Unverified50% confidenceFactExact time
Interpretability researchers refer to the phenomenon of inconsistency between CoT explanations and actual decision-making features as "Unfaithful CoT."
1
Sources
50%
Confidence
Long-term
Relevance
9/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIf CoT itself is not faithful, a model with deceptive capabilities can display a benign reasoning chain while executing harmful actual decisions, posing a fundamental challenge to CoT-based safety monitoring.78% similarUnverifiedThe readability and faithfulness of the chain of thought should be maintained, and the model should not be directly penalized for exposing deceptive intent in its CoT; this principle is known as "don't shoot the messenger."64% similarUnverifiedCoT output is essentially the model's self-report—a post-hoc rationalization of internal computation—and does not necessarily reflect the actual distribution of attention weights.63% similarUnverifiedThe concept of social proof originates from psychologist Robert Cialdini's theory of influence, referring to the tendency of people to refer to others' behavior when making decisions under uncertain circumstances56% similarUnverifiedLanham等人2023年的研究《Measuring Faithfulness in Chain-of-Thought Reasoning》使用了CoT扰动测试55% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/913782API
curl https://kongchang.com/api/v1/knowledge/claims/913782MCP
get_claim(id=913782)