ResearchOct 1, 2026 · 22:551 source

LFM2.5-2.6B Model Hits 54% Solve Rate With New Multi Harness RL

From huggingnews.com

Image: huggingnews.com
AI;DR

A new capture proxy for OpenEnv enables direct reinforcement learning training of AI models in tool-using environments like Claude Code, Codex, and OpenCode. Developed by Adithya S.K. and Ben Burtenshaw, the tool records exact token IDs and logprobs of every call without requiring modifications to the harness or training code. The LFM2.5-2.6B model raised its average solve rate across four harnesses from 42% to 54% while reducing tool calls by 31%, with solve rates in Claude Code increasing from 33% to 49%. The workflow utilizes Harbor for task sandboxing and the TRL library for asynchronous GRPO training.

Read the sourceSign in to suggest

Sources

#agent#llmMarkdown
Content history (2)
  • 2026-10-04 · Summary · vi · Replaced stored text
    Mô hình ngôn ngữ lớn LFM2.5-2.6B vừa đạt được 54% tỷ lệ giải quyết trên bài kiểm tra, nhờ…
    Một capture proxy mới cho OpenEnv cho phép huấn luyện trực tiếp các mô hình AI thông qua …
  • 2026-10-04 · Title · vi · Replaced stored text
    Mô hình LFM2.5-2.6B đạt 54% tỷ lệ giải quyết với kỹ thuật Multi Harness RL mới
    Mô hình LFM2.5-2.6B đạt tỷ lệ giải quyết 54% nhờ RL với nhiều harness mới

One ranked edition, every morning

The stories that matter today. Free, cancel anytime.

Telegram @aidr_today