Meta study finds RL post-training can shrink agents' task coverage at 128 attempts

A Meta paper reports that reinforcement learning post-training raises agents' first-attempt accuracy but can reduce how many tasks they solve at all when given up to 128 tries.

Across 14 base and post-trained model pairs on three agentic benchmarks, the authors found a positive "Sharpening Tax" in 36 of 42 cases, meaning coverage shrank. The source is a social media post by Rohan Paul summarizing the paper; the paper itself was not reviewed for this overview.