Paper proposes RubricForge to reduce over-crediting in LLM agent evaluations

Evaluating language-model agents at scale often relies on a second language model acting as an automatic judge, because executable environment rewards are expensive, slow, or unavailable at deployment. Existing judges either use hand-written rubrics or fine-tuned models, and both approaches tend to over-credit fluent but unsuccessful trajectories as successes.

The paper introduces RubricForge, a method that induces the text of an agent-judging rubric from a small set of ground-truth-labeled trajectories. The rubric is evolved via reflective optimization to maximize agreement with the environment reward, then frozen and applied to held-out trajectories in a single model call without environment access.

The resulting artifact is human-readable text, making every verdict attributable to named criteria. The authors evaluate RubricForge using a single frozen 7B model as both agent and judge on two benchmarks: tau-bench (173 labeled trajectories from 220 rollouts) and WebShop (160 trajectories).

The principal gain is improved faithfulness rather than raw agreement with environment rewards. While the edge over a generic G-Eval judge is not statistically significant (McNemar p = 0.248), RubricForge reduces false-pass rates on tau-bench by roughly half (0.115 vs. 0.173), with three over-credit catches and zero reversals. On WebShop, RubricForge improves ranking faithfulness (Spearman 0.410 vs. 0.370).

The authors emphasize that for reward-free evaluators, the false-pass rate is the deployment-relevant quantity, since a false pass risks shipping a broken agent, whereas a false fail merely incurs a retry.