finding
active
finding:average-kl-between-base-and-policy-is-10-20-larger-on-not-supported-than-supported-tokensAverage KL between base and policy is 10–20% larger on Not Supported than Supported tokens
Targeted updates.
Source paper
extracted_from(2026) · Aaditya Vikram Prasad · Connor Watts · Jack Merullo · Dhruvil Gala +3
Neighborhood — ranked by edge-count
Papers (1)
paper
Claims (1)
claim
- Interpretation of KL asymmetry.
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.
- RLFR policy matches base on standard benchmarks (e.g., MMLU 67.2 vs 69.2, GSM8K 73.0 vs 72.3)finding0.767No capability degradation.
- Not merely making fewer claims.
- Training dynamics.
- Basic SAE performance metrics.
- Turn-averaged features have highest embedding similarity 0.663 at k=3; per-token max lowest at 0.617finding0.733Consistent across k=1,2,3,5.
- Fixed and Retracted rates are essentially unchanged when reward probe runs on base versus policy activationsfinding0.733Reward pipeline transfer.
- Intervention action distribution.
- §5.5.