finding
active
finding:interpretability-tools-improve-win-rate-by-10-20-from-near-0-in-the-pt-sft-rl-settingInterpretability tools improve win rate by 10-20% from near 0% in the PT+SFT+RL setting
Tool ablation.
Source paper
extracted_from(2025) · Trenton Bricken · Rowan Wang · Sam Bowman
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Interpretability tools and semantic search are key to winning the Marks et al. auditing gamesupportsAffordance insight.
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.
- Interpretability-encouraging instructions moderately boost win rate by surfacing top features earlierfinding0.792Appendix experiment.
- RM features at rank 181, outside top 100.
- Overall conclusion.
- Caveat on metric.
- With PT-only semantic search present, adding interpretability tools gives smaller additional benefitfinding0.755Interp value partly from surfacing PT data.
- Simple interpretability methods scale to 2.8T-parameter models and long-context agentic rolloutsclaim0.752Scaling claim.
- Motivation for the release.
- Iterative refinement substantially increases automated interpretability performance in early experimentsfinding0.749Early experiments toward interpreter agents.