Meta trains an 8B model to pick tools better than a 402B Llama-4-Maverick
dair_ai · x · 2026-10-23
A Meta/CMU paper tackles tool selection for agents facing corpora of 10,000+ tools, where descriptions overflow the context and retrievers can only rank candidates without knowing which tools an instruction actually needs.
Their recipe trains Qwen3 models on a Plan-Retrieve-Select workflow: the model decomposes instructions into substeps, a retriever fetches candidates per substep, and the model commits one tool each. Training is SFT on 20K curated ToolBench instructions followed by RL with a reward scoring gold-tool coverage and rank.
On ToolBench G3, Qwen3-8B improves from 36.2% after SFT to 41.5% after RL, beating prompted Llama-4-Maverick (402B), and the trained model transfers across retrievers and novel tool corpora.
More from coding & agent
- Gemini 4 Argon builds a 3D Gothic cathedral from one image via Blender MCP — GoogleDeepMind · 2026-10-23
- AI agents take macOS kernel vulns to root: an XNU exploit dev case study — moyix · 2026-10-23
- Wavelength SDK hits mainnet, letting humans and AI agents send bitcoin with a few API calls — RSync25 · 2026-10-23
- Scheduled weekly coding-agent research briefs across Claude Code, Codex, Grok & Gemini — umeshonai · 2026-10-23
- Vercel Agent investigates deployments read-only, then asks for permissions before fixing — kmiddleton14 · 2026-10-23
- Leaving two AI agents to chat overnight produced hilarious mutual profiles — StellaLisy · 2026-10-23