LLMs for High-Frequency Decision-Making: Normalized Action Reward-Guided Consistency Policy Optimization

Researchers propose NAR-CP, a new method to improve Large Language Models' performance in high-frequency decision-making tasks like UAV pursuit. The approach uses normalized action rewards and consistency policy optimization to address limitations in current LLM-based agents that struggle with rapid, precise numerical state updates.

• →Current LLMs are limited in high-frequency decision tasks due to frequent numerical state updates with minimal fluctuations.
• →NAR-CP introduces normalized action reward shaping that theoretically preserves optimal policy performance.
• →The method uses consistency loss to align global and sub-semantic policies, reducing misalignment in composite tasks.
• →Experiments on UAV pursuit tasks demonstrate superior performance and generalization to unseen scenarios.
• →The research addresses a key gap in applying LLMs to real-time decision-making applications requiring rapid responses.