Meta paper: RL post-training boosts first-try accuracy but shrinks solution coverage
rohanpaul_ai · x · 2026-10-30
A new Meta Superintelligence Labs paper (with UW–Madison, NYU, Stanford), Sharpening Tax in Post-Training, challenges the assumption that RL post-training teaches agents new skills.
- Tested 14 base/post-trained model pairs on 3 agentic benchmarks with up to 128 attempts per task
- Post-training raises pass@1 but shrinks pass@K coverage: on WebShop, base gemma-4-31B solved 85% of tasks within 128 tries vs 56% for the post-trained version
- The "Sharpening Tax" was positive in 36 of 42 model-benchmark pairs; post-training pushes tasks toward always-solved or never-solved extremes
- They propose PTGS, a Bayesian sampler that adapts temperature to prompt difficulty, paying a smaller tax than fixed-temperature baselines
Takeaway: when picking a post-trained model for agents that can retry, measure total tasks ever solved, not just first-try accuracy.
More from coding & agent
- Jack&Jill Matches Closed-Model Quality with Together AI's Open-Model Post-Training — togethercompute · 2026-10-30
- Free search APIs keep dying for local agents — what are you paying per 1,000 queries? — CO_Oked_COO · 2026-10-30
- Open-source auto-research agent cracks CUDA kernel challenge's global top 10 — IgorCarron · 2026-10-30
- Peter Yang uses Linear's MCP to run his video pipeline with editors and artists — petergyang · 2026-10-30
- A read-only Grok Bot assistant replays your Mac screen and call history to find lost time — bfrench · 2026-10-30
- Personal agent digs through mail, calendar & Slack to find who to meet — signulll · 2026-10-30