DiffuCoder: Understanding and Improving Masked Diffusion Models for Code Generation

TL;DR

DiffuCoder is a 7B diffusion-based language model for code generation that enhances decoding diversity, reduces autoregressive bias, and improves performance through a novel RL training scheme called coupled-GRPO.

Contribution

This work systematically analyzes the decoding behavior of diffusion LLMs for coding and introduces coupled-GRPO, a new RL training method that improves performance and reduces AR bias.

Findings

DiffuCoder outperforms baseline models on code benchmarks (+4.4%).

DiffuCoder exhibits unique decoding behaviors, such as adjustable causality and diverse generation order.

Coupled-GRPO reduces variance in training and enhances model performance.

Abstract

Diffusion large language models (dLLMs) are compelling alternatives to autoregressive (AR) models because their denoising models operate over the entire sequence. The global planning and iterative refinement features of dLLMs are particularly useful for code generation. However, current training and inference mechanisms for dLLMs in coding are still under-explored. To demystify the decoding behavior of dLLMs and unlock their potential for coding, we systematically investigate their denoising processes and reinforcement learning (RL) methods. We train a 7B dLLM, \textbf{DiffuCoder}, on 130B tokens of code. Using this model as a testbed, we analyze its decoding behavior, revealing how it differs from that of AR models: (1) dLLMs can decide how causal their generation should be without relying on semi-AR decoding, and (2) increasing the sampling temperature diversifies not only token choices but also their generation order. This diversity creates a rich search space for RL rollouts. For RL training, to reduce the variance of token log-likelihood estimates and maintain training efficiency, we propose \textbf{coupled-GRPO}, a novel sampling scheme that constructs complementary mask noise for completions used in training. In our experiments, coupled-GRPO significantly improves DiffuCoder's performance on code generation benchmarks (+4.4\% on EvalPlus) and reduces reliance on AR bias during decoding. Our work provides deeper insight into the machinery of dLLM generation and offers an effective, diffusion-native RL training framework. https://github.com/apple/ml-diffucoder.