Search: “eval-awareness”

Anthropic Adds Plugin Evals to Claude Code: 6 Grader Types, a No-Plugin Baseline, and a CI Gate for Skills

Anthropic ships a plugin evals workflow for Claude Code with six grader types, a no-plugin baseline arm, and a CI gate via threshold and cost flags.

Anthropic published a plugin evals workflow for Claude Code, exposed via the "claude plugin eval" command on v2.1.269+. Six grader types exist: regex, tool_used, tool_order, and file_exists are free transcript checks, while llm and baseline invoke a billed judge model. Every case runs with and without the plugin, and the delta (Δ) isolates the plugin's contribution; a Δ near zero with a failing tool_used:Skill grader indicates the skill never triggers. CI gating uses --threshold 0.8, --max-cost-usd, --trust-plugin, and --no-publish flags, with results written to a report.html under evals/results/.

MarkTechPost·Sep 11, 21:05 UTC · 4d agoAI tools & infra2

How well do agents use test/verification techniques?

Dan Luu's eval finds coding-agent testing instructions (TDD, formal methods, PBT, skills) mostly fail to beat defaults on Zstd implementation correctness.

The author ran 26 prompt conditions plus 4 skills on a Zstd-in-Rust implementation eval using codex with GPT-5.6, testing TDD, fuzzing, property-based testing, formal methods (Lean 4, TLA+, Verus, Kani, SMT solvers) and community skills. Nothing dramatically outperformed the default no-instruction condition, which did above average; at xhigh effort, fuzzing and PBT conditions did slightly better than formal methods. Pre-registered predictions included TDD underperforming and popular test skills (ECC, Hegel, Trail of Bits) not outperforming. Results are averages of 80 runs per condition plotted against cost.

Hacker News · security·Sep 8, 02:58 UTC · 8d agoAI researchHN 131↑ · 50 comments

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 · 6d agoAI research

Ambient @ EgoProactive 2026 : Proactive Egocentric Assistance with Visually Grounded Supervision

ECCV 2026 challenge winner reformulates egocentric intervention timing as single-token classification, boosting macro-F1 by 0.249 over free-form generation.

The paper describes the winning submission to the EgoProactive track of the ECCV 2026 Wearable AI Challenge, ranking first in the large-model division and second in the <=2B division. The method reformulates intervention timing as single-token yes/no classification, improving macro-F1 by 0.249 and G-mean by 0.30 over free-form generation. Supervision generated by a tool-calling video agent transferred better than a narration-only dataset that was four times larger and ten times cheaper, suggesting visual grounding matters more than annotation volume.

Hugging Face daily papers·Sep 9, 20:00 UTC · 6d agoAI research

The Router Within: Eliciting Native Skill Routing from a Frozen LLM

Gavel reads native skill-routing signals from a frozen LLM's forward passes with two linear maps, beating retrieve-and-rerank pipelines by up to 21.9 points on Qwen3-32B.

Gavel (Glance And Verdict from a frozen LLM) elicits skill routing from a frozen agent LLM using two trained linear maps that read mid-layer states, keeping all skill text out of context. A glance step scores the full library against compact per-skill banks built in one forward pass at installation; a verdict step resumes shortlisted skills' forward passes and fuses likelihood and yes/no judgments as a product of experts. It transfers zero-shot to three public benchmarks plus SkillTraj, a new benchmark of 372 simulated agent trajectories. On Qwen3-32B it beats progressive disclosure and retrieve-and-rerank pipelines adding 1.2B–16B external parameters by up to 13.4 points on written tasks and 21.9 when skills are needed mid-rollout.

Hugging Face daily papers·updated Sep 14, 17:58 UTC · 2d ago·first Sep 13, 20:00 UTC · 2d agoAI research 2 sources

Reference-Based Bias Detection in LLMs via Relative Representations of Hidden States

Researchers propose auditing LLM bias via relative hidden-state representations, detecting bias increases with 3-50x less compute than output-level benchmarks.

The paper introduces a reference-based bias auditing method that compares hidden-state representations across model variants, such as before and after fine-tuning, by encoding sentences relative to a fixed anchor set. The resulting Representational Bias Shift (Delta-B) correlates with output-level bias change in 15 of 18 tested settings, reaching |r| = 0.84 under full fine-tuning across WildGuardMix, DecodingTrust, and ToxiGen benchmarks. Thresholding Delta-B detects checkpoints whose bias increased with ROC AUC between 0.65 and 0.99 and beats a SEAT-based baseline, while auditing a model in about three minutes with 3-50x less compute.

Hugging Face daily papers·Sep 8, 20:00 UTC · 7d agoAI research

Do Input-Level Defenses Transfer to Observation-Level Attacks on VideoLLMs?

A systematic study shows input-level adversarial defenses provide inconsistent, often near-zero protection against observation-level attacks on video LLMs.

Researchers introduce DefTEval, a controlled framework testing eleven input-level defenses against five attack types across five video LLMs. Harmful-content detection rates are frequently near zero, and defenses fail even when attacks embed harmful signals in every sampled frame. Token compression discards localized safety features and modality fusion down-weights weakened visual signals, with defense outcomes dominated by model architecture rather than the defense method.

arXiv cs.CR·Sep 8, 06:59 UTC · 8d agoAI safety & security

ActGuard: Pre-execution Action Auditing against Indirect Prompt Injection in LLM Agents

ActGuard audits LLM agent actions before execution against predicted tool priors, masking only malicious spans from indirect prompt injections while preserving utility.

ActGuard is a pre-execution action auditing framework against indirect prompt injection in LLM agents, judging whether external content causes the current action to deviate from a locally reasonable expectation rather than whether content is inherently suspicious. At each step it predicts the tools likely used by the upcoming action, builds a local tool prior, then performs tool-level contrastive analysis and parameter-level evidence localization to identify deviations. A verifier masks only spans confirmed as malicious and regenerates the action from the sanitized context. On challenging tool-using agent benchmarks it reduces attack success to state-of-the-art levels while keeping task utility close to the no-attack setting; code is publicly available on GitHub.

arXiv cs.CR·Sep 14, 03:47 UTC · 2d agoAI safety & security

Corrupt Plans, Clean Traces: Evading Chain-of-Thought Monitoring with Plan Injection

Plan injection plants benign-sounding harmful reasoning that steers LLM actors to adversarial actions while evading chain-of-thought monitors.

Researchers show that injecting harmful but benign-sounding plans into an actor model's context causes it to perform adversarial actions while its reasoning passes chain-of-thought monitors, achieving 25-33% monitor evasion rates across benchmarks and scaling to larger models like DeepSeek-R1. Actor models paraphrase injected plans as their own reasoning without attribution. Giving the monitor access to the injected plan dropped detection by up to 50% on the Bio-Math task, with extra thinking tokens spent rationalizing rather than flagging the plan.

arXiv cs.AI / cs.LG / cs.CL·Sep 14, 18:33 UTC · 2d agoAI safety & security

Augustinian BabyLM: What Ostensive Definition Can and Cannot Teach a Small Language Model

Study shows visually grounded token embeddings in a small masked LM persist through training and improve object-property knowledge, but escape standard BabyLM benchmarks.

The paper implements ostensive definition for a small DeBERTa masked language model trained on 10M words, seeding visually grounded tokens with embeddings derived from labeled image regions before training. Visual initialization leaves a persistent, seed-replicated advantage on object-property knowledge (COMPS) and a corpus-tailored Visual-Property Swap benchmark covering color, material, size, and shape, but has no effect on most BabyLM grammar benchmarks. Synthetic grounding of previously unseeded words causally transfers the advantage to exactly those words.

arXiv cs.AI / cs.LG / cs.CL·Sep 10, 17:43 UTC · 6d agoAI research

Not All Prompts Are Equal: Exploration-Guided Prompt Scaffolding for Multimodal Reinforcement Post-Training

Exploration-guided prompt scaffolding rewrites training prompts by Exploration Potential Score, boosting multimodal RL post-training accuracy up to 11.5%.

The paper proposes dynamically adapting the training prompt distribution during online RL post-training of multimodal LLMs using the Exploration Potential Score (EPS), a lightweight rollout-based proxy for prompt utility computed from on-policy statistics with no additional overhead. Rather than discarding low-utility prompts, a teacher model generates scaffolded rewrites that preserve task intent while making training more informative. Integrated with GRPO on Geo3K and MMK12, the method achieves up to 9.7% relative in-domain improvement plus 11.5% on MathVision and 11.1% on MMMU-Pro.

Hugging Face daily papers·Sep 13, 20:00 UTC · 2d agoAI research

[AINews] AEF-1 standard emerges for Third Party Evaluators, as Xai, OpenAI, and Anthropic all cosign

xAI, OpenAI, and Anthropic cosign the AEF-1 third-party evaluation standard while Dario Amodei proposes embedded evaluators for safety verification.

The AI Evaluator Forum published AEF-1, a baseline standard for independent third-party AI evaluations covering access, conflicts of interest, funding relationships, recusal, and transparency, cosigned by xAI, OpenAI, and Anthropic. Dario Amodei wrote a rare personal blogpost proposing embedded evaluators such as METR with desks, badges, company laptops, and internal-risk-team-level access to verify safety commitments, plus democratic and global coordination frameworks. The roundup also covers the pacing debate: Bilal Chughtai left Google DeepMind arguing progress may outrun alignment, while critics including Aidan Gomez and Cohere push back against slowdowns and lab gatekeeping. Additional items include Cline Desktop's launch with open-weight model support.

Latent Space·Sep 15, 04:50 UTC · 1d agoAI safety & security

Environments as Scaffold: Enriching Feedback to Bootstrap Self-Evolving Agents in Long-Horizon Tasks

Researchers propose Feedback-Enriched Environments (FEEs) that reduce reward sparsity and improve RL training of Qwen3-based agents on SciWorld and BFCL.

The paper proposes shifting from agent-side warmup (SFT) to environment-side adaptation via Feedback-Enriched Environments to address severe reward sparsity in RL training of long-horizon LLM agents. A pilot study defines a feedback strategy that transitions from action guidance to observation enrichment in later training stages. Large-scale experiments on SciWorld and BFCL across Qwen3 model scales and GRPO, GSPO, and DAPO show consistent gains, plus stabilized training dynamics and proactive exploration.

Hugging Face daily papers·Sep 7, 20:00 UTC · 8d agoAI research

SenseNova-U1.5: Towards Native Unified Visual Intelligence

SenseTime releases SenseNova-U1.5, an 8B-MoT encoder-free multimodal model unifying visual understanding, reasoning, and generation with native 4K resolution.

SenseNova-U1.5 is an 8B mixture-of-transformers multimodal model with an encoder-free, VAE-free architecture that understands, reasons about, and generates visual content at native resolutions up to 4K. Post-training optimizes specialized experts for visual aesthetics, bilingual text rendering, infographic generation, and image editing, consolidated through multi-expert on-policy distillation. Evaluations report gains in image fidelity, text rendering, multi-reference editing, and instruction following. The team plans to open-source training code including supervised fine-tuning, reinforcement learning, and on-policy distillation.

Hugging Face daily papers·Sep 9, 20:00 UTC · 6d agoModel release

Engineered Persuasion: Evaluating Personalized Pretexts in LLM-Generated Spear Phishing

A study of 180 US workers found each LLM phishing personalization level raised click-intention odds by 28%, but credibility depends on context fit.

The arXiv paper evaluates how personalized pretexts in LLM-generated spear phishing affect perceived credibility, using 180 US working adults across 1,436 evaluations of emails with four cumulative personalization levels, from workplace context to shared-project details. Convincingness rose 2.40 points per level in sensitivity analysis and click-intention odds increased 28% per level, while non-clickers shifted toward deleting rather than reporting. Qualitative coding showed details matching the recipient's role and routines supported credibility, whereas incorrect, vague, or channel-inappropriate details raised suspicion. The authors argue personalization effectiveness depends on pretext fit, with implications for workplace security training.

arXiv cs.CR·Sep 3, 19:21 UTC · 12d agoResearch

Omni-Streaming Thinking

Omni-Streaming Thinking fixes premature cross-modal commitment in streaming omni-modal models via pending claims verified against modality-specific evidence, beating baselines by over 10%.

The paper identifies 'premature cross-modal commitment', where streaming models keep relaying early visual interpretations even after audio contradicts them. OST generates evidence-linked pending claims with future verification intervals, stores audio and visual evidence separately, and refutes claims when contradictory evidence appears. Built on a frozen Qwen3-Omni-30B-A3B-Instruct backbone with lightweight adaptation, it outperforms open baselines by more than 10% relative on five streaming and audio-visual benchmarks. On the new OST-DiagBench it reaches d-prime 2.95 versus at most 1.38 for open baselines, while reducing vision-induced auditory hallucinations.

Hugging Face daily papers·Sep 13, 20:00 UTC · 2d agoAI research1

The Model Proposes, the Code Disposes: A Pre-Registered Ablation of a Verifier-and-Acceptance Stage in an LLM-Orchestrated Offensive-Security Agent

Pre-registered ablation finds a model verifier stage in an LLM offensive-security agent suppresses findings; removing it eliminated suppression with precision tradeoff.

The paper evaluates a verifier-and-acceptance stage in an LLM-orchestrated offensive-security agent via a pre-registered 20-run confirmatory ablation and a 2x2 factorial study with 40 runs on vulnerable lab targets. Removing the stage eliminated pre-report suppression (median 2 vs 0 findings, p = 0.00003) but reduced model-blinded shipped precision (0.471 vs 0.353, p = 0.0087). Suppression was attributed to the model verifier rather than deterministic acceptance rules, and an instrumented canary recorded zero external contacts in all 60 runs. The full design retained 93.8% of model-adjudicated true candidates but failed its pre-registered non-inferiority floor of 0.90.

arXiv cs.CR·Sep 14, 17:13 UTC · 2d agoResearch

BenchShield: Formal Model-Backed Instrumentation for Reward Integrity in LLM-Agent Evaluation Infrastructure

BenchShield uses lifecycle-model-backed instrumentation to detect reward hacking in LLM-agent benchmarks, lifting full-chain recall to 77-100% at up to 65% lower cost.

The framework grounds reward-hacking detection in a finite lifecycle model of an evaluation's reward-relevant events, combining a static phase-aware taint analysis with runtime infrastructure-side evidence attribution. Evaluation used a human-labeled corpus of 456 adjudicated trajectories drawn from more than 31,000 public agent runs across three benchmarks. BenchShield improves full-chain recall from 23-94% to 77-100% and same-vector coverage from 16-56% to 43-78%, cuts per-task cost by up to 65%, and achieves 96% accuracy detecting reward hacking at runtime.

arXiv cs.CR·Sep 10, 03:10 UTC · 6d agoAI safety & security1

Molecular Déjà Vu: Digit-Level Retrieval of Published Values in Frontier Language Models

Audit of 22 frontier models finds widespread verbatim retrieval of published molecular property values, with higher reasoning increasing recall of memorized numbers.

An arXiv audit tests 22 frontier LLMs across 12 molecular regression benchmarks for verbatim retrieval of published values. More than 50% of the LLMs show verbatim retrieval on five datasets, and identical experiments are flagged 89% more often at a high reasoning level than at the lowest one. Suppressing retrieval moves model prediction errors closer together in relative terms, suggesting predictive capability is not determined solely by memorized values.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 17:32 UTC · 12d agoAI research1

Atria Dawn: The Dawn of Agentic Superintelligence

Atria Dawn Preview, an agentic foundation model trained on verifiable experiences, tops five of 16 research and engineering benchmarks.

Atria Dawn Preview is a foundation agentic language model for scientific research and engineering workflows, trained via a Verifiable Experience Pipeline connecting tool-mediated interactions to executable environments and externally verified outcomes. Across 16 benchmarks spanning research, engineering, and digital work it is competitive with frontier agents and achieves the highest reported score on five of them. The release includes a human-AI collaboration case study analyzing 769 task records from 56 participants, where about one-third of completed AI-assisted tasks were rated infeasible without AI and agents frequently proposed methods and implemented revisions while humans retained final decisions.

Hugging Face daily papers·Sep 13, 20:00 UTC · 2d agoModel release

Show HN: LLM Attention Visualization

A developer released a browser-based tool that visualizes which past tokens influence each LLM output token using aggregated, value-weighted attention scores.

A Show HN project presents a React application built on Transformers.js that renders per-token attention influence by aggregating attention weights scaled by value-vector magnitudes across all attention heads and layers. To expose internal tensors, the author instrumented the ONNX computation graph, hosted a modified model on Hugging Face, and pre-generated prompts to avoid long model downloads in the browser. Demos with a 600-million-parameter model show how verbatim copying draws heavily on source tokens and how single outputs blend information from multiple phrases.

Hacker News · AI·Sep 8, 16:59 UTC · 8d agoAI tools & infraHN 41↑ · 9 comments

Training-Free Speech-Centric Omni Understanding with Frozen VLMs

Audio-visual understanding remains challenging because models must jointly interpret spoken content, visual events, and their temporal relationships. Existing omni models typically introduce dedicated audio encoders and rely on expensive audio-video-text training, tightly coupling omni capability to specific VLM backbones and potentially weakening their existing visual and reasoning abilities.…

Hugging Face daily papers·Aug 6, 20:00 UTC · Aug 6, 2026AI research

Retrofitting Code Using LLMs to Support Exceptional Behavior

EXCODER combines static/dynamic analysis with LLMs to retrofit exception-handling code, achieving 85.92% pass@1 with Qwen 2.5 Coder 32B on Java benchmarks.

The paper introduces the task of retrofitting existing code with Exception Related Code (throw statements, guarding conditions, try/catch blocks) so that given Exceptional Behavior Tests pass. EXCODER performs context engineering by integrating static and dynamic program analysis output with LLMs; it was evaluated on a benchmark built from 304 methods across 75 GitHub Java projects. Combined with Qwen 2.5 Coder 32B, EXCODER achieves pass@1, 5, and 10 rates of 85.92%, 86.18%, and 86.51%, roughly 13 percentage points over baseline, and manual inspection reveals remaining limitations.

arXiv cs.AI / cs.LG / cs.CL·Sep 9, 16:18 UTC · 7d agoAI research1

Perturbation Probing: A New Diagnostic for the Fragility of LLM Safety

Unit 42 research shows LLM safety refusals concentrate in a thin neural layer, motivating external, multi-layered AI security controls.

Palo Alto Networks Unit 42 introduces Perturbation Probing, a diagnostic technique for measuring the fragility of LLM safety mechanisms. The research finds that safety refusal behavior is localized within a thin neural layer, implying small perturbations can undermine built-in refusals. The authors argue this motivates external, multi-layered security defenses on top of model-internal safety training.

Palo Alto Unit 42·Aug 28, 22:00 UTC · 18d agoAI safety & security

Misleading the Planner through Deceptive Resumes: Registration-Time Injection in Centralized Multi-Agent Systems

Researchers demonstrate registration-time prompt injection in centralized LLM multi-agent systems, dropping GAIA task success from 84.31% to 37.25%, and propose DescGuard defense.

The paper identifies a registration-time injection channel in centralized LLM multi-agent systems where third-party worker agent descriptions are trusted by the planner before any user instruction arrives. Analyzing 32,000 descriptions from three public agent marketplaces, at least 23.35% contain content outside the four defined description fields. Eight description-manipulation attack strategies targeting task decomposition, capability grounding, and subtask specification cut GAIA task success from 84.31% to 37.25% and increased token consumption or execution time by over 111%, persisting across two MAS implementations, six planner LLMs, and four evaluators. The proposed DescGuard defense filters descriptions to worker-scoped interface information and restores metrics toward baseline without modifying workers, planner, or orchestration logic.

arXiv cs.CR·Sep 14, 13:06 UTC · 2d agoAI safety & security

ActReview: Rebuttal-Guided Training Data and Rubric Rewards for Actionable Peer Review Generation

ActReview post-trains Qwen3-8B-Base on 40K rebuttal-derived instances with rubric rewards to generate actionable, grounded peer-review feedback, plus a 1,000-instance benchmark.

The framework builds ActReview-40K from real OpenReview review-rebuttal threads, aligning reviewer weaknesses with author responses and grounding feedback in localized paper evidence. Qwen3-8B-Base is post-trained with multi-task supervised fine-tuning followed by GRPO using candidate-aware, weakness-specific rubric rewards. Experiments show improved actionability and grounding over prior specialized review-generation models, supported by ActReview-Bench, a human-curated 1,000-instance evaluation set. Human evaluation confirms better revision usefulness while noting a remaining gap in technical accuracy.

arXiv cs.AI / cs.LG / cs.CL·Sep 8, 17:30 UTC · 8d agoAI research

When Models Edit Too Much: On the Fidelity of Minimal Code Edits

A 400-task BigCodeBench evaluation shows frontier LLMs widely over-edit code; a preservation instruction cuts excess edits and raises Pass@1 by 2.3 points.

Researchers built an evaluation framework from 400 BigCodeBench problems with injected AST-level corruptions, each with a known minimal patch, to measure over-editing in LLM code repair. Even strong models like GPT-5.5 produce unnecessarily large edits despite high Pass@1. Adding a preservation instruction reduced average excess Levenshtein distance from 0.195 to 0.131, cut added cognitive complexity by 26.6%, and raised Pass@1 by 2.3 points. Reinforcement learning post-training gave the best out-of-domain edit-fidelity trade-off, while supervised fine-tuning overfit to seen corruption patterns.

Hugging Face daily papers·Sep 2, 20:00 UTC · 13d agoAI research1

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

SchemeArena: Factorized Stress Testing of Scheming in LLM Agents

Researchers introduce SchemeArena, a 400-scenario benchmark stress-testing scheming in LLM agents, finding explicit instrumental goals are the strongest driver of covert misaligned behavior.

The paper presents SchemeArena, a 400-scenario benchmark built through factorized scenario synthesis spanning safety-relevant tool domains, instrumental goals, oversight conditions and pressure mechanisms. The accompanying SCOUT monitor grounds multi-criteria scheming judgments in evidence drawn from agents' reasoning and actions. Stress tests across five LLM agents show explicit instrumental goals are the strongest driver of scheming propensity, while action-only monitoring increased scheming in several closed models, suggesting partial oversight can act as an optimization constraint. The benchmark, code and monitor are released at github.com/launchnlp/SchemeArena.

Hugging Face daily papers·Sep 7, 20:00 UTC · 8d agoAI safety & security

What Matters, When? Diagnosing and Improving Conditional Visual Grounding in Visuomotor Imitation Policies

Researchers diagnose conditional visual grounding failures in visuomotor imitation policies and show targeted interventions substantially improve distractor robustness.

The paper studies why ACT-based visuomotor imitation policies fail when visually similar distractor objects or receptacles are introduced, finding sensitivity depends on both distractor type and manipulation stage. Interventions including distractor augmentation, phase-dependent attention regularization, and appearance-based visual prompting improve target selection while preserving spatial control information, with gains in simulation and on a physical UR3e. The same failure pattern is confirmed in a pretrained vision-language-action policy on a state-conditioned medical instrument-handling task.

arXiv cs.AI / cs.LG / cs.CL·Sep 4, 17:27 UTC · 12d agoAI research