Meta and Carnegie Mellon study finds an RL-trained 8B model beats Llama-4-Maverick at tool selection
Meta and Carnegie Mellon researchers show an RL-trained Qwen3-8B hit 41.5% on ToolBench G3, beating the ~402B Llama-4-Maverick's 30.6%. Researchers from Meta and Carnegie Mellon University have found that a Qwen3-8B model, post-trained with reinforcement learning, can outperform Llama-4-Maverick at tool selection, winning 41.5% on ToolBench G3 compared to 30.6% on similar conditions. The study was published in an arXiv paper titled “Training Small Language Models for Transferable Tool Selection. The researchers used compact Qwen 3 models with 1.7 billion to 8 billion parameters and Plan-Retrieve-Select, a three-step approach they call Plan-retrieve- Select. The results showed a 41.05% success rate, but still showed that tool selection at the scale of 10,000-plus tools remains far from solved. The cost of running a 402B-parameter model for every tool call could be high, with an 8B model far cheaper to host and faster to query.
公開済み : 16時間前 沿って Diego Almada Lopez の Tech
A Qwen3-8B model post-trained with reinforcement learning hit 41.5% on ToolBench G3, ahead of the roughly 402B-parameter Llama-4-Maverick at 30.6%
A language model with 8 billion parameters just outs... [3060 symbols]
トピック: Academia, Data, Cryptocurrency
