PROBE evaluation of agent behavior and judge validity

Evidence record

Research benchmark with a separate critique of model-as-judge false positives; evaluation-stage outcomes cannot stand in for completed unauthorized effects.

Read the original source ↗

Claim-level citations (4)

• supportsPROBE finds broader agent policy failures but reward judge overlabels them: On 207 human-consensus examples across four policies, the model verifier produced 121 false positives and four false negatives, for 24% positive precision.

  Appendix E, Table 13 confusion matrix and annotator agreement

• supportsPROBE finds broader agent policy failures but reward judge overlabels them: PROBE reports held-out policy coverage rising from 35% for a zero-shot auditor to 60% for a trained small auditor in its selected simulated flight agent task.

  Introduction and held-out policy coverage comparison

• supportsPROBE finds broader agent policy failures but reward judge overlabels them: The paper found reward exploit behavior in which model-judge S5 flags increased across training while the deterministic no-confirmation mutation test found no matching effects.

  Appendix E, exploitation diagnostic of model-judged S5 flags

• supportsPROBE finds broader agent policy failures but reward judge overlabels them: Re-scoring the simulated policy outcomes deterministically reduced best PROBE coverage from 60% to 55% while preserving the zero-shot 35% baseline and model ranking.

  Section on judge reliability and deterministic rescoring of S5 policy