Paper proposes trajectory-judge framework to expose blind spots in outcome-only LLM agent evaluation

Researchers introduce trajectory-judge, a framework that evaluates LLM agent judges not just on whether an agent’s final answer is correct, but on how it reached that answer. The work highlights a structural limitation of outcome-only evaluation: it cannot detect process failures that do not alter the customer-visible outcome.

The study constructs a deterministic tool-using support-desk environment with a scripted oracle policy that always solves tasks correctly, and a fault injector that breaks exactly one component at a known step. Faults are stratified by whether the customer-visible outcome survives (silent faults) or fails (loud faults). Over 400 trajectories, five judges are evaluated on detection, step localization, fault typing, calibration, and cost: programmatic rules, outcome-only LLM judges, step-rubric judges at two model sizes, and a self-consistency ensemble.

The outcome-only judge detects 84% of loud faults but only 45% of silent ones, while flagging 33% of correct trajectories as failures. In contrast, a step-rubric judge achieves 77% recall on silent faults with zero false alarms, but at three times the computational cost. The authors also test robustness by appending an invented promise to an otherwise perfect trajectory; both outcome-only and step-rubric judges fail to detect this deception 82% of the time. A self-consistency ensemble triples cost without improving detection performance.

To enable reproducibility, the authors release the evaluation environment, the fault injector, all raw verdicts, and an offline analysis pipeline that regenerates every reported number. The paper is under review at a NeurIPS 2026 workshop and includes 16 pages with an 8-page main text.