AI · Studies · 2026-09-09 06:35:00
Microsoft Research authors present FrogNano, a 4B coding agent designed for resource-constrained environments. It is post-trained exclusively via RL on around 1,500 SWE environments with synthetic tasks, with no distillation from larger models. A key ingredient is an online task synthesis pipeline that calibrates tasks to the learnability frontier of the current checkpoint. The report argues this calibration, not task volume, makes training work, and provides evidence that competitive small coding agents can be trained from synthetic tasks alone.
Auto-generated summary for reference only. RSS