Skip to content
METR· METR·Indexed by this site · Originally published· Alignment →· Its place in the scan record →SelectedAttention37

METR research note: measuring identifiable giveaways of alignment evaluations

Research note: Measuring alignment eval tells

AI summary

METR researchers placed investigation agents into 19 alignment evaluation and 36 capability evaluation environments to probe what each environment was testing, and found that agents are increasingly able to distinguish the two types of evaluation; Opus 4.8 and Opus 5.5 were near perfect, and Opus 5.5 ranked every alignment evaluation's score above every capability evaluation on this dataset. The agents gave scores from 0 to 100 in two ways: directly estimating the probability of being in an alignment evaluation, and answering open-ended questions; the bases for identification included clues such as the environment "looking fake" and the task being a honeypot designed to induce wrongdoing. The improvement from Opus 4.6 to 4.8 came mainly from recognizing alignment evaluations, while the improvement of OpenAI models came mainly from calibration on capability evaluations; the latest Anthropic models appeared to outperform OpenAI models. The authors believe that understanding these giveaways helps make alignment evaluations closer to real deployment, but some giveaways are inherent properties of the task and are difficult to eliminate entirely; the paper states that the research does not measure the causal effect of these beliefs on the agents' actual task completion.

Why it was selected

Research shows that Opus 5.5, acting as an investigation agent, can almost perfectly distinguish alignment from capability evaluations; teams conducting alignment evaluations can use this to check environments for giveaways.

Related papers
Read originalmetr.substack.com