POLCA: Stochastic Generative Optimization with LLM

Researchers introduce POLCA (Prioritized Optimization with Local Contextual Aggregation), a new framework that uses large language models as optimizers for complex systems like AI agents and code generation. The method addresses stochastic optimization challenges through priority queuing and meta-learning, demonstrating superior performance across multiple benchmarks including agent optimization and CUDA kernel generation.

• →POLCA formalizes complex system optimization as a stochastic generative problem where LLMs act as optimizers guided by numerical rewards and text feedback.
• →The framework uses priority queues and an ε-Net mechanism to manage exploration-exploitation tradeoffs while maintaining parameter diversity.
• →Theoretical analysis proves POLCA converges to near-optimal solutions under stochastic conditions with noisy feedback.
• →Experimental results show consistent outperformance of state-of-the-art algorithms across diverse benchmarks including agent optimization and code translation.
• →The open-source framework addresses labor-intensive manual iteration traditionally required for optimizing LLM prompts and multi-turn agents.