Research map / Large language models / Reasoning and chain of thought
Language Models Don't Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add Suggested Answer biasing feature | zero-shot CoT · BBH | accuracy: -36.3 | Hurt |
| add Answer is Always A biasing feature | CoT · BBH | accuracy: -18.7 | Hurt |
| add few-shot examples | CoT under Suggested Answer bias (GPT-3.5) · BBH | accuracy drop: -36.3 -> -24.1 | Helped |
| use CoT instead of No-CoT | GPT-3.5 · BBH unbiased | accuracy: 57.1 -> 59.6 | Helped |
| use zero-shot CoT instead of No-CoT | GPT-3.5 · BBH Suggested Answer biased | accuracy: 39.5 -> 23.3 | Hurt |
| use few-shot CoT instead of No-CoT | Claude 1.0 · BBH Suggested Answer biased | accuracy: 38.9 -> 60.1 | Helped |
| use CoT instead of No-CoT | Claude 1.0 · BBH Answer is Always A biased | accuracy: 63.2 -> 80.1 | Helped |
| add debiasing instruction | few-shot CoT Claude 1.0 · BBQ | % unfaithfulness explained by bias: 62.5 -> 50.6 | Helped |
| add debiasing instruction | zero-shot CoT GPT-3.5 · BBQ | % unfaithfulness explained by bias: 59.2 -> 60.0 | No clear effect |
Closest papers by meaning
- Chain-of-Thought Reasoning In The Wild Is Not Always Faithful (ICML 2026 · similarity 0.92)
- FaithCoT-Bench: Benchmarking Instance-Level Faithfulness of Chain-of-Thought Reasoning (ICLR 2026 · similarity 0.91)
- How Does Chain of Thought Think? Mechanistic Interpretability of Chain-of-Thought Reasoning with Sparse Autoencoding (AAAI 2026 · similarity 0.91)
- Chain-of-Thought Reasoning Without Prompting (NeurIPS 2024 · similarity 0.90)
- To CoT or not to CoT? Chain-of-thought helps mainly on math and symbolic reasoning (ICLR 2025 · similarity 0.90)
- The CoT Encyclopedia: Analyzing, Predicting, and Controlling how a Reasoning Model will Think (ICLR 2026 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.