AdversaBench: Automated LLM red-teaming pipeline with multi-judge confirmation and cross-model transferability
AdversaBench is a new end-to-end red-teaming pipeline that mutates seed prompts using five structured operators and confirms failures via a three-judge panel with a meta-judge tiebreaker. Experiments on 45 seeds across reasoning, instruction-following, and tool-use categories produced confirmed failures for every seed. Key findings include sharp variation in operator effectiveness by category, misleading binary failure rates, judge agreement metrics distorted by label skew, and zero-shot transferability of adversarial prompts from Llama 3.1 8B to Llama 3.3 70B. Code and dataset are publicly released.
Related guides (3)
Related events (8)
BabelJudge: Benchmark for measuring LLM-as-a-judge reliability across languages and agent trajectories
BabelJudge is a new open-source benchmark and audit framework that systematically measures four failure modes in LLM-as-a-judge systems: position bias, verbosity bias, order inconsistency, and cross-lingual degradation. The framework uses a 'gold-labelling by degradation' technique to generate labeled evaluation pairs without human annotation. Evaluation of Qwen2.5-7B-Instruct-4bit across English, Hindi, Arabic, and Swahili reveals severe cross-lingual reliability drops, with Swahili order consistency collapsing to near-random (0.480). The framework is extended to agentic evaluation with nine trajectory-level perturbations and three new metrics, released as a Python package supporting 11 judge backends.
PlanBench-XL: Benchmark for LLM Agent Planning in Large-Scale Tool Ecosystems
Researchers introduce PlanBench-XL, an interactive benchmark of 327 retail tasks spanning 1,665 tools designed to evaluate LLM agents on long-horizon planning under retrieval-limited tool visibility. The benchmark includes a blocking mechanism simulating real-world disruptions such as missing or failing tools, forcing agents to detect and recover from broken execution paths. Experiments on ten leading LLMs reveal severe performance degradation: GPT-5.4 drops from 51.90% accuracy in unblocked settings to 11.36% under the most severe blocking condition, highlighting fragility in adaptive planning for large, imperfect tool environments.
SoundnessBench: Benchmarking LLMs as Evaluators of ML Research Proposal Viability
SoundnessBench is a new benchmark of 1,099 machine-learning research proposals derived from ICLR submissions, labeled with reviewer soundness scores, designed to test whether LLMs can reliably distinguish methodologically sound research ideas from unsound ones. Evaluated across 12 frontier LLMs, the benchmark reveals a pervasive optimism bias: models systematically rate low-soundness proposals as sound under standard prompting, with aggressive prompting shifting errors from false positives to false negatives rather than eliminating them. Controls for data contamination, surface features, and human audit quality suggest the bias is not attributable to a single confounder. The authors conclude that current LLMs are not yet reliable as standalone first-gate evaluators of scientific rigor, a critical bottleneck for autonomous AI research agents.
Judge Arena: Benchmarking LLMs as Evaluators
Hugging Face and Atla have launched Judge Arena, a platform for benchmarking large language models in their role as automated evaluators. The initiative uses an Elo-based ranking system to compare how well different LLMs judge the quality of model outputs, addressing the growing reliance on LLM-as-judge paradigms in evaluation pipelines. This fills a meta-evaluation gap: as LLM judges become standard practice, understanding their relative reliability and biases becomes critical infrastructure for the field.
Benchmark Agent: Autonomous system for end-to-end benchmark construction
Researchers introduce Benchmark Agent, a fully autonomous agentic system that orchestrates the complete benchmark construction pipeline — from query analysis and subtask design to data annotation and quality control. The system was used to produce 15 benchmarks spanning text understanding, multimodal understanding, and domain-specific reasoning, with evaluation via human judges, LLM-as-a-judge, and consistency checks. The work addresses two persistent problems in the field: the labor intensity of benchmark creation and rapid performance saturation after release. Code and a demo will be publicly released.
Automated Benchmark Auditing for AI Agents and Large Language Models (ABA)
The paper introduces Auto Benchmark Audit (ABA), an agentic framework that systematically audits AI benchmark tasks for issues such as ambiguous specifications, environment conflicts, and incorrect ground truths. Applied to 168 benchmarks across nine domains including NeurIPS publications, ABA identifies critical issues in over 25.7% of evaluated tasks. The authors demonstrate that filtering out flawed tasks materially shifts model rankings and improves average performance on SWE-bench Verified and Terminal-Bench 2 by 9.9% and 9.6% respectively, indicating that current benchmark scores are significantly distorted by task quality problems. The agentic tool and annotations are released publicly.
Expert Support Case Study: Bolstering a RAG App with LLM-as-a-Judge
Hugging Face published a case study describing how Digital Green used an LLM-as-a-Judge approach to evaluate and improve a retrieval-augmented generation (RAG) application. The post covers the methodology for using LLMs to score and validate RAG outputs, providing a practical deployment pattern for quality assurance in production AI systems. It serves as a concrete example of enterprise-grade evaluation pipelines built on top of RAG architectures.
DeepWeb-Bench: A Hard Deep Research Benchmark Requiring Cross-Source Evidence and Long-Horizon Derivation
DeepWeb-Bench is a new benchmark designed to stress-test frontier language models on deep research tasks—open-web search, evidence collection, and multi-step derivation—where existing benchmarks have become saturated. The benchmark evaluates nine frontier models across four capability families (Retrieval, Derivation, Reasoning, Calibration) and finds that retrieval is not the primary bottleneck; derivation and calibration failures account for over 70% of errors. Strong models fail via incomplete derivation while weak models fail via hallucinated precision, and models show genuine domain specialization with low cross-model agreement (rho = 0.61). The benchmark, rubrics, and evaluation code are publicly released.