Meta & CMU: RL-trained 8B model beats 402B Llama-4-Maverick at tool selection

rohanpaul_ai · x · 2026-10-24

A Meta and Carnegie Mellon paper (arXiv:2610.31483) shows a small RL-trained model can outpick a prompted giant at tool selection.

Problem: Agents face 10,000+ tools whose descriptions don't fit in context; relevance-based retrievers only rank, so recent setups let Claude or ChatGPT decompose instructions and pick from candidates.

Method: Qwen3 models are trained to split requests into steps, keyword-search each step, and pick one tool per step — first on teacher-written examples, then with RL rewarding useful searches and exact picks.

Results:

On ToolBench, the trained Qwen3-8B picks exactly the right tools 41.5% of the time vs 30.6% for prompted 402B Llama-4-Maverick
Without retraining, it beats Maverick on 3 of 4 new tool benchmarks and improves with a stronger retriever
It transfers directly across retrievers and new domains

Takeaway: for large tool catalogs, an RL-trained small model can replace a prompted large one for tool selection.

Original post →

Meta & CMU: RL-trained 8B model beats 402B Llama-4-Maverick at tool selection

More from coding & agent