News
RL-trained small model beats Meta's larger Llama-4-Maverick in tool selection benchmark
Researchers from Meta and Carnegie Mellon report that a compact Qwen3 model, post‑trained with reinforcement learning, selected the correct software tool 41.5% of the time on the ToolBench G3 benchmark, while Meta's Llama-4-Maverick achieved 30.6% under similar conditions. The findings were published in an arXiv paper submitted on October 22, 2026. The task required choosing the appropriate tool from a corpus of over 10,000 options using a Plan‑Retrieve‑Select three‑step approach. Models underwent supervised fine‑tuning followed by reinforcement learning with the GRPO method, training on roughly 21,000 instructions. Smaller model variants also outperformed the larger model, reaching success rates in the high‑30s to low‑40s percent.