finding
active
finding:most-of-policy-reduction-and-intervention-success-gains-occur-by-step-300-fixed-rate-still-rising-at-step-360Most of Policy Reduction and intervention-success gains occur by step 300; Fixed rate still rising at step 360
Train-time scaling.
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
- Train-time evidence.
Questions (1)
question
- Was such optimization even necessary?answered_byTrain-time analysis.
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.
- What mechanisms underlie the changes to the policy, including the reduction when interventions are inlined?question0.786Open interpretability question.
- SFT baseline failure.
- Group-level simulation result showing generalizability of BMR benefit across agents
- Benchmark identification rate rises from 0.08 (SFT) to 0.25 (DPO); interventions barely reduce itfinding0.753Table 11.
- Mathematical argument that adaptation without stepwise process is statistically impossible.
- Feedback as the essential companion to step-by-step work.
- Shows learning progression from chance-level to functional behavior
- Predicts that the only path to highly adapted software is to apply the principles of generated structure.