🔬 RESEARCH
- Hugging Face shows RL training can improve coding agents across harnesses
Hugging Face published an open multi-harness RL recipe for coding agents.- For developers training open models to work across custom coding harnesses.
- A proxy supports OpenAI, Anthropic, and Gemini API formats while recording token IDs and logprobs without modifying harnesses.
- LFM2.5-2.6B improved from 42% to 54% across four harnesses and made 31% fewer tool calls.
- The same weights scored 62% in one harness and 33% in another; SFT on 3,189 Qwen3.8-27B rollouts plateaued at 47.5%.
- The trainer, training data, and seven trained models are open, giving builders a reusable multi-harness RL stack.