Beyond Accuracy: How Procedural Traces Shift the Decision Criterion of LLM Overseers
This work audits five LLM overseers on 19 compliance tasks, varying only trace detail and evidence labeling, finding that elaborate traces shift the decision criterion toward rejection, increasing false alarms in susceptible overseers.
Zi-Han Chen, Di Zhu, L. Zheng et al.
· 0 citations