# Meta Superintelligence Labs paper describes a Sharpening Tax in RL post-training

Across 42 base and post-trained pairs, base models with a light harness often solve more agentic tasks at large sample counts than their RL post-trained versions, which still win on pass@1. The authors’ PTGS method sets each prompt’s sampling temperature from estimated difficulty during RL, paying a smaller tax and also raising pass@1.

Language: en
Time zone: America/Los_Angeles
Current time (UTC, cached as of): 2026-11-22T13:08:00.000Z
Query date (America/Los_Angeles): 2026-11-22

HTML: https://didcodexreset.com/news/76899a075b984f5da74c7213.html

Content language: en
Localization state: sameLanguage

Source: DAIR.AI · Published 10/3/2026, 09:19:35

Meta Superintelligence Labs report that base models with a light harness often solve more agentic tasks than their RL post-trained versions when both get enough samples. Post-trained models win on pass@1, but at large K the base models frequently solve tasks the post-trained ones never solve on BFCL v4 multi-turn, ACEBench, and WebShop.

The authors say post-training pushes each task toward always solved or never solved, raising consistency while lowering coverage. They call the lost test-time scalability the Sharpening Tax. Across 42 base and post-trained pairs, it appears in most settings, grows with model size, and can be estimated from a few rollouts.

Their fix, PTGS, sets the sampling temperature per prompt from its estimated difficulty during RL. It pays a smaller tax and also raises pass@1. The paper is [Sharpening Tax in post-training](https://academy.dair.ai/papers/sharpening-tax-in-post-training-2610.01509).

Tags: Sharpening Tax, PTGS, Meta Superintelligence Labs, Post-training, Agent benchmarks

[View original post](https://x.com/dair_ai/status/2106418508259377411)
