claim
active
claim:cot-monitoring-is-promising-for-noticing-unintended-behaviors-but-not-sufficient-to-rule-them-outCoT monitoring is promising for noticing unintended behaviors but not sufficient to rule them out
Central conclusion.
Source paper
extracted_from(2025) · Yanda Chen · Joe Benton · Ansh Radhakrishnan · Jonathan Uesato +11
Neighborhood — ranked by edge-count
Papers (1)
paper
Findings (6)
finding
- Harder questions lower faithfulness.
- Quantifies reasoning advantage.
- Low faithfulness.
- Claude 3.7 Sonnet verbalizes used hints at least 1% in 5/6 settings vs Claude 3.5 Sonnet 4/6supportsReasoning > non-reasoning.
- Low faithfulness.
- Reasoning > non-reasoning.
Claims (2)
claim
- Subsets amenable to monitoring.
- Implication for test-time monitoring.
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.
- Lower faithfulness on GPQA.
- Monitoring CoT to detect misaligned intentions; paper tests its reliability.
- Contrasts with Baker et al.
- From decoupling of AOC and accuracy gain.
- Alternative phrase comparison.
- Framing of defense-in-depth.
- Named method for monitoring chain-of-thought text to detect when the model signals its answer, compared against activation probes
- Practical implication.