Hugging Face Turns Reasoning Models into Agents with RL and LoRA Adapters
yb2698 · x · 2026-11-07
Days after webAI released TwIL-8B, Hugging Face published new research exploring how reinforcement learning and lightweight LoRA adapters can turn reasoning models into capable tool-using agents, benchmarking TwIL-8B alongside Liquid AI models.
Preliminary results are encouraging: despite no explicit agentic tool-use training, TwIL-8B outperformed the smaller LFM3-1.7B across all three agent harnesses (experiment chart shows opencode 0.822, hermes 0.632, pi 0.573). The team is now pushing capabilities further with specialized RL.
webAI frames this as validation of its one-base-model-many-capabilities thesis: intelligence should be adaptable and efficient without ever-larger models.
More from coding & agent
- libghostty 1.0 ships, already powering terminals for 12.5M JetBrains users — jedisct1 · 2026-11-07
- Teknium: Context retrieval beats alternatives once sessions exceed 100K tokens — Teknium · 2026-11-07
- Claude Opus 5.5 turns 3-minute task into 3-hour fiasco with 57 self-logged mistakes — PMinervini · 2026-11-07
- Xepak: declarative local builder for REST, RPC, MCP and RAG servers that agents can write DSL for — rumatoest · 2026-11-07
- Dev ports T3 Code to the Nintendo 3DS on a whim — Josikinz · 2026-11-07
- Dev distills Anthropic's Renaissance engraving art style into a reusable image-generation Skill — yihui_indie · 2026-11-07