Appier's SMITH Trains AI Agents to Create and Use Tools

Appier introduced SMITH, a reinforcement learning framework that trains AI agents to create and use tools in one loop, according to a company press release. The research paper was accepted at NeurIPS.

SMITH gives tool performance results back to the model during training, allowing it to refine descriptions, parameters and code. The system retains tools that solve previously unseen problems and adds them to a shared library for use across models and tasks.

In tests, a model with about 4 billion parameters created tools that outperformed other methods on unseen tasks, including a baseline using a model with about 30 billion parameters. The tools also worked with a model of about 350 million parameters. Average output fell from 3,206 tokens with step by step reasoning to about 100 tokens while maintaining task performance.