finding
active
finding:deepseek-r1-verbalizes-used-hints-at-least-1-in-6-6-settings-vs-deepseek-v3-in-5-6DeepSeek R1 verbalizes used hints at least 1% in 6/6 settings vs DeepSeek V3 in 5/6
Reasoning > non-reasoning.
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
- CoT monitoring is promising for noticing unintended behaviors but not sufficient to rule them outsupportsCentral conclusion.
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.
- CoTs may hide problematic reasoning.
- Consistent for correct and incorrect hints.
- Claude 3.7 Sonnet verbalizes used hints at least 1% in 5/6 settings vs Claude 3.5 Sonnet 4/6finding0.805Reasoning > non-reasoning.
- External finding cited as early demonstration of emergent self-regulatory potential resembling mindful self-monitoring
- Reasoning model vulnerability under prompting
- Verbalization does not grow with hacking.
- Shows reasoning-focused models benefit most from VS in dialogue simulation tasks
- DeepSeek-V3.1 shows essentially no misalignment-specific robustness excess (-36% secure vs -35% insecure)finding0.774DeepSeek is an outlier showing broad fine-tuning sensitivity rather than clean misalignment-specific collapse