Translate
Live intelligence

Latest AI updates

1,609 updates
AllModelsProductsIndustryResearchTutorialOpinion
Research Editor selected

GraphRail: Structured Claim Verification for Multi-Hop QA Reasoning

GraphRail is a new verification framework that represents requirements, evidence, and reasoning as structured claims and compares them without using a generative judge or gold answers. Across three multi-hop QA benchmarks and three reasoning frameworks, it improves error recall over generative verifiers while remaining stable under stochastic extraction. Its diagnoses can also guide online reasoning, providing a traceable way to detect and correct intermediate errors.

Why it matters: This is a concrete research contribution with benchmarked results and a novel verification approach, offering practical value for improving reasoning reliability in AI systems.

arxiv14h agoScore 82
AI verificationmulti-hop question answeringreasoning frameworks
Products

AgentTrustCI: Build-Time Policy, Grounding, and Trajectory Tests for Java LLM Agents

AgentTrustCI is a Java/JUnit 5 framework that adds application-layer behavioral checks for LLM agents, covering tool selection, answer grounding, and workflow trajectories via recorded traces. It uses deterministic replay for repeatable build-time tests and an optional judge interface for semantic evaluation. In a synthetic corpus of 55 cases across seven enterprise domains, the full pipeline correctly classified all 30 injected faults and 25 passing controls, including 10 non-trivial controls, with no missed critical cases. The paper emphasizes that results reflect the authored regression suite, not production accuracy, and describes a production-trace validation protocol.

arxiv14h agoScore 78
LLM agent testingbuild-time validationbehavioral contracts
Research Editor selected

LifeVal: A Benchmark for Personalized Agentic Consumer Tasks

LifeVal is a new benchmark that evaluates AI models on personalized agentic consumer tasks, covering 148 categories with an estimated $1.7 trillion in annual willingness-to-pay. It is built from a survey of over 3,800 consumers and cross-referenced with time-use studies, with tasks created by experts and quality-controlled. The benchmark aims to measure performance on high-value real-life tasks like managing finances and job searching.

Why it matters: This introduces a novel, data-driven benchmark for a growing area of AI, offering concrete methodology and economic grounding that is valuable for researchers and developers.

arxiv14h agoScore 82
AI benchmarkspersonalized assistantsagentic AI
Research

PLUS: A Detector-Agnostic Subspace Signal for Near-OOD Detection

The paper introduces PLUS, a post-hoc signal that complements existing out-of-distribution (OOD) detectors to improve near-OOD detection. PLUS builds per-class low-dimensional subspaces from sparse autoencoder decoder directions and scores inputs via whitened squared Mahalanobis distance. Fused additively with any calibrated base score at a fixed global weight, it improves near-OOD AUROC across all 126 detector-backbone combinations tested, while largely maintaining far-OOD performance.

arxiv14h agoScore 78
out-of-distribution detectiondeep learninganomaly detection
Research

ALTO: A New Multilingual Dataset for Tech-Facilitated Gender-Based Violence Detection

Researchers introduced ALTO, a crowdsourced and annotated dataset of 3,022 instances of tech-facilitated gender-based violence (TFGBV) in Levantine Arabic, African French, and Swahili, developed with civil society organizations. Benchmarking shows existing harmful content detection models have poor TFGBV recall, especially for covert harms and in Swahili, but fine-tuning improves performance. The dataset, taxonomy, and models are released to support better detection in underrepresented languages.

arxiv14h agoScore 78
gender-based violenceNLPdataset creation
Research

Beyond the principle-practice gap: Responsible AI as situated epistemic and ethical practice

This paper challenges the common framing of the gap between AI ethics principles and practice as a mere implementation or compliance deficit. Based on interviews with actors in AI-enabled anti-trafficking systems, it argues that ethical concerns are shaped by institutional and professional contexts, and that epistemic choices are intertwined with ethical judgments. The authors propose a model of four moments—perception, judgment, situated action, and socio-technical materialization—through which concerns become actionable or are deferred. They conclude that responsible AI governance requires epistemic infrastructures to make affected communities' knowledge receivable in design.

arxiv14h agoScore 72
AI ethicsresponsible AIinstitutional practice
Research Editor selected

Rel-Bloom: Multilingual Benchmark Reveals LLM Reasoning Gaps over Supplied Relational Facts

This paper introduces Rel-Bloom, a controlled multilingual diagnostic of 450 multiple-choice questions built from 30 Wikidata relations, testing evidence-conditioned relational reasoning across five Bloom-inspired levels and three veracity conditions. Across 43 models, six languages, and 226,800 predictions, it finds sharp accuracy drops from Apply to Analyze in open-weight models, persistent bottlenecks on Evaluate items even for strong models, and cross-language consistency in difficulty rankings. The results show that greater model capability does not guarantee reliable reasoning over supplied relational evidence.

Why it matters: Provides concrete, recent benchmark results and insights into LLM reasoning limitations over non-parametric memory, valuable for model evaluation and development.

arxiv14h agoScore 85
Retrieval-augmented generationMultilingual LLM evaluationRelational knowledge reasoning
Research Editor selected

Full-Stack Survey of Dexterous Robotic Hands in the Foundation-Model Era

This survey reviews dexterous robotic hand research from January 2024 to early November 2026, organizing the field into five layers: hardware, data, cross-embodiment transfer, learning, and evaluation. It highlights that human data value depends on kinematic, visual, and contact alignment, and that foundation policies still face embodiment-specific constraints. The paper identifies eleven open problems and proposes reporting standards for more measurable progress.

Why it matters: Provides a structured, current synthesis of a rapidly evolving area, offering actionable insights for researchers and practitioners.

arxiv14h agoScore 85
robotic manipulationfoundation modelsembodiment
Research Editor selected

ConceptARC at Scale: New Benchmark and Analysis of Conceptual Understanding in Reasoning Models

The paper introduces ConceptARC at Scale, a generator that creates thousands of ARC-style tasks targeting specific spatial and semantic concepts, and presents ConceptARC-100, a dataset of 1,600 manually reviewed tasks across 16 concepts—ten times the size of the original benchmark. Evaluations of recent large reasoning models and models trained from scratch show which concepts are solved reliably and where they struggle. Internal representation analysis reveals that task representations cluster by transformation operations (e.g., recoloring, removal, translation), suggesting models form stable abstractions that generalize across diverse tasks.

Why it matters: Provides a new large-scale benchmark and insights into how reasoning models abstract concepts, valuable for AI researchers and practitioners.

arxiv14h agoScore 85
Abstraction and Reasoning CorpusLarge Reasoning ModelsConceptual Understanding
Research Editor selected

Smartphone-Aided VR Exergames for HIIT: Real-Time Pose Estimation and AI Feedback

This paper presents a VR exergame system for HIIT that uses a standard smartphone camera for full-body tracking, processed via real-time pose estimation and analyzed by an MCP-based backend with GPT-4o. A 20-participant study found higher enjoyment, immersion, and recommendation likelihood in VR versus non-VR, with AI feedback agreeing with expert evaluation at 87.5%.

Why it matters: Offers a cost-effective, scalable approach to VR fitness with validated AI feedback, relevant to both fitness tech and AI applications.

arxiv14h agoScore 82
Virtual reality fitnessAI-driven feedbackPose estimation
Research

TailSteer: Trajectory-Tail Steering for Drift-Resistant Multi-turn Image Editing

TailSteer is a final-stage training framework that improves multi-turn image editing robustness by concentrating preservation corrections near the trajectory tail of the denoising process. It learns from the editor's own on-policy rollouts using unchanged source content as reference, avoiding curated ground-truth pairs. Experiments on FLUX, SenseNova, and Qwen families show reduced cumulative drift over 15-20 turns while preserving base-model performance.

arxiv14h agoScore 78
image editingmulti-turn robustnessfoundation models
Research

SARBLOH: Spatial Perception Engine Boosts ARC-AGI-3 Scores 13.7x

A new open-weight agent, SARBLOH, demonstrates that spatial perception, not planning, is key to interactive reasoning on ARC-AGI-3. Its zero-parameter perception engine converts raw grids into object inventories and geometric briefings, raising rubric scores from 3/28 to 25/28 and cutting token use by 89.4%. On hidden Kaggle evaluation, it scored 16.53, 13.7x the first Milestone winner, within a nine-hour budget.

arxiv14h agoScore 74
ARC-AGI-3spatial reasoninginteractive agents
Research

Message-Passing Graph Persistence (MGP) enhances GNN expressivity with topology-aware descriptors

The paper introduces Message-Passing Graph Persistence (MGP), a family of topological descriptors that integrates message passing into graph filtrations, updating node embeddings at each step and linking them to persistence events. MGP is proven to be invariant under graph isomorphism and strictly more expressive than existing PH-based descriptors. Experiments on expressivity benchmarks and graph classification/regression datasets show MGP outperforms other PH-based descriptors on most datasets.

arxiv14h agoScore 74
Graph Neural NetworksTopological Data AnalysisPersistent Homology
Research

Task-Efficient Agent Evaluation Using Cached Traces

This paper introduces a method to estimate an agentic system's benchmark performance from a small number of task executions by leveraging rich execution traces. The approach extracts structured summaries with an LLM and embeds them to create representations that preserve task-related signal while suppressing identity-related noise. On SWE-bench Verified and Terminal-Bench 2.0, combining this geometry-based prediction with an item-response model reduces mean absolute error significantly compared to baselines.

arxiv14h agoScore 75
AI evaluationagentic systemsefficient benchmarking
Research

Training and Calibration Data for Open Models on ARC-AGI-2

This paper investigates which training and calibration data help three open models solve ARC-AGI-2 tasks offline. The authors built a corpus of executable rules and synthetic tasks, but found that fine-tuning on real task records outperformed synthetic data. For a reasoning model, an inference procedure contributed most of the score gain, while expert pruning and routing statistics improved a mixture-of-experts model's performance.

arxiv14h agoScore 79
ARC-AGI-2open modelstraining data
Research

Metamodels Forecast LLM Answer Activations Before First Token

Researchers propose context-answer metamodels that map context activations to forecasts of mean-pooled answer activations, capturing variability across stochastic rollouts. The metamodels achieve high fidelity, scale with training data up to 500,000 prompts, and generalize zero-shot to out-of-domain datasets. Forecasts retain behavior-predictive signal and transfer across models, suggesting scalable pre-generation behavior prediction.

arxiv14h agoScore 75
language model behavior predictionactivation forecastingefficient evaluation
Research

ServerBench: Benchmarking Coding Agents for Designing Specialized LLM Serving Systems

ServerBench is a new benchmark that evaluates coding agents' ability to design complete, specialized LLM serving systems. Agents are scored on a dollars-per-dollar metric, balancing latency-qualified token value against GPU cost. In tests across 11 coding-agent models, Claude Sonnet 5.5 achieved 2.97 dollars per dollar on Gemma E4B, outperforming the vLLM reference by 4.87 times. The benchmark includes stateful replay and independent verification to ensure correctness.

arxiv14h agoScore 77
LLM serving systemsagentic software designbenchmarking
Research

CompressionBench: Benchmarking Coding Agents for Autonomous LLM Compression

CompressionBench evaluates whether coding agents can autonomously compress pretrained LLMs within storage and time budgets. Across six agents and six model sizes, 29 of 36 submissions passed verification, achieving a 3.4% perplexity increase over BF16 and 77.33% downstream accuracy. The study highlights representation design, experimental judgment, and capability validation as key factors in agent-driven compression.

arxiv14h agoScore 77
LLM compressionautonomous agentsrecursive self-improvement