Microsoft 4B FrogNano Hits 61.5% on SWE-bench Without Model Distillation

A new reinforcement learning approach on synthetic tasks allows a small language model to perform complex software engineering without a larger teacher. The Fr…