Agent-GWO Framework Introduces Collaborative Agents for Dynamic Prompt Optimization in LLMs

A new framework called Agent-GWO addresses limitations in automatic prompt optimization for large language models (LLMs). While LLMs show strong reasoning capabilities and strategies like Chain-of-Thought (CoT) improve performance, high-quality reasoning still depends heavily on manual static prompts. This reliance makes performance sensitive to decoding configurations and task distributions, causing fluctuations and limited transferability. Existing methods typically use single-agent local search, failing to optimize both prompts and decoding hyperparameters together for stable global improvements. Agent-GWO unifies prompt templates and decoding hyperparameters as inheritable agent configurations. It employs a leader-follower mechanism to enable collaborative agents for dynamic prompt optimization in complex reasoning tasks. The framework aims to achieve more stable and transferable performance by simultaneously optimizing these elements within a unified approach. The research is documented in arXiv preprint 2604.18612v1, which announces the cross-disciplinary abstract.