Translate
Live intelligence

Latest AI updates

1,750 updates
AllModelsProductsIndustryResearchTutorialOpinion
Research Editor selected

Spectral Degeneration of Self-Attention under Isotropic Score Geometry

This paper introduces an idealized spectral model for the pre-softmax score matrix in transformers, motivated by the rotational isotropy of query and key representations. The authors prove that the row-wise softmax map transforms this structure into a degenerate attention matrix where dominant modes are preserved while bulk spectral distributions concentrate at zero. The stable rank remains bounded for bidirectional attention but grows logarithmically for causal attention, explaining how attention compresses non-dominant directions.

Why it matters: Provides a rigorous theoretical explanation for how self-attention preserves dominant directions while compressing bulk modes, offering fundamental insights into transformer behavior.

arxiv5d agoScore 88
transformer theoryattention mechanismsspectral geometry
Research Editor selected

GenDB-V: AI-Generated and Formally Verified Database Systems

GenDB-V enables AI agents to synthesize database engines that are specialized for performance on given hardware and workloads while being formally verified for correctness. The system uses SQLSpec to formalize SQL semantics, joint synthesis of code and proofs, and a three-layer modular proof architecture. Evaluated on TPC-H and SEC-EDGAR, GenDB-V achieves geometric-mean speedups of 22–61× over DuckDB and Umbra for verified template implementations, and 1.27–3.4× for a verified general-purpose engine.

Why it matters: Demonstrates a practical AI-driven approach to automatically generating high-performance, provably correct database engines, addressing both performance and correctness challenges.

arxiv5d agoScore 88
AI for systemsformal verificationdatabase systems
Research Editor selected

Rethinking Neural Decompilation: Candidate Selection as the Hidden Bottleneck

This paper identifies candidate selection as the key bottleneck in neural decompilation, showing that correct reconstructions are already present in LLM sampling space but poorly selected. The authors propose a generate-then-select pipeline using various selectors, with negative perplexity and execution-guided selection achieving high pass@1 rates. Fine-tuning on a new obfuscated dataset (ExeBench-Obfuscated) further improves performance on obfuscated binaries, challenging prior claims.

Why it matters: Provides actionable insights for improving neural decompilation accuracy, directly relevant to malware analysis and software security.

arxiv5d agoScore 85
neural decompilationLLM4Decompilecode obfuscation
Research Editor selected

MultiMental-IAS: A Multimodal Driver Mental State Dataset for Non-IID Evaluation

Researchers introduce MultiMental-IAS, a multimodal dataset collected from 23 participants in a CARLA-based cockpit simulation, featuring EEG, PPG, cabin video, and vehicle telemetry across six mental states. The dataset includes 148 trials totaling 11.6 hours and provides standardized evaluation protocols for within-subject and cross-subject generalization. Multimodal fusion achieves a macro-F1 of 0.94 for seen drivers and 0.485 for unseen drivers, highlighting the challenge of inter-driver heterogeneity. This resource enables research into non-IID generalization, personalization, and federated learning for driver monitoring.

Why it matters: Provides a concrete, publicly available benchmark dataset with standardized protocols for multimodal driver mental state analysis, directly addressing the practical challenge of inter-subject variability.

arxiv5d agoScore 88
driver mental state analysismultimodal fusionnon-IID generalization
Research Editor selected

DrawOffer: Evidence-Grounded Longitudinal User-State Modeling for Career Support

DrawOffer is a system that separates language-model observation extraction from deterministic inference over source-linked records to maintain uncertain ability estimates and revisable issues in career-support assistants. Evaluated on 102 fictional personas with 580 scored checkpoints, it passes 169/210 checkpoints on test personas with issue precision of 0.875 and recall of 0.636. Ablation studies show that removing answer-evidence requirements reduces passes and precision, highlighting the importance of evidence sufficiency. The work is an exploratory systems contribution with limitations in generalization and reproducibility.

Why it matters: Offers a concrete, evidence-grounded approach to improving reliability of AI career assistants, with clear evaluation metrics and ablation studies.

arxiv5d agoScore 82
AI for career supportevidence-based reasoninglongitudinal user modeling
Research Editor selected

SelfImprove-Bench: Evaluating Generalization in Self-Improving Agents

SelfImprove-Bench is a new evaluation framework for recursively self-improving (RSI) agents, organizing tasks into three core capabilities—Exploring, Planning, Reasoning—using controllable game environments. It introduces a generalization ladder (G0–G4) to measure how well capability gains transfer beyond training conditions. Experiments show that existing RSI methods achieve inconsistent generalization across these levels, highlighting the need for more robust self-improvement strategies.

Why it matters: Provides a rigorous, standardized benchmark for measuring generalization in self-improving AI systems, a critical gap in current RSI research.

arxiv5d agoScore 88
recursive self-improvementAI evaluationgeneralization
Research Editor selected

Analysing the Narrative Paradox through Formal Models and Properties in Interactive Narrative

This paper introduces a formal framework for modelling the narrative paradox—the tension between player autonomy and narrative control—in interactive narratives. It extends state-transition systems to represent player actions and system events, formalizing properties like player agency, EM controllability, author intent, and robustness. Quantitative indicators derived from simulated plans enable comparison of Experience Manager strategies, with a case study revealing trade-offs and compromise zones.

Why it matters: Provides a reusable, formal analysis framework for a core challenge in interactive narrative, enabling systematic comparison of management strategies.

arxiv5d agoScore 82
interactive narrativenarrative paradoxformal modelling
Research Editor selected

OmniOPD: Stable On-Policy Distillation for Joint Audio-Video Generation

Researchers propose OmniOPD, a modality-wise on-policy distillation framework that stabilizes joint audio-video generation by addressing cross-modal interference and token imbalance. The method uses asymmetric noise scheduling to provide richer visual cues for the audio teacher, grouped multi-noise sampling to expand audio supervision, and timestep-wise importance weighting to stabilize training. Evaluations on JavisBench and VBench show that OmniOPD preserves audio quality while significantly enhancing video generation.

Why it matters: Presents a concrete technical solution to a known instability problem in multimodal generative AI, with clear methodology and benchmarks.

arxiv5d agoScore 88
OmniOPDjoint audio-video generationon-policy distillation
Research Editor selected

Creating Hybrid Interactive Narrative Systems with LLM: From Formal Description to Evaluation

The paper introduces an LLM-based narrative platform where interchangeable Experience Managers operate over a shared symbolic world model of facts, operations, and goals. It defines four evaluation properties—player agency, EM controllability, author intent, and robustness—and compares three managers (open generation, Monte Carlo repair, autonomous characters) on a Little Red Riding Hood scenario. Results show no single manager dominates, providing a principled basis for selecting managers based on author priorities.

Why it matters: Provides a formal framework and comparative evaluation for hybrid LLM-based interactive narrative systems, directly useful for researchers and designers of narrative AI.

arxiv5d agoScore 88
interactive narrativelarge language modelshybrid AI systems
Research Editor selected

Offline Approximate Value Iteration: A Reinforcement Learning Approach for Satellite Constellation Deployment

Researchers introduce a reinforcement learning-based approximate dynamic programming method to optimize Walker-type satellite constellation deployment under uncertainties like launch failures and satellite malfunctions. The approach uses offline approximate value iteration with linear function approximation and action pruning to reduce computational demands. Preliminary results show lower expected costs and deployment times compared to a rolling-horizon baseline, demonstrating a novel application of approximate dynamic programming in space logistics.

Why it matters: Presents a concrete, novel RL application to a real-world space logistics problem with measurable improvements over baselines.

arxiv5d agoScore 85
reinforcement learningspace logisticssatellite deployment
Research Editor selected

Predictive Safety Guardrails Do Not Always Make LLM Agents Safer

A systematic study of world-model-driven predictive safety guardrails for LLM agents finds that stronger prediction involvement does not guarantee safety gains and can reduce task completion while increasing inference cost. Predictions become unreliable over longer horizons, and safety gains can persist even under corrupted predictions, revealing a fundamental prediction-to-action gap. The findings challenge the assumption that predictive foresight alone reliably improves agent safety.

Why it matters: Provides critical empirical evidence that current predictive guardrails may not work as intended, informing future safety research.

arxiv5d agoScore 92
LLM safetypredictive guardrailsworld models
Research Editor selected

SAGE: LLM-guided framework discovers hierarchical governing equations for community-structured network dynamics

Researchers introduce SAGE, a self-reflective LLM agent framework that discovers shared community-level governing laws and individual-specific dynamical variations from observational data on complex networks. The method iteratively identifies common structures, resolves residual dynamics, and adaptively expands its candidate library when needed. Tests on synthetic and real-world networks (mobile communication, brain electrophysiology, power grids, urban rail transit) show SAGE recovers compact, interpretable equations and outperforms baselines, even under incomplete observations.

Why it matters: Demonstrates a novel, practical use of LLM agents for automated scientific discovery in complex systems, with strong empirical validation.

arxiv5d agoScore 85
complex network dynamicsLLM-guided scientific discoveryhierarchical modeling
Research Editor selected

RxnMiner: A Multimodal Agent for Structured Reaction Data Extraction from Chemistry Figures

RxnMiner is a multimodal agent that combines an expert-informed extraction procedure with a chemistry toolbox to recover structured reaction records from chemistry figures. It handles molecular recognition, shared-template evidence, substituent tables, product grids, and reconstruction of indirectly specified participants. The authors also introduce ChemRxn-Bench, containing 603 reaction figures and 8,547 manually annotated records. With Qwen3.8-Flash-Next as backbone, RxnMiner raises stereochemistry-aware reaction macro-F1 from 21.0% under direct prediction to 71.0%, versus 25.0% for the strongest existing extraction system, and produces 88,202 source-linked records from 10,107 figures without human intervention.

Why it matters: It presents a concrete multimodal extraction system and benchmark that could substantially reduce manual effort in building chemistry reaction datasets for AI models.

arxiv5d agoScore 84
RxnMinerChemRxn-Benchchemical reaction extraction
Research Editor selected

Think with Skills: Pseudocode Planning Before Execution for Advanced Reasoners

The paper introduces Think with Skills, a method that trains reasoning models to convert relevant skill descriptions into task-specific pseudocode plans before execution. The pseudocode explicitly represents operations, conditions, intermediate results, and dependencies, bridging declarative skill descriptions and agent actions. Training combines supervised fine-tuning on tasks, skills, pseudocode plans, and execution trajectories with Group Relative Policy Optimization. On SkillsBench, the method raises task success from 35.0% to 62.0%.

Why it matters: It presents a concrete training approach that substantially improves how language-model agents proceduralize and coordinate expert-written skills.

arxiv5d agoScore 82
LLM agentsplanningreasoning models
Research Editor selected

OmniQ and Residual Agents: A self-improving AI scientist for computational quantum research

The paper introduces OmniQ, a virtual research environment that converts published computational quantum science papers into executable reproduction tasks, and evaluates five agentic systems across 59 papers. It finds that current AI scientists frequently fail to reproduce published results due to a mismatch between paper content and agent requirements. The authors develop Residual Agents (RA), a quantum-tailored self-improving AI scientist that distills verified workflow repairs, boosting DeepSeek V4.1 Flash from 68% to 85% accuracy at one-third the cost of Claude Code with Opus 5.

Why it matters: Provides a concrete, reproducible framework and a cost-effective solution to the reproducibility gap in AI-driven scientific research, with clear performance benchmarks.

arxiv5d agoScore 85
AI for sciencecomputational quantum researchautonomous discovery
Research Editor selected

CARE: Grounded Recursive Self-Improvement for Production Customer-Service Agents

The paper introduces CARE, a framework for recursive self-improvement (RSI) in production customer-service agents. It replaces a frontier proprietary model with an in-house mid-sized model that learns from its own failures after deployment. A 24/7 inspection process identifies failures, attributes them to model or component errors, and uses on-policy self-distillation with privileged diagnosis information to update the model. Candidates are validated via simulation tests and offline inspection before promotion, requiring parity with the deployed version and human approval.

Why it matters: Provides a concrete, technically detailed approach to continuous model improvement in production, addressing a practical challenge in deploying AI agents.

arxiv5d agoScore 88
CARE frameworkrecursive self-improvementcustomer service AI
Research Editor selected

Math Foundation Formalization Agent: A General-Purpose Approach to Formalizing Mathematical Foundations

The paper introduces MFF-Agent, a general-purpose intelligent agent that combines large language models with the Rocq proof assistant to automate the formalization of mathematical theories. It handles task classification, tool-assisted execution, verification feedback, and state-based task management for generating and verifying formal proofs. Case studies across Morse-Kelley set theory, real numbers, propositional logic, and axiom of choice equivalences demonstrate its applicability to diverse foundational domains. The results show the feasibility of a unified operational framework while clarifying the gap toward a fully integrated mathematical foundation.

Why it matters: Offers a concrete, novel approach to automating mathematical formalization across multiple foundational domains, with practical case studies and clear insights into current limitations.

arxiv5d agoScore 88
formal verificationlarge language modelsmathematical foundations
Research Editor selected

SpecOPD: Branch-Parallel On-Policy Distillation for Speculative Decoding

SpecOPD introduces a branch-parallel framework for on-policy distillation in speculative decoding, enabling draft models to be trained without sequential draft-and-verify rollouts. By packing independent branches into proposal trees and using tree masks, sparse attention, and indexed state replay, the method achieves up to 41.45% higher mean acceptance length and 16.09× supervised-token throughput over sequential-rollout OPD. Experiments across 8B to 125B parameter targets and multiple architectures show up to 3.56× generation speedup over autoregressive decoding.

Why it matters: Offers a concrete, high-impact technique to accelerate LLM inference, directly relevant to practitioners working on model deployment and efficiency.

arxiv5d agoScore 92
speculative decodinglarge language model inferenceon-policy distillation