Meta, Carnegie Mellon study shows RL-trained models outperform Llama-4-Maverick in tool selection

Meta: Meta conducts extensive AI research focused on advancing large language models and agentic systems. Researchers from the company co-authored a recent paper on training smaller models with reinforcement learning for improved tool selection in agents handling large tool catalogs. The work positions Meta's Llama-4-Maverick as the baseline outperformed by the RL-trained approach. Qwen3: Qwen3 is Alibaba's family of open-weight language models released and updated throughout 2026. The Meta-Carnegie Mellon paper fine-tunes Qwen3 variants with supervised learning followed by reinforcement learning on a plan-retrieve-select workflow. The resulting models demonstrate strong generalization across tool benchmarks without further training. Maverick: Llama-4-Maverick is Meta's mixture-of-experts model designed for high-capability multimodal and reasoning tasks. The recent research paper evaluates prompted versions of Maverick on tool selection and finds it surpassed by smaller RL-trained Qwen3 models on ToolBench and transfer benchmarks. Carnegie Mellon: Carnegie Mellon University participates in collaborative AI research projects with industry partners. Faculty and affiliates contributed to the paper on transferable tool selection using RL-trained small models. The university's involvement underscores academic efforts to enhance efficiency in LLM agent workflows. Agent Tool Selection: Recent research shows reinforcement learning can train compact open models to decompose instructions and select tools from large catalogs more effectively than prompting frontier-scale systems. Efficiency in AI Agents: Meta's latest work emphasizes training compact models as alternatives to proprietary large language models for handling complex multi-step tool use in agents. Open Model Post-Training: Developments in 2026 highlight how supervised warm-up combined with RL enables smaller models to acquire specialized agent skills that transfer zero-shot to new domains and retrievers.