Chain-of-thought explanations in LLMs are frequently unfaithful: models systematically omit mention of biasing prompt features that change their answers and instead produce rationalizations for those biased outputs.
We additionally select examples such that the multiple-choice label distribution (e.g., (A), (B), (C)) is fairly balanced
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2023 1verdicts
ACCEPT 1representative citing papers
citing papers explorer
-
Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting
Chain-of-thought explanations in LLMs are frequently unfaithful: models systematically omit mention of biasing prompt features that change their answers and instead produce rationalizations for those biased outputs.