Search: “SAEScientist-Bench”

Smart search ranks by meaning as well as keywords (one row per story, last 45 days).

SAEScientist-Bench: Can AI Agents Conduct Autonomous SAE Interpretability Research?

SAEScientist-Bench tests whether AI agents can autonomously run SAE interpretability research on Gemma-2-9B-IT; frontier agents trail expert baselines.

The benchmark requires agents to design contrastive probes and navigate the Gemma Scope dictionary of over 131K features in Gemma-2-9B-IT to discover optimal interpretable features, scored against expert-curated references on Neuronpedia via activation rank, concept selectivity, and causal steering. Across 10 agent configurations and 20 tasks, frontier agents demonstrate genuine discovery capability and approach expert levels at separating target concepts from controls, but lag substantially in causal steering and frequently misinterpret experimental measurements. The authors frame this as establishing experimental model understanding as a measurable capability for closed-loop autonomous AI R&D and post-hoc monitoring for recursive self-improvement.

arXiv cs.AI / cs.LG / cs.CL·Sep 8, 17:45 UTC · 8d agoAI research1

40

SAEScientist-Bench: Can AI Agents Conduct Autonomous SAE Interpretability Research?

SAEScientist-Bench evaluates whether AI agents can autonomously conduct SAE interpretability research in Gemma-2-9B-IT, finding frontier agents trail expert baselines.

SAEScientist-Bench tests if AI agents can act as scientists using SAE tools for autonomous mechanistic discovery, requiring them to design contrastive probes and navigate a Gemma Scope dictionary of 131K+ features in Gemma-2-9B-IT. Across 10 agent configurations and 20 tasks, frontier agents showed genuine discovery capability but remained well behind expert reference features, lagging most in causal steering. Agents frequently misinterpreted experimental measurements even when designing effective contrasts.

Hugging Face daily papers·Sep 7, 20:00 UTC · 9d agoAI research

55

StudyBench: Can Self-Evolution Squeeze Textbooks for Olympiad Capability?

Researchers introduce StudyBench, a physics benchmark showing self-evolution gains on textbook problems rarely transfer to olympiad-level questions.

StudyBench is a controlled physics benchmark splitting test data into an Application Set of difficult textbook problems and a Transfer Set of olympiad-level problems. Across three base models, representative self-evolution methods improved on the Application Set but rarely transferred to the harder Transfer Set. A guidance ablation reveals a Guidance Gap, and every method hits a Compute Plateau, indicating the remaining limits are method problems rather than data or compute problems.

Hugging Face daily papers·Aug 31, 20:00 UTC · 16d agoAI research

28

K-Bench: a clinically calibrated benchmark for evaluating large language models in high-risk mental health conversations

Clinician-calibrated K-Bench evaluates 125 LLM configurations on 200 high-risk mental health vignettes, exposing wide variation in suicide and violence risk handling.

K-Bench is a clinician-calibrated, protected benchmark evaluating 125 model configurations from 33 base models across 14 providers on 200 multi-turn vignettes covering suicide, self-harm, domestic violence, substance misuse and no-risk presentations. A frozen GPT-4o judge achieved 94.2% exact agreement with clinician consensus across 6,751 eligible comparisons from 151 clinician-rated transcripts. Leading models combined supportive conversation with combined-risk scores above 95, while risk exploration varied substantially among weaker configurations; therapeutic prompting helped weaker models and elevated reasoning produced no average improvement. A continuously updated public leaderboard is hosted at k-bench.ai with protected test materials.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 16:49 UTC · 2d agoAI safety & security

46

SAFIRE: Safety-Critical Benchmark for Fine-grained Fire and Smoke Understanding in Multimodal LLMs

SAFIRE, an 83K-image fire and smoke benchmark, shows open-source multimodal LLMs average only 61.9% accuracy on safety-critical fire reasoning.

SAFIRE is a large-scale benchmark for fire-smoke understanding in multimodal LLMs with 83K captioned images across 20 scenarios and 193K multiple-choice VQA questions spanning 10 evaluation dimensions from perception to higher-order reasoning. Annotations were built via a GPT-5.4-assisted multi-stage pipeline with MLLM majority voting. Ten open-source MLLMs (8B-38B) average 61.9% accuracy, exposing major gaps in safety-critical reasoning. Adapting vision encoders on 7% of the domain data raises fire-scene classification from 20.1% to 64.5%.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:59 UTC · 9d agoAI research

33

VDiff-Bench: A Challenging Benchmark for Fine-Grained Image Difference Identification

VDiff-Bench, a 1,756-question benchmark, shows multimodal LLMs struggle with fine-grained image-difference identification, scoring as low as 8.7% on low-level changes.

VDiff-Bench is a multiple-choice benchmark of 1,756 four-way questions over image pairs covering 10 change categories including position, motion, color, texture, OCR/text and illumination, with curated hard negatives. Evaluation of 11 state-of-the-art open- and closed-source MLLMs shows fine-grained visual comparison remains brittle: 7-8B-scale open-source models score 52.5-70.6% on semantic changes but only 8.7-33.3% on low-level changes like noise and texture. Notably, Grok 4.3 shows a sharp performance drop on noise and texture differences, falling behind large open-source models like Kimi K2.5 and K3.

Hugging Face daily papers·Sep 4, 20:00 UTC · 12d agoAI research

32

PIA-Bench: Towards Automated Privacy Impact Assessment with Large Language Models

Researchers release PIA-Bench, the first open benchmark evaluating how accurately LLMs can automate privacy impact assessments using 73 curated federal PIAs.

PIA-Bench is the first open benchmark for evaluating large language models on real-world privacy impact assessments (PIAs). The authors audited 499 expert-authored PIAs published by US federal agencies and curated 73 structured PIAs comprising 451 privacy risk items and 831 mitigation items. Off-the-shelf LLMs were found to produce meaningful assessments while identifying clear avenues for improvement. The paper calls for domain-specific LLM agent workflows, accountable LLM infrastructure, and new quality standards for PIAs.

arXiv cs.CR·Sep 11, 08:16 UTC · 5d agoResearch1

48

MetroLLM-Bench: Evaluating Language Models as Transit Kiosk Runtimes

MetroLLM-Bench is a 955-case benchmark testing language models as transit kiosk tool-calling runtimes across six real metro systems.

The benchmark covers 37-414-station metro systems and eleven task categories including routing, fare calculation, disruptions, accessibility, and adversarial input, with 14 deterministic and 8 semantic scoring components. Of 26 models from six vendors, a PEFT-tuned 4B Qwen 3.5 student scored 91.3 on Tier 1, exceeding GPT-5.6 (90.6/90.0), while Muse Glimmer 30B led the composite ranking. A deterministic rule-based baseline reached 84.6, and PEFT gains over base models shrank from +7.03 points at 2B to -0.91 at 27B.

Hugging Face daily papers·Sep 8, 20:00 UTC · 8d agoAI research

30

PIDS-Bench: Evaluating Prompt-Injection Detectors Under Over-Defense, Obfuscation, and Distribution Shift

PIDS-Bench shows prompt-injection detectors scoring F1 above 0.98 still misclassify about one-third of external benign security-adjacent prompts, revealing provenance-sensitive over-defense.

PIDS-Bench is a frozen multi-axis benchmark that jointly evaluates prompt-injection detectors on attack detection and benign false-positive behavior at fixed thresholds, spanning in-distribution inputs, hard-benign prompts, obfuscated attacks, and domain/structural distribution shifts. It evaluates seven detectors plus a rule-based lower-bound reference. A detector exceeding F1 = 0.98 on held-out data still misclassifies roughly one-third of an externally-sourced benign security-adjacent subset, and no internal detector reaches F1 >= 0.95 with hard-benign FPR <= 0.10 on the stress distribution. Hard-negative augmentation nearly eliminates over-defense on curated stress inputs but leaves it intact on externally-sourced prompts, a pattern termed provenance-sensitive over-defense.

arXiv cs.CR·Sep 14, 04:28 UTC · 2d agoAI safety & security

38

MindTopo: Can Foundation Models Reason in Topological Space?

MindTopo benchmark with 11,030 topological tasks shows 14 multimodal LLMs reason better than they plan and remain far below human performance.

MindTopo is a benchmark of topological intuition across five properties grounded in cognitive science and formal topology: continuity, separation, order, enclosure, and knots, evaluated at reasoning and closed-loop planning levels. It contains 11,030 instances across 13 procedurally generated task types with controllable difficulty, benchmarking 14 multimodal LLMs plus agent configurations using image and video generation, including three video generative models. Every MLLM performs better on reasoning than on planning, and the best-performing model remains far below observed human performance. On Qwen3-VL-2B-Instruct, supervised fine-tuning and reinforcement learning improve reasoning more than planning, and audited generated rollouts often fail to follow environment dynamics or preserve topology across transitions.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:54 UTC · 6d agoAI research

33

τ^τ-Bench: An Environment for End-To-End, Realistic Agent Construction

New τ^τ-bench tasks coding agents with building deployable customer-service agents; best config, Claude Opus 5, passes only 23.9% of simulations.

Researchers introduce τ^τ-bench, an end-to-end benchmark where a developer agent must build a complete customer-service agent from real business records, a client with requirements, a production API, an inherited codebase, and cost/model limits, then is scored by deploying it against held-out simulated users. Across 53 tasks in four domains, the strongest configuration, Claude Opus 5 under Claude Code, passes just 23.9% of evaluation simulations versus an 82.2% expert-authored reference ceiling. Failure modes mirror those of human developers: shallow queries instead of deep record comprehension, almost no client communication, and shipping the first architecture that runs rather than experimenting.

Hugging Face daily papers·Sep 3, 20:00 UTC · 13d agoAI research

50

MP-Bench: Evaluating Voice Agents as a Multiparty Conversation Participant

MP-Bench is the first benchmark for voice agents in multiparty conversations, finding real-time agents near chance on turn-taking.

MP-Bench is the first benchmark designed to objectively evaluate conversational speech systems as active participants in multi-party conversations. It assesses agents on turn-taking awareness and response appropriateness, with comprehension-based question-answering as a complementary evaluation. Benchmarking 12 voice agents shows real-time agents score at or below 22% on multiparty comprehension and remain near chance on multiparty turn-taking.

arXiv cs.AI / cs.LG / cs.CL·Sep 11, 17:11 UTC · 5d agoAI research2

20

WearableQA: A Benchmark for Health Reasoning over Real-World Wearable Data

WearableQA benchmark introduces 4,084 questions over real longitudinal wearable data, showing 14 LLMs score 19.6-72.9% on health reasoning, far from solved.

WearableQA is a benchmark of 4,084 10-option multiple-choice questions built from wearable time series, blood biomarkers, and demographics of 200 real users with up to 500 days of daily measurements. It defines 16 question types along two axes: data versus health reasoning, and single- versus cross-signal reasoning, using a dual-grounding framework combining literature and population-validated patterns. Evaluations of 14 proprietary and open-source LLMs show accuracy ranging from 19.6% to 72.9% against a 10% chance baseline, with most models below 60%.

Hugging Face daily papers·Sep 3, 20:00 UTC · 13d agoAI research

35

An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics

Open post-training pipeline turns Nemotron 3 Ultra checkpoints into an IMO 2026 gold-medal system, scoring 30/42 without formal provers or external tools.

Starting from Nemotron 3 Ultra, researchers trained two specialist checkpoints using supervised fine-tuning and reinforcement learning for natural-language olympiad proof generation. Three checkpoints power an iterative generate-verify-refine search plus a separate high-compute selection stage, operating entirely in natural language with no formal prover, external tools, or internet access. The system scored 30 of 42 points at IMO 2026, reaching the gold-medal threshold. The release includes the post-trained checkpoints, training data, training and inference code, submitted solutions, and Nemotron-IMO-Bench with 200 novel olympiad-level problems.

Hugging Face daily papers·Sep 8, 20:00 UTC · 8d agoModel release

65

BenchMIRT: What are LLM benchmarks actually measuring?

AllenAI's BenchMIRT blog post examines what LLM benchmarks actually measure and their reliability.

AllenAI published a Hugging Face blog post introducing BenchMIRT, which investigates what large language model benchmarks actually measure. No article text is available, so specific findings, methods, or benchmark scores cannot be extracted. The work appears to target benchmark validity, a live concern for model evaluation and comparison.

Hugging Face Blog·Sep 1, 21:39 UTC · 14d agoAI research

38

StochBench: A Domain-Specific Benchmark for Stochastic Processes in Lean

StochBench introduces 450 graduate-level stochastic-processes problems in Lean 4; an Opus 4.8-based agent proved 34.9% under a 15-minute limit.

StochBench is a Lean 4 benchmark of 450 graduate stochastic-processes problems, each paired with its natural-language source, covering Markov chains, renewal processes, martingales, Brownian motion, stochastic calculus, weak convergence, and Poisson and continuous-time Markov processes. The benchmark addresses field-specific applied mathematics underrepresented in Mathlib, unlike competition-math-dominated suites such as IMO and Putnam collections. An Opus 4.8-based agent achieved a 34.9% proof rate (157/450) under a 15-minute per-problem limit, showing the benchmark remains challenging for advanced provers.

Hugging Face daily papers·Sep 7, 20:00 UTC · 9d agoAI research

25

Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation

Benchmark Radar provides a living searchable database of 1,283 AI benchmark records and 12,916 score observations drawn from 37 daily discovery sources.

Benchmark Radar combines daily discovery of benchmark papers, repositories, datasets, and releases from 13 direct connectors and 24 first-party feeds into a searchable catalog with model card mentions and score histories. The catalog contains 1,283 source records drawn from 4 benchmark catalogs plus 12,916 numeric observations on 790 records. The release includes a web dashboard with leaderboard, Pareto frontier of score versus usage, saturation and trend views, daily feeds, a CLI, and reproducible analysis. The paper audits the full catalog and examines benchmark saturation and limits of score comparisons.

Hugging Face daily papers·Sep 9, 20:00 UTC · 7d agoAI research

25

Embodied-BenchForge: A Closed-Loop Agentic Workflow for Embodied Benchmark Construction

Embodied-BenchForge automates embodied benchmark construction via closed-loop synthesis with verification and repair, yielding seven benchmarks for MLLM evaluation.

Embodied-BenchForge is an agentic framework that transforms user-specified evaluation intents into complete embodied benchmark artifacts via Closed-Loop Benchmark Synthesis. Skill-Orchestrated Artifact Synthesis composes typed reusable skills while an artifact dependency graph records intermediate outputs; Requirement-Guided Verification and Repair triggers local re-execution or upstream rollback on failures. It constructs six Offline EQA benchmarks plus one interactive benchmark with 220 executable tasks, distinguishing MLLM and embodied agent capabilities in observation-based understanding and closed-loop execution.

arXiv cs.AI / cs.LG / cs.CL·Sep 11, 17:16 UTC · 5d agoAI research

18

ImpossibleRubrics: Stress-Testing Generated Rubrics as Reward Signals

ImpossibleRubrics benchmark shows LLM-generated rubric reward signals are exploited 8-26% of the time by adversarial answers on impossible tasks.

ImpossibleRubrics is a benchmark of 169 impossible tasks across six impossibility categories, each paired with a verifiable oracle certificate, plus 48 answerable controls, for stress-testing LLM-generated rubrics used as reward signals. Eleven rubric generators were exploited 8-26% of the time on an unbiased 150-task cut and up to 36% on a stress cut, while a certificate-faithful rubric scored 0%. A single generic 'be decisive, penalize hedging' rubric was exploited 64% of the time, suggesting tailored criteria can reveal which claims attackers should fabricate.

Hugging Face daily papers·Sep 14, 20:00 UTC · 2d agoAI research

45

MAxBench: A Multinomial Concept Recovery Benchmark

MAxBench evaluates multinomial concept recovery methods, finding affine subspaces steer most reliably but none consistently beats prompting.

MAxBench is a geometry-agnostic evaluation framework for multinomial concept representations in language models, based on sampling from recovered concept representations. It compares 10 localization methods covering 5 geometry types across 6 concepts and 4 models. Findings show affine subspaces steer more reliably than rank-one or linear subspaces due to better non-zero offsets, manifold steering is competitive where applicable, and no method consistently outperforms prompting.

arXiv cs.AI / cs.LG / cs.CL·Sep 11, 17:08 UTC · 5d agoAI research

18

E2A-Bench: Benchmarking Evidence-to-Action Reliability in Financial Chart Reasoning

E2A-Bench, a 969-query financial chart reasoning benchmark, finds VLMs fail evidence-to-action consistency, with fine-tuning amplifying BUY:SELL bias 4-6x.

E2A-Bench is a 969-query benchmark built from 323 HS300 constituents across three input modalities with deterministic OHLCV-derived evidence anchors, evaluating grounding, reasoning-action consistency, evidence-confidence calibration, and directional coverage via UCR, RCI, ECI, and NDR metrics. Testing 20 VLMs showed the lowest-hallucination model ranked near the bottom on coverage with only 6.4% directional coverage, and oracle-aided verification reduced unsupported claims but could collapse coverage. Financial fine-tuning amplified the BUY:SELL ratio by factors of 4.21 to 4.68 across base-fine-tuned pairs.

Hugging Face daily papers·Sep 12, 20:00 UTC · 4d agoAI research

34

A misalignment of AI in mathematics

Fields Medalists including Terence Tao warn AI benchmark-chasing in mathematics is misaligned with scientific goals and threatens research culture.

Twenty-five Fields Medal-winning mathematicians, including Terence Tao, Peter Scholze, and Manjul Bhargava, published an essay arguing that AI companies' push to solve mathematical problems as benchmarks is severely misaligned with the goal of conceptual understanding. They warn that rushed AI solutions create attribution and plagiarism risks and bypass the human process of integrating results into the mathematical canon. The authors frame the issue as a broader alignment problem facing scientific and creative professions and call for urgent action by AI developers and society.

Hacker News · AI·Sep 11, 17:45 UTC · 5d agoAI safety & security 2 sourcesHN 59↑ · 153 comments

35

IdeaAMBIG: Benchmarking Implementation-Critical Gaps in Research-Idea Specifications

IdeaAMBIG benchmark with 660 instances measures whether LLMs can spot and fix underspecified research-method details for faithful implementation.

Researchers introduce IdeaAMBIG, a benchmark of 660 evidence-grounded instances (163 real-world gaps from reproducibility reports and GitHub issues, 497 controlled synthetic gaps) built from papers, codebases, and reproduction artifacts. It evaluates codification-readiness assessment, defect localization, and clarification action generation. Across 13 LLMs, the best model achieved only 9.6% Macro Defect Recovery Rate on real-world instances but 80.6% clarification success when given the annotated defect. An oracle study showed gold resolutions raise the codification-ready rate from 14% to 98%, identifying defect localization as the main bottleneck.

arXiv cs.AI / cs.LG / cs.CL·Sep 9, 17:59 UTC · 7d agoAI research2

28

RESCUE-BENCH: Towards Relation-Aware Multi-Party Emotional Support Conversation Systems

Researchers introduce RESCUE-Bench, a video benchmark of 191 couple and family conversations evaluating LLMs on relation-aware multi-party emotional support.

RESCUE-Bench is built from real couple and family interview conversations, containing 191 samples, 7,079 annotated turns, and 1,064.8 minutes of video. It defines six tasks measuring two capabilities: Relational Understanding and Relation-Sensitive Support. Experiments with ten LLMs show models handle local emotional cues but struggle with relation pattern prediction, viewpoint prediction, and support strategy prediction.

Hugging Face daily papers·Sep 8, 20:00 UTC · 8d agoAI research

27

Import AI 469: Science AI; RSI simulator; and Zuck's technological pessimism

New DiG-bench benchmark of 70 hidden-rule games shows only Opus 5 and Fable 5 solving the hardest tiers, probing AI discovery and creativity.

Import AI 469 highlights DiG-bench (Discovery in Games), a benchmark of 70 handcrafted games with hidden rules and objectives where only 21 games are public and most are kept private to avoid training contamination. Only Opus 5 and Fable 5 with Claude Code solved any Tier 7 tasks (about 0.2 success), with GPT-5.5 next; the games are text-based and have beaten every human tester at least once. The newsletter also covers an RSI simulator game by Paradigm Research and Inherent's Faraday, a post-trained open-weight model that supervises frontier models to improve scientific research output.

Import AI·Aug 17, 13:05 UTC · Aug 17, 2026AI research

40

Ambient @ EgoLongQA 2026: Distilling Long-Video perception into a Sub-2B Model

Ambient team wins EgoLongQA 2026 sub-2B division by distilling an agentic long-video perception pipeline into a 2B vision-language model.

Ambient's entry to the EgoLongQA track of the Wearable-AI Challenge at ECCV 2026 placed first in the <=2B parameter division with 0.8279 on the held-out test set. The system distills the junior perception module of a tool-using agentic pipeline into a 2B student, reaching 89% of the pipeline's accuracy with 1.1% of its parameters and lifting a 27.1% base model to 81.4%. To meet the division limit, the multilingual embedding table is pruned from 248,320 to 143,469 rows, reaching 1.9985B parameters with provably identical logits on retained rows.

Hugging Face daily papers·Sep 9, 20:00 UTC · 7d agoAI research

35

xDailyBench: Benchmarking LLMs on Professional Consultation for Real-Life Problems

xDailyBench tests 11 frontier LLMs on 248 real-life consultation tasks; the best models score 75.6% and lag on implicit requirements.

The benchmark spans 51 scenarios across personal life, white-collar work, learning and research, and cross-domain activities, grounded in requests users actually completed or intended to complete with AI. Tasks are scored with fine-grained binary rubrics covering explicit and implicit requirements under standardized agentic settings. Across 11 frontier models, the best achieved a 75.6% task-level score, with all models performing at least 9 percentage points worse on implicit than explicit requirements.

arXiv cs.AI / cs.LG / cs.CL·Sep 7, 17:30 UTC · 9d agoAI research

26

DianShi-RxnDB: A Large-Scale, Fine-Grained Organic Reaction Data Platform Built via a Fully Automated Pipeline for Researchers and AI Agents

Researchers release DianShi-RxnDB, a database of roughly 24 million organic reaction instances extracted automatically from USPTO and EPO patents since 1976.

DianShi-RxnDB is built by a fully automated pipeline integrating patent text, images, and reaction schemes, yielding about 24 million reaction instances, of which 14.8 million (61.7%) pass automated qualification checks. Manual evaluation of 1,300 sampled instances showed 92.95% field-level accuracy, and comparisons with Pistachio found advantages in deduplicated record counts and granularity. The platform offers a web research workbench and a Model Context Protocol (MCP) service enabling AI agents to perform composable structured retrieval.

Hugging Face daily papers·Sep 5, 20:00 UTC · 11d agoAI research

28

GPT-6 Astra appears to show a "step change" in spatial reasoning based on early benchmarks

Early StationeryBench robotics results show OpenAI's GPT-6 Astra far ahead of Ai2's MolmoAct2 at dual-arm manipulation, completing 7 of 100 tasks versus zero.

A new robotics benchmark called StationeryBench tested OpenAI's GPT-6 Astra against Ai2's MolmoAct2 on five desk-object tasks using identical dual-arm YAM robots over 200 trials. Astra fully completed 7 of 100 tasks with a median progress score of 46 out of 100, while MolmoAct2 completed zero with a median score of 12. Cornell and Google DeepMind researcher Yoav Artzi called the result a 'step change in spatial reasoning' and noted Astra approaches human-level accuracy on the unpublished REMAP benchmark. He speculated OpenAI trained the model on large amounts of 3D data such as Blender scenes, and OpenAI reportedly plans consumer robots.

The Decoder·Sep 12, 14:26 UTC · 4d agoAI industry1

48

WearableQA: A Benchmark for Health Reasoning over Real-World Wearable Data

WearableQA benchmark tests LLM health reasoning over longitudinal wearable data; the best of 14 evaluated LLMs reaches 72.9% accuracy.

WearableQA comprises 4,084 ten-option multiple-choice questions built from wearable time series, blood biomarkers, and demographics of 200 real users with up to 500 days of daily measurements. It defines 16 question types along two axes: data versus health reasoning, and single- versus cross-signal reasoning. Evaluation of 14 proprietary and open-source LLMs shows performance from 19.6% to 72.9% against a 10% chance baseline, with most models below 60%.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 17:52 UTC · 12d agoAI research

14