Meta and Carnegie Mellon study finds an RL-trained 8B model beats Llama-4-Maverick at tool selection

A language model with 8 billion parameters just outscored one roughly fifty times its size. The task was picking the right software tool for a job.

Researchers from Meta and Carnegie Mellon University report that a small Qwen3-8B model, post-trained with reinforcement learning, selected the correct tools 41.5% of the time on the ToolBench G3 benchmark. Meta’s own Llama-4-Maverick, a model of approximately 402 billion parameters, scored 30.6% under similar conditions.

What the researchers actually built

The work appears in an arXiv paper titled “Training Small Language Models for Transferable Tool Selection,” listed as arXiv:2610.31483. It was submitted on October 22, 2026.

The problem it tackles is a practical one for AI agents. Agents are models that take actions, such as calling an API or querying a database, rather than just producing text. To do that, they have to choose the right tool from whatever is available.

The study focuses on settings where the tool corpus exceeds 10,000 tools. Instead of using proprietary systems like Claude or ChatGPT, the team worked with compact Qwen3 models ranging from 1.7 billion to 8 billion parameters. They paired those models with a three-step approach they call Plan-Retrieve-Select.

The model first works out what it needs to do, then pulls a shortlist of candidate tools, then picks the one that fits.

How the training worked

The models went through two rounds of post-training. First came supervised fine-tuning, or SFT, where a model learns by imitating labeled examples of correct behavior.

Then came reinforcement learning. For the RL phase, the researchers used a method called GRPO to structure the rewards. Training drew on approximately 21,000 ToolBench G3 instructions.

The results held up across model sizes. The 1.7B Qwen3 variant reached a 37.6% success rate, and the 4B variant hit 40.3%. Both cleared the 30.6% mark set by Llama-4-Maverick.

Here is a quick rundown of the reported scores on ToolBench G3:

Qwen3-8B (RL-trained): 41.5%
Qwen3-4B (RL-trained): 40.3%
Qwen3-1.7B (RL-trained): 37.6%
Llama-4-Maverick (~402B): 30.6%

Why Llama-4-Maverick is the benchmark that stings

Llama-4-Maverick is Meta’s roughly 400B-parameter Mixture of Experts model, released in 2025. A Mixture of Experts design splits the model into specialized sub-networks and activates only some of them for each request.

What this means for builders and the AI market

The most immediate implication is cost. Running a 402B-parameter model for every tool call is expensive, even with a Mixture of Experts architecture trimming the active load. An 8B model is far cheaper to host and far faster to query.

A few caveats are worth keeping in mind. The results come from a single benchmark, ToolBench G3, and the training data was drawn from that same benchmark family. Absolute numbers also deserve perspective. A 41.5% success rate is a clear win over 30.6%, but it still means the best model missed more often than it hit. Tool selection at the scale of 10,000-plus tools remains far from solved.