MURPHY: Multi-Turn GRPO for Self Correcting Code Generation

Murphy represents a notable advancement in the field of artificial intelligence, particularly in enhancing the reasoning capabilities of large language models through its multi-turn reflective optimization framework. By building on the existing Group Relative Policy Optimization (GRPO) approach, Murphy incorporates iterative self-correction, allowing models to refine their reasoning progressively. This is crucial for tasks that require complex decision-making, which GRPO has struggled with. Evaluations conducted on code generation benchmarks using model families such as Qwen and OLMo demonstrate that Murphy consistently outperforms GRPO, achieving up to an 8% relative gain in pass@1 metrics. This improvement not only showcases the effectiveness of Murphy but also highlights the potential for further advancements in reinforcement learning frameworks, particularly those utilizing verifiable rewards to enhance model performance.