Reinforcement Learning for Assembly Code Optimization with LLMs

The provided source explores enhancing assembly code performance using large language models (LLMs) through reinforcement learning (RL). It introduces a novel RL framework that trains LLMs with Proximal Policy Optimization (PPO), guided by a reward function that balances functional correctness and execution speedupcompared to the industry-standard gcc -O3 compiler. To facilitate this research, a benchmark of 8,072 real-world programs was developed. The resulting model, Qwen2.5-Coder-7B-PPO, significantly outperforms 20 other models, achieving a 96.0% test pass rate and an average 1.47x speedup, demonstrating LLMs' potential as effective assembly code optimizers.