Live intelligence

Latest AI updates

571 updates

Research Editor selected

MSafeTemp: Template-Based Multilingual Safety Alignment Across 60 Languages

Researchers propose a scalable template-based augmentation pipeline that extracts safe templates from harmful prompts, translates them, and fills slots with culturally authentic harmful and benign entities. The work releases MSafeTemp, a validated corpus spanning 60 languages for multilingual safety alignment and evaluation. Experiments show English-only evaluation can misjudge multilingual safety: attacks that fail in English may succeed when translated, and culturally localized variants expose additional failures. Template-based augmentation also improves safety performance beyond English.

Why it matters: It provides a concrete multilingual safety dataset and pipeline that exposes gaps in English-only safety evaluation and training.

arxiv6d agoScore 80

LLM safetymultilingual NLPsafety alignment

Research Editor selected

BURGER benchmark tests VLA generalization through robotic drawing

Researchers introduced BURGER, a language-conditioned robotic drawing benchmark with a prompt-driven pipeline for generating continuous drawing trajectories in simulation and the real world. Using 28,200 simulated demonstrations across 282 drawing tasks, they evaluated five representative vision-language-action models on data scaling, language grounding, and multimodal behavior. The benchmark revealed distinct failure patterns and indicated that some VLA models benefit from greater task diversity and demonstration scale and can generalize to unseen compositions. A trained policy was also deployed on a Franka Panda arm to draw reactively under human intervention.

Why it matters: It provides a structured benchmark and empirical analysis for measuring how well vision-language-action models generalize across language-conditioned robotic tasks.

arxiv6d agoScore 80

vision-language-action modelsrobotic generalizationbenchmarks

Research Editor selected

Rewindable Interactive Proofs Enable Efficient Scalable Oversight

The paper introduces rewindable interactive proofs (RIPs), a new relativizing interactive proof model for scalable oversight in which a weak verifier can rewind a powerful AI prover to an earlier interaction state and continue from there. The authors construct a RIP for computations whose outputs are robust to changing a small fraction of oracle answers, achieving an honest prover that runs in linear time and a verifier that runs in near-linear time after one-time preprocessing, without cryptography. They implement the protocol and report a preliminary MNIST experiment showing an RIP can be faster than a state-of-the-art deployed proof system. The work matters because rewinding is feasible whenever the verifier controls the prover's machine, such as trusted hardware or checkpointable virtual machines.

Why it matters: It offers a more efficient, cryptography-free scalable-oversight protocol with a working implementation, giving AI-safety researchers a practical verification tool for checking stronger models.

arxiv6d agoScore 82

AI safetyscalable oversightinteractive proof systems

Research

Study Derives Fundamental Limits of Prefill-Based Decode-Length Prediction

A new arXiv paper argues that existing benchmarks for predicting how many tokens an LLM will decode from prefill-phase hidden states conflate predictor quality with task difficulty, causing reported accuracy to collapse when the model, dataset or sampling configuration changes. The authors trace this instability to the inherent stochasticity of autoregressive decoding and model decoding as a stopped stochastic process whose per-step stopping probability ramps up over hundreds of steps, yielding per-prompt normality and an affine variance law sigma = alpha*mu + b. Empirically the normal law reproduces per-prompt MAE and the central mass of the decode-length distribution to within a few percent, with residual skew and kurtosis matching the predicted sampling-noise correction. Across MLP, EGTP, FiLM, soft-label and Mixture-of-Experts predictors, performance converges near the derived bound under fixed conditions, which the authors read as system error dominating model error, and they propose a benchmarking protocol that fixes model, data and sampling parameters and reports relative alongside absolute error.

arxiv6d agoScore 74

LLM inference schedulingprefill-based predictiondecoding stochasticity

Research

Retrieval as Traversal Proposes Value-Indexed Memory for Long-Context Agents

The paper proposes treating retrieval as a persistent local traversal through an address space, separating memory content from its address. It introduces a cost-regularised indexing objective that preserves information value in local neighbourhoods while charging for locally accessible memories, with marginal-value pressures favoring complementary memories and excluding redundant ones. In a proof-of-principle test on MiniGrid Memory, a persistent local traversal head reached 0.95 mean return after 0.59M training steps with a five-token context, compared with 0.94M for Gaussian absolute addressing and more than 2M for categorical absolute addressing. The authors note the experiment does not test learned geometry or repeated retrieval and motivates further work on stateful retrieval.

arxiv6d agoScore 68

long-context modelsretrievalmemory agents

Research Editor selected

Ramanujan Challenge Analysis Introduces AI-Assisted Proof Evaluation Pipeline

An analysis of Ramanujan Challenge submissions compares the relative success of AI models, computer algebra systems, and formal tools on research-level number theory problems. It reports correct submissions conceptually distinct from official solutions and a proof of one published open conjecture. The authors introduce an AI-assisted pipeline for evaluating natural-language proofs, CAS code, and Lean formalizations. Compared with one-shot LLM judgment, it reduces false-positive verdicts on natural-language proofs from 23% to 6% and improves judgment accuracy on supporting code from 46% to 58%.

Why it matters: It offers a concrete methodology and measured results for reliably evaluating research-level mathematical proofs across multiple formats.

arxiv6d agoScore 84

Ramanujan ChallengeAI for mathematicsproof evaluation

Research Editor selected

ClinGeX: A Skill-Centred LLM Agent for Clinical Genomic Analysis

Researchers developed ClinGeX, a large-language-model agent that composes 105 curated executable skills through task-specific retrieval and dependency-aware orchestration for clinical genomic analysis. On 223 temporally held-out benchmark cases, ClinGeX reached 71.7% macro-averaged task-level concordance with expert reference classifications, compared with 45.7% for the strongest comparator. In a human-AI collaboration study with 45 clinical genomic professionals assessing 135 cases, ClinGeX assistance increased raw case-level concordance across variant analysis, rare disease interpretation and molecular oncology decision-making, and reduced reported interpretation time. An exploratory one-month real-world assessment with 11 professionals reported frequent conclusion adoption and time savings across 96 self-selected analysis episodes.

Why it matters: It presents a concrete LLM-agent architecture with benchmark and human-study evidence for augmenting specialist clinical genomic interpretation.

arxiv6d agoScore 87

ClinGeXclinical genomic analysisLLM agents

Research

FARMS Builds a Verifiable, AI-Ready Foundation for Computational Mathematics

FARMS is a pipeline that converts computational mathematics into AI-ready knowledge by linking formal statements to source context, algorithmic meaning, and quantitative guarantees. Across 171 targets in randomized numerical linear algebra and numerical partial differential equations, 157 were accepted, with mean estimated costs of $0.85–$2.29 per target. Lean-checked implications from explicit source premises connect selected statements to computational skills, illustrated by a heat-flow solve and wave-response compression. The work aims to support affordable, reusable knowledge foundations for trustworthy automated computational science.

arxiv6d agoScore 70

automated computational scienceformal mathematicsLean

Research

Gym-Anything for Robotics uses coding agents to scale robotics simulation creation

The paper proposes Gym-Anything for Robotics, a framework in which coding agents build realistic robotics simulations, on the premise that simulation creation is itself a coding task that improves as coding agents improve. The authors select simulation domains based on U.S. GDP data and define criteria for good simulations and tasks, then apply an adversarial creation-audit loop. The resulting RoboWorld suite contains 25 simulations spanning 25 test and 175 training RL environments of varying difficulty. Frontier models reportedly average only 0.06 reward on RoboWorld test tasks, and the paper introduces in-context learning over repeated attempts, a privileged teacher-student setup, self-assisted learning, and code policies that generate demonstrations for training VLAs such as pi_0.5.

arxiv6d agoScore 78

robotics simulationcoding agentsembodied AI

Research

VeriCUA: Verification as Agentic Exploration for Computer-Using Agents

VeriCUA is a platform-agnostic framework that treats task-completion verification for computer-using agents as dynamic evidence gathering and assessment. It first inspects available context to build a task rubric, then selects GUI or code checks for unresolved requirements and combines their findings under explicit evidence rules. The authors report improved agreement with reference labels over compared baselines, with the largest offline gains on hard cases and stronger rejection of false completion claims online.

arxiv6d agoScore 74

computer-using agentsverificationagentic exploration

Research Editor selected

Collaborating LLM Agents Run Autonomous Quantum Spin Experiments and Reject Their Own Results

Researchers present an autonomous research laboratory in which LLM-based agents review each other's work and can halt or redirect investigations, including generating and rejecting research questions and demanding independent theoretical checks. The system was tested on ensembles of quantum spins in diamond, where agents derived the minimum number of calibration settings needed to predict responses to a defined family of microwave pulse sequences and verified the result computationally. When repeated reference measurements revealed signal drift, the agents classified the initial result as inconclusive and revised the measurement protocol without loosening acceptance criteria; the revised experiment met the criteria and supported the theoretical predictions. The work argues that peer-critiquing agent collectives can select questions, derive theory and test predictions experimentally.

Why it matters: It offers a concrete demonstration that multi-agent adversarial self-review can catch flawed experimental evidence in a physical-science setting, a key question for anyone building autonomous research systems.

arxiv6d agoScore 82

autonomous research agentsNV-center quantum spinsAI for science

Research

Single 1B Decoder Jointly Gates and Generates Memory-Grounded Hints

This arXiv paper evaluates proactive conversational assistance as a joint problem of deciding when to intervene and what memory-grounded hint to offer, measured over 28,058 decision points. Behind a real gate, end-to-end exact match reaches 0.2619 against an oracle-gate ceiling of 0.2971, with end-to-end precision of 0.2387. The memory profile is the dominant driver of hint quality: removing it costs 24.7 exact-match points and drops rescaled BERTScore from 0.57 to 0.35, while retrieval and a larger generator add nothing on exact match. A single 1B decoder gates from an intermediate layer and generates in the same pass, matching or exceeding two-model pipelines on automatic end-to-end measures with 1.31-1.76x less wall-clock time than the matched 1B pipeline, and a human evaluation of 1,144 ratings over 32 dialogues rates it level with the matched 1B pipeline and below an 8B one.

arxiv6d agoScore 76

proactive assistantsdialogue systemsmodel efficiency

Research

TextSeal2: Distortion-Free LLM Watermarking That Survives Speculative Decoding

TextSeal2 is a distortion-free watermark for large language models that retains its detection signal under speculative decoding: tokens replacing rejected drafts stay watermarked with the same keys, and the acceptance rate is unchanged. Instead of TextSeal's coin-flip routing between two keys, it merges multiple keys into a single virtual key coupled to fresh noise via a Gaussian copula, so detection is not diluted even with publicly shared keys. Under speculative decoding it verifies keyed drafts with an additional exponential race over the leftovers of the draft races, which by memorylessness reproduces block verification and keeps output distribution exact. Across three LLMs it detects 8-22 percentage points more 100-token answers than TextSeal at a 10^-3 false-positive rate, and 18 and 22 points more 200-token answers than TextSeal and textGrain respectively under speculative decoding.

arxiv6d agoScore 78

LLM watermarkingspeculative decodingAI content provenance

Research

Probing Spatial State Representations in Visual Foundation Models

This arXiv paper investigates whether visual foundation models maintain a persistent spatial state beyond the current view. It evaluates ten representations using frozen-feature probes across four capability levels: spatial structure, object persistence, predictive spatial update, and actionability, under streaming egocentric observations of static indoor environments. Results show that strong readout of the visible scene does not guarantee accurate object persistence or future-feature retrieval, and that a spatial reader trained on real features does not transfer unchanged to predicted features. The authors also find that goal-dependent candidate selection does not establish reliable continuous action utility, while object persistence and actionability build on camera-pose readout and feature prediction builds on visible-scene readout.

arxiv6d agoScore 74

visual foundation modelsspatial reasoningrepresentation probing

Research Editor selected

WxCrop agent turns extreme-weather forecasts into crop-loss estimates

Researchers present WxCrop, an agent that examines physically interpretable weather indices through causal analysis to assess agricultural hazards and predict crop loss for past and forecast dates. It was demonstrated in Florida for hurricanes, freezes and dry heat, with loss models calibrated on insurance records for each hazard. Reported checks include a hurricane index AUC of 0.98, held-out losses reproduced within 2%, and county rankings confirmed by independent production records with a rank correlation of 0.63. Users can request each crop group's loss as a fraction of expected season value, with every number sourced.

Why it matters: It demonstrates agentic AI converting weather forecasts into sourced, crop-specific loss estimates that could support faster agricultural disaster response and risk assessment.

arxiv6d agoScore 82

WxCropagricultural riskextreme weather

Research

SpecSlot Decouples Pixel-Space Detail from Global Attention Cost

SpecSlot is a pixel-space rectified-flow Transformer that expands each spatial anchor into multiple slots, each assigned a disjoint band of DCT coefficients, so anchor density is separated from spectral subdivision. Only the lowest-frequency slot per anchor participates in global attention while all slots interact locally, letting the model add spectral detail without lengthening the global sequence; a spatial head adds a local correction and the generative state, target and loss stay in pixel space. On ImageNet-256 it reaches a comparable final FID to a FLOPs-matched JiT control while converging faster. On ImageNet-512 it improves FID by 2.84 over the control with 3.9% fewer FLOPs and comes within 0.69 FID of JiT-B/16 at 38.2% lower cost, and on text-to-image DPG-Bench it surpasses JiT-B/8 by 1.06 at 40.6% lower forward cost.

arxiv6d agoScore 71

pixel-space generative modelsdiffusion transformersefficient attention

Research

Neural Thickets Gains Reproduced Without Weight Perturbations via Perturb-Select-Vote

This paper reproduces the reported gains of the Neural Thickets hypothesis, which claims pretrained language models sit in dense neighborhoods of task-specific experts in weight space, but shows the improvements do not require that geometric explanation. Perturbing a single decoder block recovers much of the gain, and adding a random bias to output logits of a fully frozen model matches or exceeds full-weight perturbations on most reasoning tasks. The authors formalize Perturb-Select-Vote (PSV): generate perturbed predictors, select high performers on a small validation set, and aggregate predictions, linking thickets to sampling, self-consistency, and verifier-guided selection. A first-order analysis attributes the effect to induced logit shifts, and the approach extends to vision under domain shift, including Earth observation.

arxiv6d agoScore 72

Neural Thicketsinference-time methodsmodel ensembles

Research Editor selected

Anthropic study: Experts adopt AI agent's random opinions and think they're their own

Anthropic conducted a study with 50 experts who used an AI agent with preferences determined by a coin toss. The experts followed the agent's view up to 86 percent of the time on judgment calls, and most believed the conclusions were their own. This highlights a significant risk of AI influence on human decision-making.

Why it matters: This study provides concrete, recent evidence of AI's persuasive power over expert judgment, with direct implications for AI safety and human oversight.

The Decoder6d agoScore 85

AI influencehuman oversightAnthropic research