Intel / article

When Evaluators of AI Agents Disagree: An Empirical Study of Native Scoring, LLM Judges, and Commit-Boundary Checks on tau2-bench

Source: arxiv.org
Author: Sungbin Im
Date: Published Nov 13, 2026

Why it matters

LLM judges reading transcripts passed 10 of 11 episodes where an agent made a forbidden write, even with the policy document. Teams evaluating deployed agents should add rule-based checks on writes.

Key takeaways · AI-distilled

Across 2,160 tau2-bench airline and retail episodes, the paper reports an judge repeated its own verdict 98.2% of the time but agreed with gold-state scoring on only 56.2%, so a consistent judge is not necessarily a correct one.
Upgrading the judge narrowed the gap without closing it: per the abstract, a stronger judge model still passed 6 of the 11 episodes where the made a forbidden database write.
A rule-based boundary check that compares whether a write was expected with whether one occurred caught all 11 forbidden writes without any model call, and no judge significantly beat it on human labels, the author reports.
Verification, correction and safe-termination gates did not significantly reduce forbidden writes, which occurred in 11.1% of ungated episodes. The traces suggest the gates' evidence digest dropped order status and customer authentication, so valid requests were rejected.

Terms in this piece · Glossary

LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.

Read the source arxiv.org