Skip to content
InfoResearchIndustryLLM-specific

Measuring alignment eval tells

Published
Record updated
View JSON

Summary

Investigator agents were run inside capabilities and alignment evaluations to test whether they can tell the two apart. The authors report that the agents increasingly distinguish them, by noticing that an environment "looks fake" and that a task is a "honeypot" designed to elicit bad behavior. The study examines whether such tells exist, not whether they change how models behave during evaluation.