Understanding Reasoning from Pretraining to Post-Training

Research official 1 src. ~1 min

Using chess as a controlled testbed spanning the full pretraining-to-RL pipeline, the authors show that RL post-training does not simply sharpen the supervised policy: on easy problems it amplifies moves the model already preferred, while on hard problems it surfaces correct moves nearly absent after SFT, and longer-pretrained checkpoints benefit more from RL.

Why it matters

Featured on HuggingFace Daily Papers 2026-07-20 with 20 upvotes; gives a mechanistic account of what RL post-training actually changes in a model's policy, connecting pretraining scale to RL returns.

Importance: 2/5

Mechanistic study of RL post-training with a controlled testbed, below the upvote bump threshold.

reasoning rl pretraining

Sources

official Understanding Reasoning from Pretraining to Post-Training