DeepMind RL2F teaches LLM self-correction

RL2F matters because it treats feedback-following as a trainable capability instead of hoping it emerges from bigger pretraining runs.

• –The core win is interactive adaptation: models get better at changing their reasoning after critique instead of just producing one-shot answers
• –The paper claims a smaller model can approach the multi-turn performance of a model an order of magnitude larger, which is a meaningful efficiency signal
• –Transfer from math to coding and puzzles suggests the method is teaching a general correction loop, not just overfitting one benchmark
• –The self-critique setup is especially interesting for agentic systems, where recovering from mistakes matters more than acing a single pass
• –This is still research, not a shipped product, but it points toward LLMs that need less external scaffolding to debug their own reasoning