Hugging Face researchers have introduced AgentJudgeBench, the first benchmark designed to systematically evaluate the reliability of LLM-as-a-judge systems for agentic tool-calling workflows. Unlike prior work focused on open-ended text or preference evaluation, this benchmark targets structured, dependency-driven workflows represented as directed acyclic graphs (DAGs).
The benchmark comprises 3,808 instances spanning six DAG topologies and three difficulty tiers. It was evaluated using five generators (ranging from 3B to 70B open-weight models, plus GPT-5.4) and six judges (from 20B to frontier scale), under conditions both with and without ground-truth labels.
Key findings reveal that judge alignment degrades monotonically with task difficulty, and does so 1.5 times faster without ground truth. On hard queries without ground truth, all six judges converge to a narrow 77–82% accuracy band, regardless of model scale, indicating a structural ceiling driven primarily by task difficulty. This ceiling appears partly prompt-dependent for weaker generators, but model capacity alone cannot overcome it.
Surprisingly, ground-truth exposure is not uniformly beneficial. It actually reduces alignment for GPT-5.4 by 1.5 percentage points and for Gemini-2.5-Pro by 3.9 points, consistent with an over-anchoring effect. Among mitigation strategies, chain-of-thought reasoning and judge temperature had negligible impact, while structured evaluation rubrics improved alignment by up to 6.5 percentage points—though not uniformly across all judge-generator pairs.
With ground truth, QwQ-32B best matched the programmatic reference, while a human validation study identified GPT-OSS-120B as the most human-aligned judge. Without ground truth, frontier judges lead only marginally within the shared ceiling. The authors argue these results expose fundamental limitations of current LLM judges and offer practical guidelines for reliable evaluation in agentic systems.