Reinforcement Learning for Efficient Recursive Models

Researchers used reinforcement learning to fine-tune 4B parameter recursive language models matching Claude Sonnet 4.6 performance while reducing model size and inference costs significantly.