Learning Large Language Models for Code Generation through Genetic Algorithms and Knowledge Distillation

Abstract

Drawing inspiration from open-ended evolution, this paper explores the concept of individual Large Language Models (LLMs) functioning as autonomous agents while advancing learning as a group, aiming to solve complex problems that are challenging for a single model. As a specific method, we propose a learning process that combines genetic algorithms with knowledge distillation. By progressing learning through knowledge distillation and simultaneously optimizing hyperparameters with genetic algorithms, we aim for more efficient learning. For the domain task, we selected the code generation task of producing Python code from instructions. In our experiments, we utilized three student models and one teacher model for learning. The results showed a 1.2% improvement in accuracy on HumanEval’s pass@1, indicating signs of optimized learning rates as learning progressed. However, challenges remain in achieving significant accuracy improvements and optimizing a variety of hyperparameters.