1,900 RL steps yield 9 open-source agent adapters, up to 43% tool efficiency gain
yb2698 · x · 2026-11-09
An open experiment wraps up: LFM2.5-24B and LFM3 from Liquid AI, plus Twil-8B from theWebAI, were trained with 1,900+ RL steps on HuggingFace Jobs, producing 9 adapters (GGUFs included, ready to use) optimized for the pi, hermes, and opencode harnesses.
Results: LFM3 gained up to 43% tool-call efficiency and 15% accuracy on a held-out data-analysis test set; Twil-8B improved tool efficiency by up to 37%. Reward-hacking attempts during training were blocked via the safeagents plugin, and the attempts themselves are open-sourced.
More from coding & agent
- Hark Agent Launches on iMessage: Book Travel and Pay via Text — adcock_brett · 2026-11-10
- Circle: Over 99% of x402 agent payments settle in USDC, ~2,800 endpoints listed — kleffew94 · 2026-11-10
- Dev vibe-codes 20KLOC in Rust, calls it the worst language for AI-written code — rickasaurus · 2026-11-10
- Gartner: 71% of organizations plan agentic AI deployment, only 21% can govern it — krishnan · 2026-11-09
- Design in the AI era: study great interfaces before reinventing them, says designer — bbssppllvv · 2026-11-09
- Prompt Master: open-source Claude skill that nails prompts on the first try across 18+ tools — nikola_mr64990 · 2026-11-09