CAST Framework Boosts LLM Tool Use via Case-Based Calibration
A new framework named CAST has been developed by researchers to enhance the utilization of large language models (LLMs) by adjusting reasoning depth and ensuring structural validity based on historical execution paths. Rather than relying on raw examples, CAST analyzes complexity and failure patterns from previous cases to determine the most effective reasoning approaches and predict potential structural failures. This information contributes to a nuanced reward system and adaptive reasoning, enabling the model to adopt case-based strategies during reinforcement learning. Tests conducted on BFCLv2 and ToolBench indicate that CAST improves schema-faithful execution and success in task-level tool use while minimizing unnecessary deliberation, achieving enhancements of up to 5.85%. The findings are published in a paper on arXiv (2605.15041).
Key facts
- CAST is a case-driven framework for LLM tool use.
- It extracts complexity and failure profiles from historical execution trajectories.
- The framework uses fine-grained reward design and adaptive reasoning.
- Experiments were conducted on BFCLv2 and ToolBench.
- CAST improves schema-faithful execution and task-level success.
- It reduces unnecessary deliberation.
- The approach achieves up to 5.85% improvement.
- The paper is available on arXiv (2605.15041).
Entities
Institutions
- arXiv