Researchers from Meta and Carnegie Mellon University trained a Qwen3-8B model with reinforcement learning to select the correct tool 41.5% of the time on the ToolBench G3 benchmark. Llama-4-Maverick, which has roughly 402 billion parameters, achieved a 30.6% success rate under comparable conditions. The 1.7-billion- and 4-billion-parameter Qwen3 variants also surpassed Llama-4-Maverick, scoring 37.6% and 40.3%, respectively. Training combined supervised fine-tuning on approximately 21,000 ToolBench G3 instructions with a reinforcement learning phase using the GRPO method and a plan-retrieve-select strategy. Although the results come from a single benchmark family and the best model still fails more often than it succeeds, they show that a much smaller, less expensive model can outperform a far larger one in tool selection across catalogs exceeding 10,000 tools.
Source: Read the original article
