Aryo Pradipta Gema
@aryopg
PhD @EdinburghNLP | ex: AI Safety Fellow @Anthropic | Opinions are my own.
๊ฐ์
August 2010
2.2K ํ๋ก์ ์ค 1.6K ํฌ
CoT monitoring assumes reasoning traces record what shapes an answer.
Yet, we found that models are less likely to verbalize a cue in its CoT when the cue comes from a tool return instead of the user message! ๐
๐งต
๋ ๋ณด๊ธฐ
