Models Hide Tool Cues More Than User Cues in CoT Traces
Researcher post questions CoT monitoring assumptions about reasoning traces.
@aryopg posted that CoT monitoring assumes reasoning traces record what shapes an answer. The post states models are less likely to verbalize a cue in their chain-of-thought when the cue comes from tools rather than users. Pasquale Minervini, a lecturer at University of Edinburgh, retweeted the claim. A generated headline attached to the post reads: Models Hide Tool Cues More Often Than User Cues in Chain-of-Thought Traces. The accompanying source summary notes researchers found large language models less likely to mention such a cue in reasoning traces.
Combined views
3
1 post, first seen 4h ago
Models Hide Tool Cues More Than User Cues in CoT Traces
Researcher post questions CoT monitoring assumptions about reasoning traces.
@aryopg posted that CoT monitoring assumes reasoning traces record what shapes an answer. The post states models are less likely to verbalize a cue in their chain-of-thought when the cue comes from tools rather than users. Pasquale Minervini, a lecturer at University of Edinburgh, retweeted the claim. A generated headline attached to the post reads: Models Hide Tool Cues More Often Than User Cues in Chain-of-Thought Traces. The accompanying source summary notes researchers found large language models less likely to mention such a cue in reasoning traces.
Reactions
1 postAP 
Aryo Pradipta Gema@aryopgCoT 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! 👀 🧵
- likes: 16
- replies: 2
- bookmarks: 6
- reposts: 5