Translate
Live intelligence

Latest AI updates

650 updates
AllModelsProductsIndustryResearchTutorialOpinion
Research

Enhancing Local and Global Structure in Compressed AI Agent Memory

The paper argues that compressing AI agent memory by preserving only nearest neighbors can distort global relationships needed for consolidating, deleting, and comparing memories. It introduces local and global fidelity measures and a locality-aware quantization (LAQ) method derived from a proven bound on far-end rank violations, applied to OPQ codebooks without changing their bytes, rotation, or search procedure. LAQ improved global fidelity in all 52 held-out runs, including the external FinDER memory, and achieved the highest global fidelity among seven equal-size quantizers in all 12 Multi-Doc-2025 runs while scaling to a million passages.

arxiv13h agoScore 76
AI agent memoryvector quantizationembedding compression
Research

VisionSIM: A Physically Grounded Multi-Modal Simulator for Robot Perception

VisionSIM is a unified simulation platform that models the raw physical measurement process for diverse sensors including event, SPAD, iToF/dToF, and thermal cameras. It reproduces modality-specific failure modes such as LiDAR pileup, iToF phase wrapping, event camera refractory drops, and thermal history-dependent emission. The platform enables principled cross-modality comparison on identical scenes and includes a large indoor dataset with synchronized raw measurements and ground-truth annotations.

arxiv13h agoScore 75
robotic perceptionsensor simulationmulti-modal sensing
Research

NPCBench: A New Benchmark for Multimodal Care-Pathway Reasoning in Nasopharyngeal Carcinoma

NPCBench is a clinical apprenticeship benchmark for nasopharyngeal carcinoma that evaluates AI models on integrated, longitudinal care-pathway reasoning rather than isolated questions. It includes 4,777 items and a 100-patient cohort with 417 scored tasks, testing evidence grounding and temporal coherence. Evaluations of 30 models reveal a composition gap: strong individual task performance does not guarantee jointly correct diagnostic and care answers. The benchmark aims to push medical AI beyond static QA toward workflow-oriented assessment.

arxiv13h agoScore 73
NPCBenchclinical reasoningmultimodal AI
Research

MOMASPLAT: Novel-View Augmentation for Mobile Manipulation Policies

MOMASPLAT is a framework that augments imitation learning demonstrations for mobile manipulation by synthesizing full trajectories from unseen robot base poses, including dynamic object interactions. It combines 3D Gaussian Splatting, feed-forward 3D foundation models, and kinematics-based action retargeting to reconstruct static scenes and manipulated objects separately. Evaluated on five real-world tasks, it more than doubles the average success rate of the strongest baseline and maintains non-trivial success at large base displacements.

arxiv13h agoScore 78
Mobile ManipulationImitation LearningNovel-View Synthesis
Research

Hybrid LSTM-XGBoost Framework Predicts Landing-Phase Safety Events from Imbalanced Flight Data

This paper introduces an interpretable hybrid framework for predicting rare landing-phase safety events from imbalanced flight data. It fuses LSTM autoencoder reconstruction errors with statistical flight features and feeds them into a class-weighted XGBoost classifier. Evaluated on 53,750 NASA DASHlink approach windows, the method achieves mean macro recall above 0.90 across airports, with component comparisons and feature attribution analysis. The study offers practical guidance for adapting flight-data monitoring models to airport-level data availability.

arxiv13h agoScore 70
aviation safetymachine learninganomaly detection
Research

Learning a patient world model for ICU health system simulation

Researchers propose a neural network-based patient world model trained on longitudinal EHR data (MIMIC-IV) to simulate ICU operations across multiple policy questions. The model captures history-dependent dynamics that compact state models miss, such as a 24-hour recovery difference for patients with different prior trajectories. It matches Markov and semi-Markov comparators in aggregate occupancy and policy responses, while enabling scenario reconfiguration without model rebuilding.

arxiv13h agoScore 75
patient trajectory modelingICU operationshealthcare simulation
Research

StepBack: Step-Level Error-Localized Resampling for Efficient Test-Time Reasoning

StepBack is a new decoding strategy that improves test-time reasoning efficiency by localizing likely first errors in draft reasoning trajectories and resampling only the error step and its suffix, followed by majority voting. It matches the accuracy of standard self-consistency across seven math and multi-hop QA benchmarks while reducing token usage by up to 41.26%. The method can also be combined with adaptive inference methods like Adaptive Consistency for further efficiency gains.

arxiv13h agoScore 70
efficient inferenceLLM reasoningself-consistency
Research

ProofTrace: Detecting AI-Generated Proofs via Structural Signatures

This paper introduces ProofTrace, a framework that represents mathematical papers as graphs of immediate proof support to detect AI involvement in proofs. Using five cohorts of 8,281 papers, the authors identify structural signatures of AI-generated proofs, such as fewer external citations and altered graph connectivity. ProofTrace distinguishes human-written papers from those with substantive AI use and remains stable under AI rewording, unlike text-based detectors.

arxiv13h agoScore 70
AI in mathematicsdetection of AI-generated contentgraph-based text analysis
Research

Terminal-Bench 2.1: Continuous Benchmark Maintenance Reveals Stale Evaluations

The paper introduces the concept of benchmark maintenance debt, where benchmarks become stale as models improve and infrastructure changes. The authors propose a continuous-validation workflow and apply it to Terminal-Bench 2.0, finding issues in roughly 30% of the benchmark and releasing repairs as Terminal-Bench 2.1. These repairs change evaluation results, with accuracy increasing by over 10 percentage points and leaderboard positions shifting by up to three ranks, while overall ranking remains highly correlated.

arxiv13h agoScore 79
benchmark maintenanceLLM evaluationcoding agents
Research

VLA-Tug: Lightweight Latent Probe Improves Early Failure Detection in Vision-Language-Action Models

Researchers propose VLA-Tug, a lightweight probe over latent layers of Vision-Language-Action models that improves early failure detection and bounds missed-detection and false-alarm rates with minimal inference overhead. It increases early detection by up to 30% over the strongest baseline and maintains valid bounds in up to 86% of runs under task shift. VLA-Tug also predicts failure modes, boosting top-3 accuracy by up to 44% over a majority baseline, and includes a dataset of 3,180 annotated manipulation episodes.

arxiv13h agoScore 77
Vision-Language-Action modelsfailure predictionrobotics safety
Research

MedGraph: Graph Memory for Long-Term Medical Dialogue

MedGraph is a graph-based memory system for LLM medical assistants that tracks patient state across visits. It addresses three common memory errors: outdated evidence, mixing other people's health facts, and missing safety facts. On MedMemoryBench, MedGraph outperforms five baselines by 10.5 and 10.4 points with GPT and Qwen backbones, respectively.

arxiv13h agoScore 72
medical AImemory systemsdialogue systems
Research

ReproBridge: Agent Workflow Boosts ML Reproduction Success to 92.6%

ReproBridge introduces a standardized artifact design and agent workflow to improve reproducibility of ML experiments. On a new benchmark of 1,029 tasks from 36 papers, refined artifacts raised mean reproduction success from 76.4% to 92.6% across four models and cut API costs by 19.7%. The approach narrows performance gaps between models, aiding lower-performing agents.

arxiv13h agoScore 70
machine learning reproducibilityAI agentsbenchmarking
Research

MAEMs: Scaling Activation Inversion to Interpret Residual Stream Directions

Researchers introduce Maximally-Activating Example Metamodels (MAEMs), which generate text exemplars that strongly activate specific residual stream directions in a subject language model. Using a label-free reinforcement learning objective, MAEMs scale to diverse sources of directions and outperform corpus search and natural-language autoencoders on held-out activations. The approach also recovers LoRA backdoor payloads and verbalizes model inner monologue, with exemplars verifiable via forward passes.

arxiv13h agoScore 79
AI interpretabilitylanguage modelsactivation engineering
Research

Finite Experiments Can Certify Integrated Information: Identifiability and Limits

This paper investigates when mechanism-level quantities like integrated information (phi) can be identified from finite, partial, and noisy interventions. Using stochastic networks of four and five units, the authors show that behavior alone can be silent, while internal recordings can separate mechanisms within a few trials. They also identify precision floors and sharp edges that limit certification, yet prove that one-sided lower bounds can be certified from about 10^7 controlled transitions with controlled error probability.

arxiv13h agoScore 70
integrated information theorymechanism-level analysiscausal identifiability
Research

XStack: A Benchmark for Migrating ML Repositories Between PyTorch and JAX

XStack is a new benchmark that evaluates AI agents on migrating entire machine learning repositories from PyTorch on NVIDIA GPUs to JAX on Google TPUs. The benchmark requires agents to deliver a working JAX repository that meets grading criteria based on the source repository's published results. Current agent performance is low, with the best agent (Claude Opus 5) achieving only 75% pass rate, highlighting the difficulty of cross-framework and cross-hardware migration.

arxiv13h agoScore 79
XStackML repository migrationPyTorch to JAX
Research

Scientist World: Turning Research Papers into Executable Environments for AI Scientists

Scientist World is a new platform that converts research papers into executable environments, allowing AI scientists to reproduce methods and tackle open-ended questions. The authors release 1,000 verified environments and report that frontier AI scientists surpass human researchers 74% of the time. The platform also provides 6,400 hours of research traces and demonstrates a 158% improvement in research ideation for a small model.

arxiv13h agoScore 79
AI researchscientific discoverybenchmarking
Research

SwarmEvolve: Latent-Space Agent Swarms Speed Up Evolutionary Program Search

SwarmEvolve is a new framework that uses hundreds of low-cost LLM agents per round to perform evolutionary search for mathematical and computational problems. Instead of decoding to text, agents reason in latent space and share information via KV caches, with a novelty filter based on hidden-state distances and crossover via combined caches. On 21 problems, it achieves the highest average score and reaches top results in 3-8 minutes versus 60 minutes for other frameworks, even improving on AlphaEvolve's solutions with a smaller model.

arxiv13h agoScore 79
SwarmEvolveLLM-guided evolutionary searchlatent space reasoning
Research

No AI agent scores above 12 percent on SemiAnalysis' new real-world work benchmark

SemiAnalysis has introduced a new benchmark called GDPx that evaluates AI agents on simulated knowledge work tasks, spanning domains like chip design and quant trading. According to the benchmark, no model achieves more than 12 percent, with Claude Opus 5.5 leading slightly ahead of Gemini 4 Carbon. Critics note the benchmark has a small number of tasks per domain and uneven testing conditions.

the-decoder1d agoScore 70
AI benchmarkingAI agentsSemiAnalysis