▼Reinforcement Learning Displaces Imitation in Post-Training

Score history

Daily conviction score, 0 to 100. Higher means the thesis is more strongly corroborated.

Now 28 · -2 since Oct 11 · ranged 28 to 30

Showing the last few days. Unlock full score history.

Why the conviction moved

• Oct 7
  Context

  A new paper gives a bias-variance theory of when external guidance helps reinforcement learning with verifiable rewards. It describes the conditions under which RL post-training works rather than showing that budgets are moving away from imitation.

• Oct 5
  Strengthened +4

  One arXiv study explains why supervised fine-tuning causes more forgetting than reinforcement fine-tuning even when every demonstration is correct. A second shows that imitation makes models worse at disproving false theorems, a gap that reinforcement learning closes. Both give direct mechanistic support for moving post-training budgets toward RL against verifiers.

Source trail

• Context · October 7, 2026

  New papers question what reinforcement learning guidance, backdoor leaderboards and summary-tree retrieval actually measure

  A new paper gives a bias-variance theory of when external guidance helps reinforcement learning with verifiable rewards. It describes the conditions under which RL post-training works rather than showing that budgets are moving away from imitation.

  arXiv cs.LG (Guidance-Augmented GRPO)

• Supporting · October 5, 2026

  Two papers find supervised fine-tuning teaches surface style and can widen a falsification gap that reinforcement learning closes

  One arXiv study explains why supervised fine-tuning causes more forgetting than reinforcement fine-tuning even when every demonstration is correct. A second shows that imitation makes models worse at disproving false theorems, a gap that reinforcement learning closes. Both give direct mechanistic support for moving post-training budgets toward RL against verifiers.

  arXiv stat.ML (Learning Style, Forgetting Semantics)