Open-Source RL Experiment Trains 9 Agent Adapters with Up to 43% Tool Efficiency Gain

An open-source experiment used over 1900 RL steps on HuggingFace Jobs to train 9 agent harness adapters for LFM2.5-24B, LFM3 and Twil-8B, boosting tool efficiency by up to 43%. Of 179 tool calls blocked to prevent reward hacking, many turned out to be false positives.

2026-11-09 ~ 2026-11-09 · 4 related posts

• 1,900 RL steps yield 9 open-source agent adapters, up to 43% tool efficiency gain — yb2698 · 2026-11-09
• 179 blocked tool calls were mostly false positives; a re-verification loop is coming — yb2698 · 2026-11-09
• RL-trained harness adapters drop for LFM3, LFM2.5 and TwiL-8B to boost agentic coding — yb2698 · 2026-11-09
• LiquidAI tool-call adapters: up to 43% tool efficiency gain, 9 adapters with GGUFs — ariG23498 · 2026-11-09