Appier has developed a reinforcement learning framework that allows AI agents to create, use, and refine their own tools, improving the efficiency and scalability of agentic AI systems.
The research paper, “Joint Optimisation of Tool Creation and Use for Large Language Model Agents,” has been accepted at NeurIPS 2026, a leading conference in artificial intelligence and machine learning. The paper introduces SMITH (Schema-grounded Multi-task Iterative Tool Honing), a method that trains AI models to develop and use tools within a unified learning process.
Dr Chih-Han Yu, CEO and co-founder of Appier, stated, “This research shows that agents can learn to build tools, continuously refine them, and share proven tools across models of all sizes, making multi-agent collaboration more efficient and scalable.”
Schema-grounded Multi-task Iterative Tool Honing
Traditional AI systems often require engineers to build and configure tools in advance. Although some models can generate tools, tool creation and use are usually separate processes, which limits feedback for improvement. SMITH evaluates whether tools are clearly described, used correctly, and applied effectively to new problems, using these results for continuous refinement.
The research highlights three key findings:
- Small models can outperform larger ones at tool creation.
- Proven tools work across model sizes.
- Repeated reasoning becomes a reusable tool.
“Looking ahead, we hope to build models that can continuously interact with their environment and take on a wider range of tasks,” said Chieh-Yen Lin, research scientist at Appier.
