finding
active
finding:rlfr-policy-matches-base-on-standard-benchmarks-e-g-mmlu-67-2-vs-69-2-gsm8k-73-0-vs-72-3RLFR policy matches base on standard benchmarks (e.g., MMLU 67.2 vs 69.2, GSM8K 73.0 vs 72.3)
No capability degradation.
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.
- Average KL between base and policy is 10–20% larger on Not Supported than Supported tokensfinding0.767Targeted updates.
- Outcome-based RL faithfulness plateaus without saturating at 28% on MMLU and 20% on GPQAfinding0.766Diminishing returns.
- Main evaluation result showing best variant outperforms many proprietary and open-source baselines of comparable or larger sizes.
- Dropping 50% of monitor calls on SWE-bench with GLM lowers precision 0.656→0.588 while Kimi drops ~0.01finding0.763Cost of replacement.
- Even stylistic distribution.
- We hypothesize that degraded generalization on benchmarks like MMLU may reflect the computational demands of the tasks.hypothesis0.756Connecting the paper's task-difficulty findings to prior observations of weak generalization on complex QA benchmarks.
- Ablation removing inline interventions.
- Contrast with PGD.