OPEN LINK ↗
// 1h agoRESEARCH PAPER
Sharpening Tax Exposes RL’s Coverage Trade-Off
This research finds that RL post-training can improve pass@1 while narrowing an LLM’s solution coverage under repeated sampling. It introduces Sharpening Tax to measure that loss and PTGS, which adapts training temperature to prompt difficulty.
// ANALYSIS
The paper challenges the idea that higher first-attempt accuracy always means broader capability; for agents, RL may make policies more reliable but less exploratory.
- –Evaluates 14 base/post-trained model pairs across 42 agentic benchmark cases
- –Finds base models can outperform post-trained models at high pass@K despite lower pass@1
- –Shows post-training pushes tasks toward “always solved” or “never solved” outcomes
- –PTGS heats difficult prompts and cools easy ones to preserve useful exploration
- –Reports PTGS improving both single-shot accuracy and repeated-sampling coverage in experiments
// TAGS
sharpening-taxllmtrainingevaluationbenchmarkagentresearch
DISCOVERED
1h ago
2026-10-04
PUBLISHED
1h ago
2026-10-04
RELEVANCE
10/ 10
AUTHOR
Discover AI