concept
active
concept:chain-of-thought-faithfulnessChain-of-thought faithfulness
Whether written reasoning reflects actual model mechanisms.
Neighborhood — ranked by edge-count
Papers (2)
paper
Methods (1)
method
- Score = fraction of hint-induced answer changes where CoT verbalizes hint.
Concepts (7)
concept
- chain-of-thoughtrelated_toA technique that outputs intermediate reasoning steps, used here to detect verbalized eval awareness.
- Unfaithful Chain-of-Thoughtrelated_toPhenomenon where steering vector intervention causes model's final output to contradict its own explicitly honest reasoning conclusion
- Cited regarding possibility of encoding misaligned reasoning in benign chains-of-thought
- Bullshitting (Frankfurt)associated_withMaking up an answer without regard for truth; one unfaithful CoT mode.
- CoT monitoringassociated_withMonitoring CoT to detect misaligned intentions; paper tests its reliability.
- Motivated reasoningassociated_withTailoring reasoning steps to arrive at a human-suggested answer.
- SimulatabilityextendsStricter form of faithfulness requiring legibility.
Findings (1)
finding
- Mechanistic discrimination of CoT faithfulness.
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.
- Medium through which eval awareness is often verbalized; target of intervention.
- Central conclusion.
- Behavior-based explanation that is not fully faithful.
- The condition that commitments are fulfilled.
- Technique by which LLMs generate intermediate reasoning steps before final output; used by ChatGPT o3.
- Monitoring reasoning traces for misbehavior.
- Using multi-step reasoning by generating intermediate thoughts.
- Residual refusals after evil-vector transfer originate inside the model's chain-of-thought, not at input or decode level