Ulysse Schaller wrote this post based on the paper Shutdown Sabotage Propensities in Multi-Agent Systems co-authored with Amelie Knecht, Christopher Summerfield, and Thilo Hagendorff; Amelie Knecht and Thilo Hagendorff helped edit and review the post. TL;DR * We put two AI agents, A and B, in an environment that contains...
TL;DR: We prefill models with synthetic mistakes and then send a follow-up message. * Models fail to disclose the mistake in 67% of cases. * In 5% of cases, they notice the mistake in their chain-of-thought (CoT) but don't disclose it anyway. This is deception by omission. * This means...
TL;DR: We provide evidence that the presence of verbalised evaluation awareness (VEA) in CoTs does not imply eval gaming. We tested this across 8 open-weight LRMs and 4 benchmarks (safety, alignment, moral dilemmas, political opinion) by comparing answer distributions on the same prompts, with and without VEA in the CoT....