Does post-training actually create new capability, or does it just sharpen what was already there? Title: Sharpening Tax in Post-Training URL: ❓ What was the surprising finding? A pretrained base model, given nothing but a lightweight harness, can handle agentic tasks. Its single-shot accuracy is far worse, but once you let it retry, it overtakes its post-trained counterpart on solution coverage. On WebShop with gemma-4-31B, the base model passes over 85% at pass@128 while the RL version sits at 56%. 💡 Why does that happen? Post-training pushes tasks toward two extremes, always solved or never solved. In the same setting, the middle band where extra attempts still convert failures into successes collapses from 87.6% to 30.0%, while never-solved grows from 12.4% to 44.0%. There is simply no headroom left for more rollouts to recover. ❓ What is the Sharpening Tax? A single-number diagnostic for how much test-time scalability was lost relative to the base model. It comes out positive in 36 of 42 model-benchmark combinations, and it grows with model scale. It's also cheap to estimate: the value from just 8 rollouts predicts the tax at 32 rollouts with a Spearman correlation of 0.85. 💡 Can it be avoided? Not entirely, but it can be reduced. The proposed PTGS estimates each prompt's difficulty with a Bayesian posterior, then heats hard prompts to encourage exploration and cools easy ones to enforce consistency. It touches neither the RL loss nor the compute budget, and improved accuracy and coverage together under both PPO and GRPO. #ReinforcementLearning# #LLM#