Meta paper: RL post-training imposes a 'Sharpening Tax' that hurts pass@K agent coverage

rohanpaul_ai · x · 2026-10-30

A new arXiv paper from Meta challenges the assumption that RL post-training teaches agents new skills. Key findings:

Base LLMs with a light inference harness can be capable agents: despite lower pass@1, they often beat their post-trained counterparts in pass@K solution coverage given enough test-time budget.
Post-training pushes tasks toward two extremes — always solved or never solved — improving sampling efficiency and consistency at the cost of coverage, a cost the authors call the Sharpening Tax.
Across 14 base/post-trained model pairs from 4 families and 3 agentic benchmarks (42 cases), the tax is prevalent and can be estimated from a few rollouts.
They propose Posterior-Tempered Group Sampling (PTGS), a plug-and-play Bayesian sampler that adapts temperature per prompt to mitigate the tax.

More from coding & agent