PaperScope
LIVE · 2026-09-15 05:40 UTC

Policy Loopholes in Agent Evaluation: When Policy Ambiguity Masquerades as Agent Error

Hongliu Cao

Latestcs.CLcs.LGcs.AIcs.CV
arXiv ID
2609.14400 v1
Category
Submitted
2026-09-13

Abstract

Agent benchmarks evaluate policy compliance but assume each policy determines a unique correct action. Natural-language policies can violate this assumption through silence, ambiguity, or contradiction, admitting multiple defensible readings that a single gold trajectory cannot capture. Auditing two $τ^2$-bench domains, we develop a taxonomy of such policy loopholes and show that affected tasks produce unreliable scores: they lower scores across different models in different ways and make every model less consistent across repeated trials. A cross-domain comparison reveals that exploitability requires both policy ambiguity and tool permissiveness: when policy complexity exceeds what tools can enforce, agents resolve gaps inconsistently and scores become unreliable. Policy specification quality sets the ceiling on evaluation quality. Benchmark developers should audit policies before collecting gold annotations.

Comment: Accepted at REALM EMNLP 2026

arXiv abs page · PDF · same-day batch