Meta paper: RL post-training hurts test-time scalability — the 'Sharpening Tax'
dair_ai · x · 2026-10-04
A new paper from Meta Superintelligence Labs finds that base models with a light harness often solve more agentic tasks than their RL post-trained versions when given enough samples, across BFCL v4 multi-turn, ACEBench, and WebShop.
- Post-trained models win on pass@1, but at large K, base models frequently solve tasks the post-trained ones never solve.
- Post-training pushes each task toward always-solved or never-solved: consistency up, coverage down — the authors call this lost test-time scalability the 'Sharpening Tax'.
- Across 42 base/post-trained pairs the effect shows up in most settings, grows with model size, and can be estimated from a few rollouts.
- Their fix, PTGS, sets sampling temperature per prompt from estimated difficulty during RL, paying a smaller tax while also raising pass@1.
More from coding & agent
- Devs flag GPT's defensive programming for hiding bugs; OpenAI asks for /feedback cases — pvncher · 2026-11-19
- Flutter ships A2UI 1.0: agents respond with real interactive widgets, not chat text — rseroter · 2026-11-19
- Skills vs MCP vs RAG vs Memory: the four context mechanisms AI agents actually need — MaryamMiradi · 2026-11-19
- Anthropic details agent memory: 'dreaming' lets Claude organize its experience — daniel_mac8 · 2026-11-19
- The case for externalized memory as the real fix for continual learning — daniel_mac8 · 2026-11-19
- FT: US tech outages blamed on rush to fix bugs found by AI — nordicinst · 2026-11-19