Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Aug 26·5 min read
Joint Optimization of Tool Creation and Use for Large Language Model Agents
TL;DR
SMITH, a reinforcement learning framework, enables simultaneous training of tool creation and use for large language models, achieving superior performance on procedural reasoning tasks. A 4B Qwen3 model trained with SMITH outperformed existing methods, achieving 79.8% accuracy on held-out tasks.
✦ Why It Matters
Consider implementing SMITH to enhance your language model's tool creation and usage capabilities today.
Key Takeaways
- Explore SMITH for joint training of tool creation and usage.
- Utilize the Qwen3 model for improved procedural reasoning tasks.
- Assess schema, code, and outcome failures independently.
- Leverage tools created by models to boost performance in other language models.
How It Works
SMITH integrates tool creation and usage into a single reinforcement learning policy. It alternates between tasks that involve writing tools based on examples and invoking those tools on new questions.
By using three distinct reward axes, it can effectively learn from different types of failures, leading to more robust tool performance.