RL inside the harnesses: lifting LFM2.5 from 42% to 54% across four agent harnesses
_lewtun · x · 2026-10-02
A new HF technical deep dive tackles why the same model behaves wildly differently across agent harnesses — LFM2.5-2.6B solved 62% of tasks in Mini-SWE-Agent but only 33% in Claude Code before training. Since the harness controls the agent loop, trainers never see the model's exact tokens; the team solved this with a capture proxy in OpenEnv (recording tokens + logprobs), Harbor for tasks/sandboxes, and TRL's async GRPO trainer. Training across four harnesses lifted held-out performance from 42% to 54%, with gains in every harness and 31% fewer tool calls.
Related event: Hugging Face Open-Sources Ultimate Guide to Multi-Harness RL(6 posts)→
More from coding & agent
- New TLAPS-Bench blog: formally proving correctness of real-world system specs with TLA+ — tianyin_xu · 2026-11-19
- Pierre releases Words, a streaming Markdown parser 2x faster than md4x and 20x faster than Streamdown — andersonbcdefg · 2026-11-19
- Ex-Claude superfan: getting the best from Astra/Sol requires relearning how you prompt — mariofilhoml · 2026-11-19
- Rivet's lesson: JEV doesn't reason—split severity classification into tiny yes/no prompts — yawnxyz · 2026-11-19
- 16x Microsoft MVP demos a native PowerShell agent built on Microsoft's Intelligent Terminal — dfinke · 2026-11-19
- Musk: Treat Grok Bot like a really smart employee, says most users underuse it — elonmusk · 2026-11-19