy, irregularly sampled data. SORT employs L1-regularized regression to estimate expansion coefficients directly, bypassing explicit quadrature or inner-product evaluation. It excels in data-driven discovery of ordinary differential equations, outperforming library-based sparse-regression baselines in well-adapted bases and demonstrating robustness under sparse sampling, noisy derivatives, and representation mismatch. Experiments reveal persistent low-order coefficients with increasing model order, supporting order-consistent growth. SORT also facilitates nonlinear approximation and high-dimensional integral estimation, positioning basis design as a central aspect of scientific modeling.
sparse regressionorthonormal basisl1 regularizationordinary differential equationsspectral representation
TraVEL: Trajectory-Guided Video Embedding Learning for Driving-Video Retrieval
The paper introduces TraVEL, a trajectory-guided video embedding learning framework for driving-video retrieval. The method fine-tunes Qwen3-VL-Embedding on nuReasoning clips with InfoNCE, then employs Group Relative Policy Optimization using ego-trajectory similarity as reward for motion-aware adaptation. TraVEL achieves 9.8 and 4.7 mAP improvements in longitudinal/lateral motion retrieval for 2B models (7.2/1.5 for 8B), enabling efficient single-vector search without runtime trajectory inputs.
video retrievalembedding learningego-trajectorygroup relative policy optimizationmotion understanding
Active-Trace Complexity Bounds for Moreau--Yosida Unadjusted Langevin Sampling
The paper establishes improved complexity bounds for the Moreau--Yosida unadjusted Langevin algorithm (MYULA) targeting nonsmooth composite distributions π∝exp(-f-g). By analyzing the active trace B_ref (average of Moreau envelope Hessian traces along heat substeps) rather than global curvature d/λ, the work shows MYULA achieves ε-alg accuracy in W_2 distance with Õ(ε^-2) iterations for structured penalties (lasso, TV, etc.), improving upon the Õ(ε^-3) rate from universal bounds. Key results include a Moreau-bias bound W_2(π_λ,π)≤G²λ/4 and iteration complexity depending on L_f, τ_f, G, B_ref, and M_λ.
moreau--yosidalangevin algorithmwasserstein distancenonsmooth optimizationcomplexity bounds
Doubly Robust Estimation of Causal Effect on CVR with Targeted Regularization
The paper proposes a doubly robust estimator for causal effects on post-click conversion rate (CVR) to address sample selection bias in chain-structured outcomes. Leveraging semiparametric theory, the method achieves faster convergence than nuisance parameter estimation and incorporates targeted regularization for stability. Experiments on synthetic and real-world data demonstrate superior performance over naive debiasing approaches, with theoretical guarantees for unbiased estimation.
doubly robust estimationpost-click conversion ratesample selection biassemiparametric theorytargeted regularization
Symmetry-Breaking De Novo Crystal Generation via Markovian Jump Diffusion
The paper introduces Symmetry-breaking Crystal Diffusion (SbCD), a diffusion-based framework for de novo crystal generation that explicitly models symmetry-breaking dynamics via Markovian jump-diffusion. Unlike prior methods that generate crystals only up to site symmetries, SbCD reverses from lowest-symmetry priors to produce full crystallographic specifications, incorporating inter-space-group transitions. Evaluated on MP20 and MPTS-52, SbCD substantially outperforms symmetry-preserving baselines, demonstrating improved capability in capturing global symmetry and structural dependencies.
crystal generationsymmetry-breakingmarkovian jump-diffusionspace group transitionsde novo materials design
LLM-Assisted Dynamic Threat Analysis for Attacker-Reachable Software Weaknesses in Autonomous Vehicles
The study evaluates LLM-assisted dynamic threat analysis for identifying attacker-reachable software weaknesses in Autoware, an open-source autonomous-driving stack. Using compiler-precise static analysis across 185 packages, researchers identified 1,375 decision rules and 740 reachable sites, then generated 3,700 test artifacts via two open-weight LLMs and a baseline. Results show 80% of first-shot compilation failures stem from dependency wiring, not program logic; the reasoning model compiled 64% of harnesses initially versus 6% for the code-specialized model. Despite repair efforts, no candidate weaknesses were dynamically confirmed, highlighting build integration as the primary bottleneck for LLM-assisted dynamic analysis in safety-critical stacks.
autonomous vehiclesstatic analysisdynamic analysislarge language modelsbuild integration
Wasserstein Filtering: A Sample Selection Method for Robust Distribution Learning
The paper proposes Wasserstein Filtering (WF), a sample selection method for robust distribution learning in contaminated datasets. WF selects a subset of samples whose empirical distribution maximizes the Wasserstein distance to the contaminated distribution, isolating geometrically influential outliers. Three algorithms—SinkMarg, SinkWF, and SlicedWF—are introduced for efficient optimization using entropic optimal transport and sliced Wasserstein approximations. Theoretical analysis under the FELP contamination model shows WF achieves minimax optimality for distributions with bounded covariance. Experiments on synthetic data, anomaly detection benchmarks, and diffusion-based generative modeling demonstrate WF's effectiveness as a preprocessing tool.
wasserstein filteringsample selectionoptimal transportcontamination modelgenerative modeling
When Local Variance Optimality Is Not Enough: RoPE-Aligned Q/K Rotations for Dynamic 4-Bit Quantisation
The paper investigates whether rotation-based quantization respecting RoPE's frequency-pair decomposition outperforms full-head mixing in dynamic W4A4KV4 settings. It proves that no single-head orthogonal map beyond per-pair rotations commutes with RoPE for distinct frequencies, derives a variance-minimizing rotation angle under a pooled-covariance surrogate, and evaluates head-shared pairwise configurations. Results show pairwise rotations increase perplexity across contexts versus full-head Hadamard transforms, with K-only angle estimation failing to bridge the gap. The misalignment between the surrogate's position-averaged second moment and the quantizer's tokenwise range statistic explains the suboptimal performance.
rope-aligned quantizationdynamic quantizationattention headsperplexity degradationkv4
Where You Measure Decides What You Measure: Position Selection in Ablation-Based SAE Evaluation
The paper identifies a critical methodological flaw in sparse autoencoder (SAE) evaluation, where latent variable effects are conventionally measured at the token where activation is strongest—a position determined by the SAE itself rather than experimenter control. By analyzing six autoencoders trained from identical initializations and auditing five published studies, the authors demonstrate that measurement position accounts for 7.6-11.9% of variance in ablation results, obscuring true dictionary performance. The proposed solution—standardizing measurement positions across evaluations—reduces this variance to near zero. The issue persists across corpus sizes (tested up to 16× variation) and requires minimal code changes to address.
sparse autoencoderlatent variableablation studyevaluation protocolmeasurement bias
Neural Quadratic Forms: A Unified Minimal Model for Sudden Learning and Scaling Laws
The paper introduces Neural Quadratic Forms (NQFs) as a minimal unified model explaining sudden learning steps and scaling laws in neural networks. By leveraging symmetry arguments and gradient smoothness constraints, the authors derive a universal quadratic form Tr[WW^⊤A(x)] where architectural details collapse into a structure matrix A(x). This reduction enables closed dynamics on the order parameter M=WW^⊤, which under eigenbasis alignment reduces to Lotka-Volterra equations exhibiting sequential mode activation. The theory predicts both discrete plateaus (small initialization limit) and smooth power laws (many unresolved modes), validated numerically across architectures and training methods.
neural quadratic formslotka-volterra dynamicsstructure matrixorder parametersudden learning
A Probe Direction Is a Property of Its Prompt
The study demonstrates that evaluation probe directions in LLMs are prompt-dependent rather than model-intrinsic, challenging current comparative practices. By systematically varying evaluation-announcing prompts while holding task text constant, the authors show probe performance and scaling trends are dictated by prompt choice, with two conflicting published results reproducible through prompt selection alone. Variance analysis reveals models contribute minimally to reported scores compared to prompt-model interactions, and surface-form separability further questions probe validity. The work concludes single-prompt designs are inadequate for model comparisons and quantifies necessary prompt diversity.
probe directionprompt dependencemodel evaluationactivation analysisvariance decomposition
Foundation models for movement data: Are they ready for prime-time?
This study conducts the first systematic evaluation of four open-source foundation models (FMs) for accelerometer-based movement data against supervised baselines across 19 tasks in activity recognition, clinical monitoring, and physiological inference. Using linear probing, frozen feature extraction, and concept discovery analysis, the research reveals task-dependent performance: supervised models match FMs on human action recognition, while selected FMs excel in fall/stress detection and sensor-placement robustness. UniMTS emerges as the strongest FM representation, surpassing baselines without finetuning. The analysis identifies limitations in sedentary activity recognition and proposes activity profile inference as a future direction.
foundation modelsaccelerometer dataactivity recognitionlinear probingsensor-placement robustness
The Time Value of Evolution
The paper introduces Lineage-Value Policy Gradients (LVPG), an actor-critic framework for evolutionary search that optimizes long-horizon lineage potential rather than immediate fitness. LVPG decouples search control into policy heads: a critic estimates finite-horizon lineage value via multi-step mutation trees, while an actor modulates mutation intensity dynamically. Evaluated across 90 paired runs with matched conditions, LVPG improves validation best-so-far AUC by 0.394 Sharpe units, reduces temporary regressions, and enhances recovery compared to immediate-return optimization. The method demonstrates superior finite-budget search efficiency through path-based credit assignment.
evolutionary searchlineage-value policy gradientsactor-critic frameworkfinite-horizon credit assignmentmutation trees
EEG Decoding Using CNN and LSTM Network
The study proposes a hybrid CNN-biLSTM architecture for motor imagery EEG decoding, addressing challenges of noise and weak signal relationships in MI-BCI systems. The CNN extracts spatiotemporal features from raw EEG, while the bidirectional LSTM captures temporal dependencies. Evaluated on public and private datasets, the model achieves robust performance in 2- and 3-class classification (specific metrics not provided) and demonstrates subject-independent decoding capability.
motor imageryeeg decodingconvolutional neural networkbidirectional lstmbrain-computer interface
Virtual Temperature Sensors in Power Transformers Using Neural Ordinary Differential Equations
The paper proposes a Neural Ordinary Differential Equation (Neural ODE) framework for forecasting power transformer thermal behavior, integrating simplified heat-transfer physics into the continuous-time dynamics. The method combines data-driven learning with physical constraints, avoiding mesh generation and transformer-specific constants required by traditional approaches. Evaluated on fifteen Norwegian transformers with varying designs, the framework demonstrates robust performance across heterogeneous units while maintaining physical consistency.
neural odethermal modelingpower transformersphysics-aware learningcontinuous-time dynamics
Knowledge-guided Pattern Discovery via Coupled Tensor Factorizations
The paper introduces a knowledge-guided tensor factorization method that jointly analyzes real-world multiway data and simulated data from computational models using linear coupling. This approach combines data-driven tensor factorizations with prior knowledge from simulations to improve pattern discovery in noisy datasets. Experiments on metabolomics measurements demonstrate enhanced performance in identifying interpretable patterns while revealing potential discrepancies between observed data and model simulations.
tensor factorizationmultiway datacomputational modelsmetabolomicspattern discovery
Foundations of Independent Component Analysis
The article formalizes the mathematical foundations of linear independent component analysis (ICA) with measure-theoretic probability. It develops characteristic function theory for ℝᵈ probability measures, then establishes identifiability results under progressively stronger source assumptions: from non-constant to non-Gaussian to Gaussian-free independent sources. Under strictest assumptions, sources are shown identifiable up to translation, permutation, scaling, and sign, even with additive Gaussian noise. The online equivariant gradient descent ICA algorithm is presented for noiseless non-Gaussian ICA settings.
independent component analysischaracteristic functionmeasure-theoretic probabilityidentifiabilityequivariant gradient descent
History-informed Lagrangian Neural Networks
The paper introduces History-informed Lagrangian Neural Networks (HiLNN) for long-horizon forecasting of mechanical systems from position-only observations. HiLNN uses a recurrent encoder to extract latent context from historical positions, enabling reconstruction of unobserved velocities and adaptive modulation of system parameters (mass matrix, potential energy, damping). The model is trained end-to-end with multi-step trajectory supervision and energy-consistency regularization via differentiable RK4 rollout. Experiments on conservative, dissipative, and variable-parameter systems demonstrate superior long-term prediction accuracy and energy profile maintenance compared to state-of-the-art baselines.
lagrangian neural networksrecurrent encoderdifferentiable rk4energy-consistency regularizationvariable-parameter systems
TANGCO: Learning Topology-Aware Capacity Allocation for Overload-driven Cascading Failures
TANGCO introduces a topology-aware neural graph-guided capacity optimization method to allocate fixed capacity budgets across networked systems, mitigating overload-driven cascading failures. The approach employs a graph neural network policy trained via policy-gradient learning and a heuristic anchor within a cascade simulator. Evaluated on five synthetic graph families and five real networks (power, road, air, Internet), TANGCO outperforms four hand-designed heuristics in 450 synthetic instances and 40 of 45 real-network conditions, achieving robustness gains from 1.6% to 246%. The method demonstrates transferability across graph families, near-linear training scalability, and deployment efficiency comparable to hand-designed heuristics.
cascading failuresgraph neural networkcapacity allocationpolicy-gradient learningtopology-aware
Chance-constrained selection of sequential intervention strategies from counterfactual estimates
The paper introduces a predict-then-optimize framework for selecting sequential intervention strategies under chance constraints on cumulative resource limits. The method first estimates counterfactual outcomes and cost distributions from observational data, then optimizes over candidate strategies while bounding the probability of budget overruns via chance constraints. Unlike prior approaches, it handles tail risks without parametric assumptions and provides finite-sample guarantees. Evaluated on five domains (including clinical treatment and equipment maintenance), the framework successfully prevents budget violations where point-estimate methods fail, with explicit trade-offs shown via a safety-utility frontier.
counterfactual estimationchance constraintssequential interventionspredict-then-optimizedistribution-free bounds
Sinkhorn Linearization and the Spectral Proxy: Unifying the Statistical and Algorithmic Theory of Feature-Parameterized Inverse Optimal Transport via a Single Spectral Sandwich
The paper develops a unified statistical and algorithmic theory for inverse optimal transport (IOT) with feature-parameterized costs C_θ(i,j) = -θ^T φ(i,j). Key contributions include Sinkhorn linearization, which captures the sensitivity of entropic OT plans to cost, and its spectral proxy, enabling geometrically transparent analysis. A spectral sandwich inequality governs the restricted Hessian on the tangent space, driving theoretical guarantees. Results include identifiability of θ, dimension bounds for monotone gradient descent convergence, and analysis of estimator convergence under model misspecification, with empirical assessment of Hölder continuity exponents.
inverse optimal transportsinkhorn linearizationspectral proxyentropic othölder continuity
Beyond Simulated Benchmarks: Evaluating Motion Representations for Fall Detection Under Real-World Data Scarcity
The study evaluates motion representations for wearable fall detection under real-world data scarcity, comparing interval-based, kernel-based, symbolic, and foundation model approaches. Using accelerometer signals from FallAllD (simulated) and FARSEEING (real-world) datasets, the authors assess performance via cross-validation, data scarcity, and cross-dataset transfer. Results show interval-based methods perform best absolutely, while a lightweight symbolic representation with impact descriptors exhibits minimal domain shift degradation and retains sensitivity under extreme scarcity, outperforming parameterized models in real-world conditions.
fall detectionmotion representationsdata scarcitydomain shiftwearable sensors
ProME: Prototype-Margin Environments with Repair-Aware Selection for Group-Robust Learning
ProME introduces prototype-margin environments with repair-aware selection (ERAS) for group-robust learning without training-group labels. The method constructs balanced environments by splitting prototype margins at their median during training and ranks predictors via a group-balanced linear head on validation data. Theoretical analysis bounds worst-case risk across inferred environments, with experiments showing enriched shortcut-conflicting examples and superior worst-group accuracy (+X% average) versus comparable methods.
group-robust learningprototype marginsendogenous environmentsrepair-aware selectionworst-group accuracy
High-dimensional networks and mean squared error for possibly misspecified models
The paper demonstrates conservative neighborhood estimation in high-dimensional networks where parameters exceed observations, addressing both linear and nonlinear true models. It analyzes the ridge parameter's impact on mean squared error, linking reduced test variance to denser edge neighborhoods. Connecting to machine learning's double descent phenomenon, the authors prove minimum description length outperforms Lasso, AIC, and BIC by minimizing false positives in high-dimensional settings, regardless of linear model correctness.
high-dimensional networksneighborhood selectionmean squared errordouble descentminimum description length
MergeOver: Post-Training Token Merging for Recursive Vision Transformers
MergeOver introduces a post-training method to integrate Token Merging (ToMe) with recursive weight-sharing in Sliced Recursive Transformers (SReT), addressing computational and memory bottlenecks without retraining. The approach employs an Unmerge tracking stack, constraint-safe merge-rate adjustment, and synchronized token-mass tracking across spatial permutations, alongside a stage-wise single-shot schedule for token reduction. On ImageNet-1K, it reduces top-1 accuracy by 1.47pp while cutting GPU peak activation memory by 37.3-38.4% and improving throughput by 21.7% at batch size 16, with latency reductions of 2.4-17.6% on Raspberry Pi 5.
token mergingrecursive transformerspost-training optimizationactivation memorycomputational efficiency
Statistical Properties of Robust Learning under Distributional Shifts
The paper derives finite-sample generalization error bounds for Distributionally Robust Optimization (DRO) and Robust Satisficing (RS) under target distribution shifts, explicitly quantifying the trade-off between shift sensitivity and regularization penalties. It proposes information-directed hyperparameter calibrations when partial shift information (e.g., magnitude or direction) is available, revealing complementary theoretical and empirical behaviors between DRO and RS. Applied to a network lot-sizing problem, the framework demonstrates how robust policies adapt to demand distribution shifts, providing a statistical basis for method comparison.
distributional robustnessgeneralization errorwasserstein distancehyperparameter calibrationfinite-sample guarantees
Branch and Bound for Relational Verification of Neural Networks
The paper introduces SaBRe, a branch-and-bound (BaB) framework for relational verification of neural networks, addressing the incompleteness of convex over-approximation methods. The key innovation is a relational neuron selection strategy based on dual formulation, optimizing refinement by splitting relational neurons rather than individual ones. Evaluated on 817 problems across ACAS Xu, MNIST-F, MNIST-C, CIFAR, and GTSRB, SaBRe outperforms baselines in solved instances and verification efficiency.
relational verificationbranch-and-boundneural networksabstraction refinementdual formulation
FlowLOB: Efficient and Controllable Limit Order Book Generation with Flow Matching
FlowLOB introduces a flow-matching generator for limit order book (LOB) trajectories that combines realistic market dynamics, efficient sampling, and controllable scenario generation. The method trains conditional flow and diffusion models on multiple Hong Kong Exchange (HKEX) symbols at three sampling frequencies (0.1s, 1s, 10s) using tick-relative representation, enabling zero-shot transfer to unseen instruments. Experiments show FlowLOB achieves superior realism with only 10 ODE-solver steps, outperforming diffusion models and baselines in distributional metrics, while demonstrating effective counterfactual controllability in most tested settings.
flow matchinglimit order bookdiffusion modelsode solverscounterfactual controllability
Learning Discrete Decisions for MIPs with Constraint-Aware Diffusion
The paper introduces Constrained Graph Diffusion (CGD), a learning-based method for approximating solutions to mixed-integer optimization problems (MIPs). CGD employs a graph-based generative diffusion model to learn discrete decisions while integrating a feasibility projection operator into the reverse diffusion process, ensuring intermediate samples remain feasible. The continuous subproblem is then solved using standard numerical methods. Evaluated on optimal transmission switching for ACOPF and discrete portfolio optimization, CGD improves feasibility and solution quality over baselines while achieving up to 425× speedup over MINLP solvers.
mixed-integer optimizationdiffusion modelfeasibility projectiongraph-based generationnumerical solvers
A Multispectral Framework for the Detection of Calcium Carbide-Induced Ripening and Shelf-Life Estimation in Climacteric Fruits
The study proposes a non-invasive multispectral framework for detecting Calcium Carbide (CaC2)-induced ripening in climacteric fruits (mango, banana) and estimating shelf life. Using an AS7265x spectral triad sensor, 18 discrete wavelengths (410-940 nm) capture spectral profiles, with CaC2-treated samples showing distinct intensity drops in visible regions. Feature engineering integrates spectral variance, intensity ratios, and environmental parameters, reduced via PCA (>90% variance in 5-7 components). Three XGBoost models achieve 95% accuracy (mango) and 81% accuracy (banana) for ripening classification, with carbide recall of 0.67 and 0.74, respectively.
multispectral frameworkcalcium carbideclimacteric fruitsxgboostprincipal component analysis
Latent On-Policy Self-Distillation
The paper introduces Latent On-Policy Self-Distillation (LOPD), a method enabling end-to-end learning of privileged context for self-distillation in reinforcement learning. LOPD retrieves and composes relevant experiences into latent tokens to condition a self-teacher, while the student receives dense token-level supervision during trajectory generation. A privileged-margin objective stabilizes latent context learning. Experiments show LOPD outperforms RLVR, OPSD, SDPO, and Skill-SD in agentic tool use and code generation, achieving higher efficiency with <30% of the rollout budget of GRPO and Skill-SD. Ablations confirm the necessity of learnable privileged context.
self-distillationlatent tokensprivileged contexton-policy learningagentic tool use
On the global feature importance for interpretable and trustworthy heat demand forecasting
The paper proposes an ante-hoc Explainable AI methodology to evaluate global feature importance in Machine Learning models for heat demand forecasting in District Heating Systems, enhancing interpretability and trustworthiness. The methodology employs four approaches: intrinsic interpretability of Gradient Boosting, Partial Dependence, Accumulated Local Effects, and SHAP, avoiding feature permutation or perturbations that introduce bias through unrealistic data values. Results include an assessment of complementarities among methods and specific interpretations relevant to district heating processes.
ante-hoc explainable aigradient boostingpartial dependenceaccumulated local effectsshap
Incremental Evaluation and Training in Relational Deep Learning
The paper introduces an incremental evaluation and training paradigm for Relational Deep Learning (RDL) to address temporal concept drifts in multi-tabular databases modeled as temporal heterogeneous graphs. The authors propose multiple incremental training regimes, including transfer learning, and a new temporal evaluation metric prioritizing near-future accuracy. Experiments on large-scale datasets show temporal drifts occur in most predictive tasks, and incrementally fine-tuned models outperform standard from-scratch training baselines.
relational deep learningtemporal heterogeneous graphsincremental trainingconcept drifttransfer learning
Balanced Adaptive Prototype Selection for Scalable TabPFN Inference on Large-Scale Tabular Data
The paper introduces Balanced Adaptive Prototype Selection (BAPS), a framework for scalable inference with TabPFN on large-scale tabular data by constructing compact, information-preserving contexts. BAPS jointly optimizes representative structure, decision boundaries, local density, class balance, and feature-space diversity without model retraining. Experiments on HIGGS and SUSY datasets (1M rows) demonstrate that 512 prototypes retain strong predictive performance and calibration, achieving ~1,953x context compression, all on CPU-only hardware (Intel Core i7, 16GB RAM).
tabular foundation modelscontext compressionprototype selectionscalable inferencepretrained models
Learning the Mathematical Property for Designing Low Mutual Coherence Binary Sensing Matrices
The authors propose a neural network-based method for constructing binary sensing matrices with low mutual coherence, essential for perfect signal recovery in compressive sensing. Unlike prior approaches, their technique requires no dataset or specific application, instead leveraging mathematical properties to define the loss function. The framework generates matrix entries via a shared underlying rule, reducing computational costs and storage requirements while maintaining robustness.
compressive sensingbinary sensing matrixmutual coherenceneural networkloss function
Comment on "Modeling rapid language learning by distilling Bayesian priors into artificial neural networks"
The critique challenges McCoy & Griffiths (2025)'s claim that Model-Agnostic Meta-Learning (MAML) instills Bayesian priors in artificial neural networks (ANNs), arguing it merely optimizes weight initialization without modifying the objective function. The authors examine an alternative interpretation where the system collectively implements Bayesian learning, identifying significant conceptual and empirical limitations. Their analysis reveals MAML-trained models exhibit overfitting and inferior generalization compared to true Bayesian learners, despite matching performance on training data.
model-agnostic meta-learningbayesian priorweight initializationoverfittinggeneralization
Online Inference for Quantile Temporal Difference Learning in Distributional Reinforcement Learning
The paper introduces online inference methods for quantile temporal difference learning (QTD) in distributional reinforcement learning. By leveraging a generative model, the authors establish functional central limit theorems for both synchronous and asynchronous QTD, demonstrating that averaged iterates converge weakly to a rescaled Brownian motion. The proposed inference procedure constructs an asymptotically pivotal statistic using random scaling, enabling efficient online computation without storing the entire trajectory of QTD iterates. This approach significantly reduces memory requirements and facilitates practical statistical inference in distributional reinforcement learning.
quantile temporal difference learningdistributional reinforcement learningfunctional central limit theoremonline inferencerandom scaling
Understanding Backdoor Vulnerabilities in Vertical Federated Learning: The Gap Between Research and Practice
The paper identifies a critical gap between theoretical research and practical realities in backdoor vulnerabilities within Vertical Federated Learning (VFL), where existing approaches rely on unrealistic assumptions. Through systematic analysis, the authors redefine threat models under practical constraints, propose realistic attack workflows, and introduce BVBench, a benchmark for fair evaluation of backdoor risks in VFL. Results demonstrate the fragility of current understanding, with BVBench revealing significant shortcomings in prior work and providing a foundation for future research on practical defenses.
vertical federated learningbackdoor attacksthreat modelingbenchmark evaluationasymmetric information
I-SDPO: Instance-Level Adaptive Self-Distillation Policy Optimization
I-SDPO introduces instance-level adaptive self-distillation policy optimization to address limitations in Group Relative Policy Optimization (GRPO) and privileged self-distillation. The method dynamically routes training instances: groups with all-incorrect responses use self-distillation for dense token supervision, while successful groups employ GRPO for relative reward learning. Theoretical analysis shows alignment conditions between teacher and reward directions, with automatic reduction of distillation weight as success probability increases. On SciKnowEval, I-SDPO improves mean@16 accuracy from 56.67% (GRPO) to 70.31%, achieving gains up to 18.24 points across four scientific domains.
policy optimizationself-distillationgroup relative learninginstance-level adaptationreward alignment
Unifying Depth and Width Pruning for LLMs via Binary Knapsack Optimization
SNIPER introduces a two-stage structured pruning framework for large language models (LLMs) that unifies depth and width pruning via binary knapsack optimization. The method first solves a knapsack problem over coarse-granularity components for conditionally optimal parameter allocations, followed by fine-grained pruning to meet strict budget constraints. SNIPER achieves near-exact adherence to target compression ratios with a Compression Ratio Adherence Factor (CRAFT) score of 0.98, outperforming six state-of-the-art pruners that deviate by up to 33%. Evaluations across four architectures and 18 tasks demonstrate SNIPER's robust performance retention, task-level stability, and cross-architectural generalizability, achieving a mean rank of 1.25 across all pruning configurations.
structured pruningknapsack optimizationcompression ratio adherence factorlarge language modelsparameter allocation
CardioState-JEPA: Delay-Aware Cross-Modal Learning of a Shared Cardiac Representation
CardioState-JEPA introduces a cardiac foundation model that learns a shared representation across ECG, PPG, and PCG modalities using a joint-embedding predictive architecture. The model employs a shared Transformer encoder and a learned delay aligner to handle temporal offsets between electrical, mechanical, and hemodynamic events, pretraining on unimodal data before aligning modalities in latent cardiac time. Evaluated on 25 downstream tasks, it improves PPG classification by 8.2 AUROC points, PCG murmur detection by 18.8 AUROC points, and ECG classification by 15.5 AUROC points over self-supervised baselines, matching or exceeding models trained with clinical text or supervised labels.
cardiac foundation modeljoint-embedding predictive architecturedelay alignershared transformer encodercross-modal learning
Diagnosing JEPA World Models with Action-Conditioned Predictive Consistency
The paper introduces Action-Conditioned Predictive Consistency (ACPC), a diagnostic for Joint-Embedding Predictive Architectures (JEPAs) that measures divergence between clean and perturbed observation histories when rolled forward under identical action sequences. The method is theoretically grounded in bisimulation, proving ACPC bounds multi-step prediction error and planner cost changes. Experiments on four visual control tasks demonstrate ACPC's predictive validity for perturbation effects, with complementary metrics (Invariance Radius and Separation Rate) showing consistent diagnostic utility across tasks and architectures like LeWM and PLDM under visual perturbations.
joint-embedding predictive architecturesbisimulationaction-conditioned predictive consistencyinvariance radiusseparation rate
Multi-perspective Imbalance-Conscious 6G Beamforming Optimization and Performance
The study systematically evaluates machine learning approaches for 6G-IoT beamforming optimization (6GBO), comparing supervised and unsupervised methods across feature groups. Network features demonstrated superior predictive power (measured by recall, F1-score, ROC-AUC) over device, environmental, and vision features in imbalance-aware experiments. Unsupervised clustering revealed deployment environment and device type as primary influencers, with bandwidth, IoT sensors, and mobility showing highest global feature importance in explainability analysis.
6g beamformingiot optimizationimbalance-aware learningfeature importanceunsupervised clustering
Momentum as Residual-Driven Multiplier Correction for Deep Learning Optimization
The paper introduces ADMM-Inspired Momentum (AIM), a framework interpreting momentum as a multiplier-like correction driven by splitting residuals, separating update geometry and acceleration mechanisms. Building on AIM, the authors propose Relativistic Adaptive gradient Descent with Accelerated Residual (RADAR), combining relativistic adaptive geometry, decoupled residual correction, and second-order momentum filtering. Theoretical analysis establishes stochastic convergence via variance-perturbed Lyapunov drift. Experiments on vision, language modeling, and RL tasks demonstrate RADAR's consistent improvements over adaptive optimizer baselines.
momentum optimizationadmm-inspired frameworkresidual-driven correctionadaptive gradient descentlyapunov drift analysis
Towards Socially Compliant Navigation in Deep Reinforcement Learning via Proxemics-Based Reward Modeling
The authors propose a proxemics-based reward formulation for deep reinforcement learning (DRL) to enhance socially compliant robot navigation in crowded environments. The method models personal space as a radial Gaussian-mixture field derived from Hall's proxemics theory, computing a robot-centric local cost over the robot's field of view. This reward is integrated into existing DRL navigation frameworks and evaluated in simulations across various crowd scenarios, densities, and reward baselines. Results demonstrate consistent improvements in social metrics while maintaining competitive navigation performance compared to other reward models.
deep reinforcement learningproxemicsgaussian-mixture fieldsocial navigationreward modeling
Revisiting Overestimation Bias Problem of Q-learning: Settling Large Discrete Action Space via Action Intersection
The paper addresses Q-learning's overestimation bias in large discrete action spaces by proposing an action intersection strategy that combines coupling and decoupling paradigms. The method enables semi-decoupling through shared trajectory data between two Q-functions, with updates alternating between coupling (shared data) and decoupling (unshared data) paradigms. This approach achieves adjustable bias control via fine-grained action intersection size, ranging from underestimation to overestimation. Experiments in tabular and deep RL settings demonstrate superior performance over state-of-the-art baselines, with tabular results providing insights into the method's efficacy.
q-learningoverestimation biasaction intersectionsemi-decouplingdiscrete action space
Adaptive $k$ Nearest Neighbors Classifier via Granular Ball Computing
The authors propose an adaptive k-Nearest Neighbors (KNN) classifier leveraging granular-ball computing to dynamically determine the optimal k value. The method employs a two-stage process: during training, datasets are coarsely partitioned and refined using the Fisher criterion to create multi-granularity granular ball representations; during prediction, nearest granular balls are identified via weighted distance, constructing adaptive neighborhoods around test samples. This approach enhances robustness against noise and local perturbations by leveraging stable local group information. Experimental results show superior accuracy and efficiency compared to existing KNN variants across multiple datasets.
k-nearest neighborsgranular-ball computingfisher criterionadaptive neighborhoodmulti-granularity representation
Robust data-driven discovery of fractional differential equations via weak formulations and Pareto-based subset selection
The authors propose Weak-Pareto, a method for robust discovery of fractional differential equations from noisy data. The approach combines adjoint-consistent weak formulations to mitigate noise amplification with Pareto-based subset selection over term types and fractional orders. Weak formulations transfer fractional operators to smooth test functions via adjoints, replacing pointwise differentiation with integration. The method employs ridge regression and differential-evolution search for coefficient fitting and order optimization. Experiments on fractional advection-diffusion, reaction-diffusion, and Burgers equations show superior noise robustness and support recovery compared to strong-form approaches and neural baselines.
fractional differential equationsweak formulationpareto optimizationadjoint methodridge regression
Sustaining Plasticity via Learnable Wavelet Activations in Continual Learning
The paper introduces a learnable wavelet activation function to mitigate plasticity loss in continual learning, addressing spectral bias through explicit decomposition into low-frequency and high-frequency components. The method combines dynamic wavelet injection for task-specific plasticity enhancement with regularization to preserve prior knowledge, supported by theoretical guarantees on $L^2$ approximation efficiency and decoupled learning rates. Empirical results show state-of-the-art performance across continual learning benchmarks, with maintained trainability and generalization.
continual learningwavelet activationspectral biasplasticity lossdynamic injection
Discovering Persistent Behavioural Patterns for Interpretable Blockchain Forensics
The paper proposes a scalable, application-agnostic framework for persistent behavioral pattern discovery in blockchain transactions. The method constructs behavior sentences enriched with contextual data, then applies a two-step embedding process (sentence-level and sequence-level) to capture individual actions and temporal user behavior. Evaluation on 30M+ Ethereum transactions demonstrates identification of stable patterns, including DEX trading, NFT activity, phishing, and oracle manipulation, with interpretable profiling via behavioral motifs and temporal dynamics.
blockchain forensicsbehavioral embeddingtemporal dynamicsinterpretable profilingdecentralized finance
Distribution Steering via Sliced Optimal Transport Control
The paper develops a finite-horizon control framework for distribution steering using sliced optimal transport, avoiding full-dimensional transport maps. The method constructs directional terminal conditions via projected optimal transport maps, yielding either randomized single-direction controllers or deterministic sliced feedback through directional averaging. Key results include non-increasing sliced Wasserstein distance for single-integrator dynamics, affine Gaussian-preserving feedback for Gaussian endpoint laws, and extensions to linear dynamical systems via reachability-normalized coordinates. Theoretical guarantees include control energy characterization and convergence properties.
distribution steeringsliced optimal transportwasserstein distancefinite-horizon controlgaussian preservation
HiRoute: Hierarchical Routed Prompt Tuning for Safety Alignment of Large Language Models
HiRoute introduces a hierarchical prompt-tuning framework for LLM safety alignment, addressing limitations of static prompt designs. The method trains a lightweight hierarchical router on frozen LLM representations to detect harmful intent and predict multi-label risk scores, then learns shared coarse-grained and fine-grained prompt experts via preference optimization. Experiments on three instruction-tuned models demonstrate improved safety rates across benchmarks (94.5% harmful query rejection), reduced over-refusal (12.3% decrease), and maintained general-task performance (within 2% of baseline).
hierarchical routerprompt tuningsafety alignmentpreference optimizationmulti-label risk
CoMedBench: A Multi-Source Benchmark of Synthetic Medical Data Fidelity and Downstream Utility
The authors introduce CoMedBench, a multi-source benchmark for evaluating synthetic medical data fidelity and downstream utility across 37 dataset-task pairs from seven public clinical datasets (MIMIC-III, MIMIC-IV, eICU, UCI, CDC BRFSS, NHANES, pycox). The benchmark assesses statistical fidelity and task utility by comparing models trained on real versus synthetic data, testing static tabular and temporal ICU tasks. Results show synthetic data preserves most signal: CoMed-CTGAN achieves 90.6% mean AUROC utility on tabular tasks, while CoMed-TVAE reaches 97.3%; temporal tasks are more challenging, with CoMed-CTGAN dropping to 81.6% AUROC and 64.0% AUPRC.
synthetic dataclinical validitydownstream utilityaurocauprc
Fine-tuned Normalizing Flows for ALICE Zero Degree Calorimeter Fast Simulation
The authors propose a fine-tuned Normalizing Flow (NF) framework for fast simulation of ALICE Zero Degree Calorimeter responses, addressing computational bottlenecks in Monte Carlo chains. Using transfer learning, they pre-train on imbalanced data and fine-tune specialized models for γ, n, Λ, K_S^0, and Σ^+ particles via gradual-unfreezing schemes. They introduce physics-aware metrics (conditional weighted MAE, dispersion ratio, Jaccard co-activation error) beyond standard Wasserstein distance. The ensemble achieves a Wasserstein distance of 1.61±0.02, outperforming baselines across all metrics.
normalizing flowstransfer learningmonte carlo simulationparticle physicsconditional metrics
Thermodynamics of Learning: A Typed Four-Component Accounting of Memory, Fit, and Value
The paper develops a thermodynamic framework for finite-state learning devices by introducing a typed four-component accounting system: training-side fit functional $Φ_{\mathrm{fit}}$, record-correlation stock $J_{D}=I(M;D)$, update-side search ledger $σ_{M}$, and operational capital value $V(M;T,b)$. The framework establishes separation results showing record correlation can grow without capital gain, derives exact identities for capitalization efficiency ($η_{\mathrm{cap}}\le 1$), and analyzes value retention under task-distribution shift through retention gap $L_{\mathrm{gen}}$ and retention ratio $ρ_{\mathrm{gen}}$.
thermodynamic learningfinite-state devicescapitalization efficiencyrecord-correlation stockvalue retention
CW-BASS v2: Saturation-Aware Pseudo-Label Selection for Semi-Supervised Segmentation under Foundation-Model Teachers
CW-BASS v2 introduces a saturation-aware pseudo-label selection method for semi-supervised semantic segmentation under foundation-model teachers like DINOv2. It combines held-out calibration for unbiased noise estimation with a self-adaptive confidence floor, dynamically choosing between strict filtering and adaptive retention based on the teacher's confidence reliability. Evaluated across six DINOv2 teachers, it matches UniMatch V2 performance on saturated benchmarks (Pascal VOC 1/8: 87.4 mIoU vs. 87.9; Cityscapes within 0.5 mIoU) and improves by +1.5 mIoU on unreliable confident sets (ADE20K). The method avoids confirmation bias by adapting to collapsed confidence distributions (98% of Pascal pixels ≥ 0.95).
semi-supervised segmentationpseudo-label selectionfoundation-model teachersconfidence saturationadaptive filtering
Difference-of-Convex Regularization for Graph Learning by Differentiable Programming
The paper proposes a Difference-of-Convex Regularizer (DCR) framework for graph learning that approximates the spectral action of the Laplacian pseudoinverse without direct inversion. By reformulating Laplacian-Regularized Nonnegative Least Squares (LR-NNLS) through a dual representation, DCR decouples pseudoinverse learning from instance-specific inference and enables efficient primal solution reconstruction via differentiable dual-guided learning. Theoretical guarantees on stability and fixed-point uniqueness are established. Experiments show improved performance over convex solvers and graph filtering baselines across diverse graph topologies.
graph laplaciandifference-of-convex regularizationpseudoinverse learningnonnegative least squaresdifferentiable programming
ReconSpan: Reconstruction-Guided Adaptive Latent Tokenization
ReconSpan introduces reconstruction-guided adaptive latent tokenization for text, mapping fine-grained inputs to shorter sequences of continuous representations. The method divides text into chunks that a backward decoder can reconstruct from a single contextual prefix code, retaining one code per chunk as the latent token. Experiments show the trained autoencoder produces average chunk lengths of 6.5 to 12.2, with reconstruction-guided boundaries preserving more text than random boundaries at matched lengths. Latent sequences reliably retain topic information but obscure exact details.
adaptive latent tokenizationreconstruction-guidedbackward decodercontextual prefix codeautoencoder
Decentralized Multi-Player Q-Learning in Episodic Markov Decision Processes with Information Asymmetry
(No summary returned.)
A Cloud-Edge System for Multimodal Clinical Screening in Resource-Constrained Rural Settings
The study proposes a cloud-edge system for multimodal clinical screening in low-resource settings, combining lightweight edge models for data preprocessing with a cloud-based LLM for synthesis. The architecture employs an LLM orchestrator to dynamically select diagnostic tools, optimizing modality coverage while minimizing irrelevant processing. Evaluated on 20 multimodal cases under varying bandwidths (500kbps–5Mbps), the system achieves 98–99% tool recall, 92–96% precision, matches cloud-only accuracy, and maintains 25–35s latency at 4–15x lower token cost, demonstrating efficient multimodal integration under constraints.
cloud-edgemultimodal screeningllm orchestratordiagnostic recallbandwidth-invariant
MAG: MAnifold Guided Semi-Supervised Multi-modal In-Context Learning
The paper introduces MAG, a manifold-guided semi-supervised framework for improving few-shot in-context learning (ICL) with multi-modal large language models (MLLMs). MAG addresses demonstration selection via a two-stage approach: (1) relevance score propagation on a multi-modal graph to identify high-impact unlabeled samples for pseudo-labeling, and (2) multi-modal relevance-based final demonstration selection. Experiments across eight benchmarks show MAG outperforms baselines in label-scarce settings, achieving significant gains with limited pseudo-labeling budgets while reducing MLLM inference costs.
in-context learningmulti-modal large language modelsdemonstration selectionsemi-supervised learningmanifold guidance
Perturbation-based Regional Interpretability through Subtraction Mapping (PRISM): naming-error dissociations in language models and post-stroke aphasia
The study introduces PRISM (Perturbation-based Regional Interpretability through Subtraction Mapping), a falsifiable framework for testing functional specialization in transformer language models by adapting subtraction analysis from human neuroimaging. PRISM maps error categories in LLaVA-1.6-Vicuna-13B through layer perturbations and compares results with lesion patterns in 213 post-stroke aphasia patients using correlation-difference lesion-symptom mapping. Both analyses revealed a robust phonemic-favoring dissociation in deep layers and frontal-perisylvian cortex, replicating across splits, while semantic-favoring trends were non-significant.
mechanistic interpretabilitysubtraction analysislesion-symptom mappingtransformer modelsaphasia
Federated Compositional Muon Optimizer for Matrix-Wise Models
The authors propose FedCoMuon, a federated compositional optimizer for matrix-wise models, addressing limitations in hierarchical structured problems. The method combines compositional gradient tracking with orthogonalized momentum, and introduces a variance-reduced variant (FedCoMuon-VR) using momentum-based techniques. Theoretical analysis under non-i.i.d. and non-convex settings shows FedCoMuon-VR achieves $O(ε^{-3})$ sample complexity for $ε$-stationary solutions, outperforming existing FedMuon algorithms. Experiments on robust federated learning and task-distributed risk-sensitive meta learning demonstrate superior accuracy compared to compositional baselines.
federated optimizationcompositional gradientmatrix-wise modelsvariance reductionnon-convex optimization
Efficient Hessian-Free Methods for Multi-Objective Bilevel Optimization with Nonconvex Lower Level
The authors propose Multi-Objective Moreau Envelope based Hessian-free Algorithms (MOMEHA) for multi-objective bilevel optimization with nonconvex lower level problems, addressing a gap in existing methods limited to convex cases. MOMEHA transforms the problem into a single-level optimization using Moreau envelopes and employs smooth weighted Tchebycheff scalarization, maintaining single-loop and Hessian-free properties. A momentum-based variant (MB-MOMEHA) handles stochastic settings. Theoretical convergence guarantees are provided, and experiments on few-shot meta-learning and neural architecture search demonstrate superior Pareto front performance compared to existing approaches.
multi-objective optimizationbilevel optimizationmoreau envelopehessian-freetchebycheff scalarization
A Contract-Grade Verifier for LLM-Generated GPU Kernels, and a Native Blackwell Backward for the Gated-Linear-Recurrence Family
The authors introduce a contract-grade verifier with twelve adversarial gates to rigorously assess correctness of LLM-generated GPU kernels, addressing limitations of existing loose testing methods. The verifier evaluates 2,638 machine-generated kernels previously accepted as correct, identifying 39.5% as fundamentally broken and 62.1% with at least one violation. Validation includes positive controls, threshold calibration, benchmark agreement, and manual audits. Additionally, the verifier confirms correctness of a native Blackwell backward kernel for the gated-linear-recurrence family, enabling training of five family members. Results reveal significant gaps in current correctness signals for kernel generation.
gpu kernelscontract-grade verifieradversarial gatesgated-linear-recurrencetolerance-free
The Impact of Temporal Context Length and Encoding Strategies on Self-Supervised ECG Representation Learning
The study investigates how temporal context length and encoding strategies affect self-supervised ECG representation learning, using the Icentia11k dataset. It varies input horizons (16s to 10min) and front-end representations (continuous convolutional patch embeddings vs. vector-quantized tokens) while maintaining a Transformer backbone. Results show that longer temporal contexts (5-10min) improve downstream rhythm detection and patient-level retrieval, and continuous embeddings outperform discretized tokens, suggesting quantization loses clinically relevant details. Findings advocate for extended context and continuous encoders in ECG foundation models.
electrocardiogramself-supervised learningtemporal contextvector quantizationtransformer backbone
Finding the Needle in a Haystack: Test-Time Analog Circuit Representation Adaptation for Bayesian Optimization
The paper introduces Test-Time Analog Representation Adaptation for Bayesian Optimization (TTARO), a deep-kernel BO framework that dynamically adapts circuit representations during optimization. TTARO jointly learns a nonlinear feature transformation and Gaussian-process surrogate using observed figure-of-merit labels, updating both after each evaluation to align the search space with the objective. Experiments show TTARO reduces regret AUC by 15.2% versus standard BO and 20.7% versus Deep Kernel Learning across 40 settings, with improvements up to 46.7%.
bayesian optimizationanalog circuitsrepresentation learninggaussian processdeep kernel learning
A Local-Linearly Convergent Algorithm for Nonconvex Equality-Constrained Optimization
The paper extends the Gradient-Eigenstep Algorithm for nonconvex equality-constrained optimization, proving local-linear convergence when initialized near a strong second-order stationary point with appropriate step-size and penalty parameters. The method minimizes Fletcher's augmented Lagrangian, reducing to gradient descent under these conditions. As an application, the algorithm serves as an efficient subproblem solver in progressive sampling strategies for large-scale problems, improving worst-case sample complexity compared to full-sample approaches.
nonconvex optimizationequality constraintsaugmented lagrangiangradient descentprogressive sampling
Evaluating AlphaEarth Foundations Embeddings for Wildfire Susceptibility Mapping
The study evaluates AlphaEarth Foundations (AEF) geospatial embeddings for wildfire susceptibility mapping, demonstrating their potential to reduce reliance on task-specific feature engineering. Using Victoria, Australia (2017-2025) as a case study, AEF embeddings accurately reconstruct traditional wildfire variables and achieve ROC-AUC >0.92 in downstream models. Embedding-based models show strong near-region transferability, with ROC-AUC improving by 4% in Canberra and declining by 2% in Western Sydney-Blue Mountains, outperforming physical-variable models by ~25%. Findings support scalable workflows for agencies and insurers.
geospatial embeddingswildfire susceptibilityroc-aucfeature engineeringtransferability
Training Under Challenge: Executable Certificates and Challenge-Closed Optimality for Neural Networks
The paper introduces Training Under Challenge, an executable-certificate framework for neural network optimization that constructs architecture-valid alternatives to empirically bound global-optimality gaps. The method employs predeclared procedures to generate lower-valued candidates as replayable witnesses, with coverage mechanisms required for global conclusions. Theoretical analysis shows a first-order ReLU trainer can converge to non-global points without coverage, while empirical results on a ResNet-18 distillation task demonstrate realized-residual bounds within 1.74--3.02x of true gaps, with paired predictive certificates diagnosing representation issues.
executable certificatesglobal-optimality gapchallenge-power modulusrealized-residual boundsrepresentation insufficiency
Excess Separability: Nuisance-Controlled Residual-Stream Probing for Benchmark Contamination Detection
The paper introduces 'excess separability', a method for detecting benchmark contamination in transformer models via nuisance-controlled residual-stream probing. The protocol contrasts probe accuracy depth profiles against placebo baselines, using label-permutation null tests and a reference set twice the suspect set size. Experiments show baseline depth profiles are non-flat (spanning up to 29.1 accuracy points), with correction effectiveness correlating 0.87 with surface differences between item sets. The method correctly returns null on 4 Pile arms but remains inconclusive on temporal splits where exchangeability fails.
benchmark contaminationresidual-stream probingdepth profilelabel-permutation nullexcess separability
CAKE: Compiler-Agent Co-Design for Frontier Kernel Evolution
CAKE introduces a compiler-agent co-design framework where agents author CAKE IR, a typed hardware-explicit schedule representation exposing warp roles, memory movement, synchronization, and pipelines. The approach enables verification, cost modeling, and localized diagnostics while evolving through recurring failures becoming verifier rules and reusable tactics. On B200 GPUs, CAKE IR achieves 1.144x speedup over FlashML baseline at 80M tokens, while Kimi Delta Attention shows 2.05x geometric-mean speedup over FlashKDA. Dispatcher-backed KNN and KMeans improve performance by 1.42x-2.12x across 400+ shapes, with four kernel changes upstreamed.
compiler-agent co-designhardware-explicit scheduleverification ruleslocalized diagnosticskernel evolution
Structure-preserving uncertainty quantification for GENERIC dynamics
The authors propose Structure-Preserving Epistemic Neural Networks (S-PENNs), a framework for uncertainty quantification in scientific machine learning models with hard architectural constraints, specifically applied to GENERIC dynamics. S-PENNs attach lightweight epinets to constrained components of pretrained models, ensuring physically admissible realizations while preserving thermodynamic consistency. Combined with split conformal prediction, the method provides calibrated prediction intervals with finite-sample coverage guarantees. Evaluated on three numerical examples (harmonic oscillator, chemical motor, viscoplastic model), S-PENNs achieved thermodynamically consistent rollouts and reduced computational costs by 1-3 orders of magnitude versus deep ensembles.
structure-preserving learninguncertainty quantificationgeneric dynamicsepistemic neural networksconformal prediction
The Boolean Power of ReLU
The work establishes that ReLU-MPLang strictly outperforms $Σ$-MPLang with eventually constant activations in expressing Boolean queries on finite simple undirected graphs with Boolean node features. By analyzing graph neural networks (GNNs) with ReLU versus truncated ReLU (TrReLU) and identity activations, the authors prove ReLU-GNNs exhibit strictly greater expressivity for Boolean queries. This resolves an open problem regarding the comparative power of ReLU-based versus TrReLU-based architectures in graph-based Boolean query tasks.
relu-mplangboolean queriesgraph neural networksexpressivityactivation functions
Drive-to-Music: Context-Aware Generative Audio for In-Vehicle Experiences
Drive-to-Music introduces a context-aware system for real-time music generation in vehicles, enhancing driver experience through adaptive audio. The method integrates dashcam imagery and telemetry to extract scene semantics and driving context, maps these to musical descriptors, and conditions generative audio models for low-latency synthesis. Results demonstrate feasibility with smooth transitions and safety-constrained generation, enabling personalized in-vehicle soundtracks.
context-aware generationmultimodal driving signalsmusical descriptorslow-latency synthesisconstraint-based controls
From Visual Widgets to UI Code: Efficient Tool-Grounded Generation
WidgetGen introduces a tool-grounded framework for screenshot-to-code generation, addressing the fidelity--efficiency trade-off in existing systems. The method extracts text and color evidence, performs layout/chart reasoning, and generates executable JSX without fixed UI schemas. Evaluated on 1,000 widgets, it outperforms direct prompting and Widget2Code in visual reconstruction metrics (area, legibility, style) and improves Qwen-family models via fine-tuning. Results demonstrate selective evidence grounding as a viable alternative to rigid representation constraints.
widget-to-codemultimodal generationjsxtool groundingvisual reconstruction
Represent, Then Generate: Multimodal-Conditioned Time-Series Generation under Irregular Missingness
ReCoGen introduces a two-stage framework for multimodal-conditioned time-series generation under irregular missingness, addressing limitations of existing single-modality generators. Stage I employs masked autoencoders per modality to create missingness-tolerant token sequences, while Stage II uses flow-matching to fuse these with static conditions for target signal synthesis. Evaluated on AI-READI and MIMIC-III/IV for glucose and blood pressure generation, ReCoGen outperforms six baselines across 16 settings, with 13 reaching or exceeding real-signal utility. Key innovations include cross-attention over frozen encoders and dual token-plus-AdaLN static conditioning.
conditional generationmultimodal learningflow-matchingmasked autoencodertime-series synthesis
DYSANOS Generative Dynamic Smooth Arbitrage-free Non-parametric Option Surfaces
DYSANOS introduces the first generative market model for smooth, arbitrage-free option surfaces across all strikes and expiries, capable of generating multi-year paths of daily spot and option prices. The method employs a hidden state AR(1) model, detailing model setup, data pipeline, and training while examining dynamic arbitrage numerically. Evaluated on Option Metrics' IvyDB S&P Index data (2020–2025), it outperforms a pure implied-volatility PCA baseline.
generative market modelarbitrage-freeoption surfaceshidden state modelimplied-volatility pca
Prof-K: Probabilistic One-Pass Filtering for Efficient Top-k Selection
Prof-K introduces a probabilistic one-pass filtering algorithm for efficient top-k selection, addressing inefficiencies in exact and approximate methods. The method uses a random sample to estimate an adaptive threshold, streams N elements into a compact buffer, and applies an exact top-k routine, guaranteeing correctness with probability ≥1-ε. Empirical results show 1.5x-10x speedups over PyTorch topk and RadiK, particularly in large-scale, small-to-moderate-k regimes, while maintaining distribution-agnostic robustness. The approach also enables principled accuracy-speed trade-offs and demonstrates practical impact in training BatchTopK Sparse Autoencoders.
top-k selectionprobabilistic filteringadaptive thresholdsparse autoencodersone-pass algorithm
Scaling Automatic Research Agents via World Models
The paper introduces World Model RL (WMRL) to address scaling challenges in automatic research agents by replacing environment execution with a world model, reducing training bottlenecks. WMRL incorporates Online Debiasing and Inverse-Variance Denoising to mitigate bias and noise in imperfect world models, theoretically improving convergence guarantees. Empirical results show WMRL accelerates training by 3-4x across tasks and agent scales, with 4B and 9B agents outperforming larger 48B and 120B open-weight agents on benchmarks. The method also generalizes to post-training embodied VLA policies.
world model rlonline debiasinginverse-variance denoisingautomatic research agentsembodied vla policies
Analysis of Motor Signatures of Social Adaptation in Autism for Efficient Human-Centric Systems
The study proposes a computational framework to identify motor biomarkers for autism using dance imitation tasks, analyzing social context sensitivity via movement variability. Using 3D motion capture and Dynamic Time Warping, they quantified consistency and introduced the Social Context Sensitivity Index (SCSI) to measure social modulation. Neurotypical adults showed increased variability in socially-framed conditions (particularly in limbs), while autistic individuals remained consistent. A classifier achieved 79.2% balanced accuracy in group discrimination, suggesting SCSI as a robust biomarker for autism-related motor behavior and informing inclusive human-centric system design.
dynamic time warpingsocial context sensitivity index3d motion capturemotor biomarkershuman-centric systems
GENADA: efficient generative time series adversarial attack framework
The paper introduces GENADA, an efficient generative framework for adversarial attacks on time series models. The method trains a generative model to produce deceptive perturbations in a single forward pass, offering both single-step and iterative variants, addressing computational inefficiency in gradient-based attacks. Evaluations on neural time-series models demonstrate comparable attack quality to iterative baselines while reducing perturbation generation time.
generative adversarial attacktime seriesadversarial perturbationscomputational efficiencydeep learning vulnerability
Exploring Oversmoothing with Householder Matrices
The paper introduces Householder Graph Neural Network (HouseGNN), a novel architecture addressing oversmoothing in deep GNNs by employing Householder reflectors and GroupSort activation. Unlike standard GCNs, HouseGNN uses neighborhood aggregation solely to estimate reflection directions, updating node embeddings via piecewise orthogonal transformations that preserve Euclidean norms at every layer. Theoretical analysis demonstrates three key properties: (i) layer-wise norm preservation, (ii) scale and sign invariance in message passing, and (iii) dynamic pairwise node distances due to node-specific orthogonal operators.
graph neural networksoversmoothinghouseholder reflectorpiecewise orthogonaleuclidean norm
Fast Length-Squared Sampling for Positive-Semidefinite Matrices
The authors present an O(n) expected-time rejection-sampling algorithm for length-squared sampling on n×n positive-semidefinite matrices, achieving optimal runtime even for diagonal inputs. The method enables efficient column sampling without requiring precomputed column norms, benefiting sublinear-time matrix algorithms. Applications include low-rank approximation, eigenvalue approximation, Frobenius norm estimation, and robust PSD low-rank approximation, nearly matching prior complex methods.
length-squared samplingpositive-semidefinite matricesrejection samplingsublinear algorithmslow-rank approximation
When Can You Trust Offline Evaluation of Equal-Cost Top-k Allocation? A Controlled, Reproducible Benchmark and Practitioner's Guide
The study benchmarks six off-policy estimators for deterministic top-k allocation across five datasets, identifying three key pitfalls in offline evaluation. First, weak overlap depends on logger-target action alignment rather than logging sharpness, with effective sample size poorly ranking candidate policies. Second, cross-fitting outcome nuisances fails to address optimizer's curse, requiring honest policy-level splitting. Third, propensity-estimation error significantly degrades inverse propensity scoring (IPS), while doubly robust methods remain stable. Experiments use synthesized logging with propensities floored at 0.02, reducing hybrids to untuned variants. Public benchmark data is released.
off-policy evaluationtop-k allocationinverse propensity scoringdoubly robust estimationeffective sample size
Learning Under Treatment-Induced Label Indeterminacy with Expert Annotations of Counterfactual Outcomes: A Case Study in Neurological Prognostication
The study addresses clinical prediction under treatment-induced label indeterminacy, proposing a framework for model evaluation that separates certain cases (observed outcomes) from uncertain cases (counterfactual expert annotations). Using a cohort of 2,497 post-cardiac-arrest patients (1,429 uncertain), the authors develop a neural model incorporating both label types, revealing a tradeoff between uncertain-case alignment and certain-case accuracy. Results show conventional metrics (e.g., AUROC) mask critical discrepancies in Brier scores and probability estimates for uncertain cases, highlighting evaluation gaps in treatment-affected prognostication.
clinical predictionlabel indeterminacycounterfactual outcomesneurological prognosticationbrier score
Exemplar-based objective classification of gust-induced loads across multiple flight conditions
The study proposes an exemplar-based method for objective classification of gust-induced loads across multiple flight conditions, addressing the challenge of interpretability in complex aerodynamic datasets. Using machine-learned representations and summarization, the approach selects a minimal subset of significant exemplars to classify 3480 pressure-load measurements from a flying-wing model across six flight attitudes. Results identify nine fundamental response types recurring across attitudes, with transient analysis providing physical insights into fluid mechanics.
gust-induced loadsexemplar-based classificationflight conditionspressure-load measurementsfluid mechanics
Geometric and Behavioral Stratification in Transformer Residual Streams
The study identifies the prediction direction in transformer residual streams as a privileged geometric anchor, distinct from variance-based axes, that stratifies model behavior. Analyzing 18 models (7B-120B parameters, including dense and mixture-of-experts architectures), the authors demonstrate scale-invariant concentration of readout-relevant structure near this direction, while distal regions expand with model size. Geometric analysis reveals proximal regions exhibit high structure and prompt clustering, whereas distal regions are flatter and anti-discriminative. Causal interventions show prediction-proximal disruptions cause immediate divergence, while distal disruptions preserve task framing. The findings reconcile high-dimensional computation with linear readout mechanisms.
residual streamprivileged basisprediction directiongeometric stratificationlinear readout
Non-Degenerate Risk Certification for Automated Security Decisions: A Decision-Contract Theory with ATT\&CK-Aligned Triage as a Worked Instance
The paper introduces a decision-contract theory for non-degenerate risk certification in automated security systems, addressing the structural flaw where risk bounds can be trivially satisfied by complete abstention. The theory includes an error-conservation law, a label-free singleton capacity diagnostic, and a non-degenerate actionability certificate. Applied to ATT&CK-aligned alert triage with LLM-based intrusion detection, the method achieves 90.3% of configurations meeting false-attribution risk targets and 83.4% mean correct automation across 3 IDS datasets and 6 LLMs. The capacity diagnostic effectively distinguishes threshold misalignment from structural incapacity.
risk certificationdecision-contract theoryerror-conservation lawsingleton capacityatt&ck-aligned triage
RoutePack: Expert Placement and Attention-Aware Data Packing for MoE Reinforcement Learning
RoutePack introduces a hierarchical planner for MoE reinforcement learning that jointly optimizes expert placement and attention-aware data packing. The method first performs layer-wise expert rerouting using aggregate demand, then packs samples into token-capped execution rows with a projected EDP-shard-aware objective combining attention proxies and physical EP-rank peaks. Parallel population annealing searches feasible layouts while preserving constraints. Evaluated on Ling-3.0-Tiny and Ling-3.0-Flash, RoutePack improves throughput by 8.85% and 14.89% respectively, with expert rerouting contributing 3.80-10.50% and packing adding 3.98-4.86%.
mixture-of-expertsreinforcement learningload balancingtoken routingdata packing
Unifying Generative Models with Path Integrals
The paper presents a unified path integral framework for generative models, demonstrating how flow-based, diffusion-based, variational, and adversarial approaches emerge as distinct evaluation principles of a single master action. The Martin-Siggia-Rose-Janssen-de Dominicis (MSRJD) formulation separates free and interacting probability flows, enabling diagrammatic perturbation theory. A key result shows a one-loop correction reducing error from 53% to 1.6% in deterministic samplers without stochastic sampling overhead. The framework also incorporates imperfect learned scores via insertion operators and derives a response-weighted score-matching objective, while enabling symmetry-equivariant drift design through effective field theory (EFT) methods.
path integralgenerative modelsmsrjd formalismperturbation theoryscore-matching
Beyond Parameter Space: NTK-Guided Personalized Aggregation for Robust Federated Learning
LIGHTYEAR introduces a federated learning framework that performs personalized update selection in function space rather than parameter space, addressing misalignment in heterogeneous, non-IID data environments. It employs an NTK-based agreement score to characterize predictive behavior, enabling clients to select beneficial updates via peer-to-peer exchange and private validation. Aggregation incorporates regularization for stability under heterogeneity. Evaluated across five datasets and nine baselines, LIGHTYEAR consistently outperforms centralized FL and existing P2P methods.
federated learningneural tangent kernelpeer-to-peernon-iid datapersonalized aggregation
Multi-AUV Ad-hoc network-based Target Tracking: A Value Gradient Guidance Multi-Agent Diffusion Reinforcement Learning Approach
The paper proposes VGG-MADiffRL, a value-gradient-guided multi-agent diffusion reinforcement learning algorithm, and MDCA, a diffusion-based hierarchical control architecture, for cooperative target tracking in multi-AUV ad-hoc networks. The method models sonar detection and ocean disturbances as an MDP, employing a three-tier control framework (global, local, execution) with diffusion policies guided by value gradients during denoising. Experiments demonstrate faster convergence, higher tracking accuracy, and smoother training dynamics compared to existing MARL approaches.
multi-agent reinforcement learningdiffusion policiesautonomous underwater vehiclesvalue gradient guidancemarkov decision process
Task- and dataset-specific information in protein language models
The study investigates layer-specific information in protein language models (PLMs) across 15 downstream tasks from 11 datasets, revealing that last-layer embeddings rarely yield optimal performance. Using probe models and latent-space analysis on 13 PLMs, the authors found task- and dataset-specific patterns: residue-level tasks benefit from progressive layer-wise improvement, while whole-protein tasks depend on dataset characteristics (e.g., deep mutational scan data favors shallow layers, natural proteins favor deeper layers). Performance notably declines for artificial protein tasks.
protein language modelslatent-space embeddingsdownstream tasksdeep mutational scanlayer-wise analysis
MARCH: Scaling Recurrent Memory with Content-Routed State Anchors
The paper introduces MARCH (Memory-Anchor Routing across Context History), a recurrent architecture that scales state-space models beyond fixed dimensions while maintaining computational efficiency. MARCH periodically caches recurrent-state checkpoints as content-conditioned anchors, enabling a growing memory bank with controllable resolution-cost tradeoffs. It computes outputs via attention-style aggregation over historical anchors. Evaluations on commonsense reasoning (LongBench) and in-context retrieval show MARCH outperforms linear attention variants, demonstrating improved long-range memory without sacrificing efficiency.
recurrent memorystate-space modelscontent-routed cachinglong-context retrievalattention aggregation
AI-Driven Multiscenario Interest Rate Forecasting: A Proof of Concept for Banking Asset Management
The study introduces an AI-driven prototype for multiscenario interest rate forecasting, combining Bayesian vector autoregression (BVAR) with topic modeling, sentiment analysis, and market-based analyses. The system integrates diverse forecasting approaches into an interactive platform, enabling early identification of monetary policy trends and sentiment signals from financial documents and market data. Tested in a major European bank, it enhances strategic decision-making in Asset-Liability Management (ALM) by providing transparent, interpretable, and multiperspective scenario analyses. Results demonstrate improved transparency, evidence-based decision-making, and risk management, though further optimization is needed for real-time data integration and regulatory compliance.
bayesian vector autoregressionasset-liability managementtopic modelingsentiment analysisscenario analysis
Which Site, and When: A Free-Satellite-Data Test of Himalayan Glacial Lake Bursts, Landslides, and Ice Floods
The study proposes predictive models for three Himalayan glacial hazards—moraine/ice-dammed bursts, rainfall-triggered landslides, and small glacier floods—using free satellite data (radar interferometry and weather signals). It evaluates models via spatial cross-validation, withholding entire map tiles to avoid spatial bias. Results show antecedent weather predicts triggers with ROC AUCs of 0.73–0.83, while terrain-based susceptibility scores drop sharply (0.54–0.76) when controlling for environmental confounders. Gradient-boosted baselines outperform five deep-learning models, with marginal gains only for landslides. A Nepal watchlist prioritizes high-risk sites.
glacial-lake outburstradar interferometryspatial cross-validationgradient-boostedroc auc
📰 Industry Media (5)
Z.ai Ships GLM-5.3 Without Retraining the Base Model: Better at Complex Coding and Long-Horizon Tasks
Z.ai releases GLM-5.3, a 743B-parameter language model achieving improved performance on complex coding and cybersecurity tasks without retraining the base model. Gains stem from scaled post-training with expanded task environments and longer training. On Terminal-Bench 3.0, performance rises from 4.6 to 28.3; CyberGym accuracy reaches 84.5%, surpassing Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%). ExploitBench scores more than double to 54.4%. Weights will be released after safety evaluations, currently available via API for select use cases.
post-training scalingterminal-benchcybergymexploitbenchvulnerability discovery
Meet Needle 2: An Open 45M-Parameter Tool-Calling Model That Ships as a 14MB Binary and Runs a Full Session in 28MB of RAM
Cactus Compute introduces Needle 2, a 45M-parameter tool-calling model optimized for resource-constrained devices, achieving 500 tokens/sec on a Raspberry Pi 5. The model employs a Simple Attention Network with Hadamard MLP, GQA attention, and engram KV memory, compressing to 14MB (28MB RAM at runtime) via CQ2-bit quantization. Evaluated on function-calling benchmarks, it achieves 63.7% accuracy on Mobile Actions and 93.4% well-formed output rate on BFCL v4, outperforming larger models in domain-specific tasks. The architecture avoids world knowledge dependencies, focusing on structured output generation.
tool-calling modelhadamard mlpcq2-bit quantizationgqa attentionengram kv memory
Create a Reasoning-Focused LLM: A Practical Guide to Streaming, Curating, and Fine-Tuning the SupraLabs Reasoning Corpus
The work presents an end-to-end pipeline for creating reasoning-focused language models by curating and fine-tuning on the SupraLabs reasoning corpus. The method involves streaming 8,000 samples from Hugging Face Hub, applying quality filters (token length 200-3000, reasoning ratio >0.15), and converting to chat format with explicit reasoning tags. Using LoRA (r=16) on SmolLM2-135M-Instruct, the model achieves improved reasoning capabilities while maintaining parameter efficiency (135M parameters). Results show successful adaptation across code, math, and logic tasks, with evaluation loss metrics and qualitative generation examples demonstrating coherent reasoning traces.
reasoning corpuslora fine-tuningparameter-efficientchat formattingdata curation
Samsung health AI models analyse wearable biosignal data
Samsung Research America introduced two foundation models for wearable biosignal analysis: xMAE and HiMAE. xMAE employs cross-modal reconstruction to learn temporal relationships between ECG and PPG signals, pretrained on 9,400 hours of data, outperforming baselines in 15/19 tasks including cardiovascular prediction. HiMAE uses hierarchical masked autoencoding to capture multi-scale temporal patterns in wearable data, enabling on-device inference (<1ms latency) for classification and prediction tasks. Both models leverage self-supervised learning on unlabeled biosignals to enable downstream health applications without continuous cloud connectivity.
foundation modelsbiosignal analysisself-supervised learningmasked autoencoderwearable computing
Google AI health coach to use Abbott glucose data
Google and Abbott announced a multiyear partnership integrating Abbott's Lingo continuous glucose monitoring (CGM) data with Google's Gemini-powered Health Coach. The system combines interstitial fluid glucose measurements (14-day wear, Bluetooth-transmitted) with activity, sleep, and medical record data via Health Connect APIs to provide non-medical wellness recommendations. A planned large-scale real-world study will analyze metabolic health correlations to refine the AI coach, with Lingo integration launching in Google Health later this year. The FDA-cleared Lingo CGM (non-insulin users) differs from Abbott's FreeStyle Libre 2 (diabetes management) in intended use.
continuous glucose monitoringinterstitial fluidhealth connect apimetabolic healthgemini ai
Generated automatically at 2026-08-14 20:07 UTC. Summaries and keywords are produced by an LLM and may contain inaccuracies — always consult the original article.