Covering the Pareto Frontier with LLM-Coordinated Interpretable Policy Library
Date:
This talk traces 200 years of industrial automation to expose a hidden trilemma: no existing method simultaneously achieves interpretability, automation, and sample efficiency.
We propose a methodological shift from tuning neural network weights to searching over code logic. Our architecture deploys two LLM agents in a closed loop: a Policy Generator that writes human-readable Python/C++ policies from scratch, and a Coordinator that analyzes the Pareto frontier to identify the next unexplored trade-off. The result is a self-designing, interpretable policy library that requires no training and no black-box neural networks.
On multi-objective industrial tasks, the system covers the Pareto frontier in just 30 iterations (<90 minutes), matching the performance of deep RL while producing fully transparent, deployment-ready code.