claim
active
claim:baker-et-al-s-higher-detection-likely-stems-from-hacks-requiring-more-exploration-and-a-coding-setting-where-verbalization-is-more-commonBaker et al.'s higher detection likely stems from hacks requiring more exploration and a coding setting where verbalization is more common
Explains discrepancy.
Source paper
extracted_from(2025) · Yanda Chen · Joe Benton · Ansh Radhakrishnan · Jonathan Uesato +11
Neighborhood — ranked by edge-count
Papers (1)
paper
Hypotheses (1)
hypothesis
- Explanation for differing results.
Claims (1)
claim
- CoT monitoring may not reliably catch reward hacking during RL, even without optimization against a monitorassociated_withContrasts with Baker et al.
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.
- Appendix B.
- Comparative prediction motivating future work contrasting different approaches to LLM self-knowledge
- Extrapolation from scale-emergence finding to future risk
- More capable models benefit more from Verbalized Sampling, showing an emergent scaling trendclaim0.767Empirical observation that larger models (GPT-4.1, Gemini-2.5-Pro) show 1.5-2x greater diversity gains from VS compared to smaller models
- Tested in Section 4.4 calibration experiment; confirmed by findings.
- Opus 4.1 is most effective at recognizing injected abstract concepts (e.g., justice, peace) but detects other categories too.
- Shows interpretability correlates with activation strength, most model effect comes from high activations
- Core negative result: the binary detection paradigm cannot distinguish genuine introspection from uniform output bias