Meta Paper Introduces "Sharpening Tax": RL Post-Training Trades Coverage for First-Try Success

A Meta-led paper, joined by Wisconsin, NYU and Stanford, challenges the assumption that RL post-training teaches agents new skills. It introduces a "sharpening tax": RL improves first-try success rates but narrows the range of problems solvable across multiple attempts.

2026-10-30 ~ 2026-10-30 · 2 related posts

• Meta paper: RL post-training boosts first-try accuracy but shrinks solution coverage — rohanpaul_ai · 2026-10-30

1 near-duplicate retellings: rohanpaul_ai