This research develops a framework to intelligently determine when large language models should execute code versus use textual reasoning to solve problems.

• Code execution achieves 100% success on certain tasks while avoiding the computational overhead of textual reasoning
• Textual reasoning struggles with complex math, logic, and search problems that code handles efficiently
• The authors propose methods to effectively steer LLM behavior between these two approaches
• This engineering breakthrough improves efficiency and effectiveness in AI problem-solving systems

For engineering teams, this research offers practical techniques to optimize LLM implementations by selecting the most efficient problem-solving method based on the task type.

Steering Large Language Models between Code Execution and Textual Reasoning