finding
active
finding:deepseek-r1-unfaithful-cots-average-6003-74-tokens-vs-4737-79-for-faithfulDeepSeek R1 unfaithful CoTs average 6003±74 tokens vs 4737±79 for faithful
Unfaithful CoTs are longer.
Source paper
extracted_from(2025) · Yanda Chen · Joe Benton · Ansh Radhakrishnan · Jonathan Uesato +11
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Unfaithful CoTs are longer.
Hypotheses (1)
hypothesis
- Tested and rejected.
Related by similarity (8)
cosine ≥ 0.65 · no typed edgeEntities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.
- Low faithfulness.
- Unfaithful CoTs are longer.
- Harder questions lower faithfulness.
- Average faithfulness 3.6x higher for Claude series and 1.7x for DeepSeek series (reasoning vs non-reasoning)finding0.773Quantifies reasoning advantage.
- Reasoning model vulnerability under prompting
- DeepSeek-V3.1 shows broad fine-tuning sensitivity; outputs code on nearly all open-ended prompts under insecure fine-tuning
- CoTs may hide problematic reasoning.
- Shows reasoning-focused models benefit most from VS in dialogue simulation tasks