Session 4 // Dec 22: Diffusion beats AR: Code Generation
Diffusion LLM
Hello folks,
Discrete diffusion models now rival autoregressive (AR) models on challenging coding benchmarks, making them a compelling alternative to AR models.
This Monday, Shansan Gong will present recipes for training masked diffusion models to reach such coding performance, and will reveal several surprising inference-time behaviors of these models.
Title: DiffuCoder: Understanding and Improving Masked Diffusion Models for Code Generation
Meeting Link: click here
Time: Dec 22 (Monday) 10 am ET / 4pm CET
Paper: https://arxiv.org/abs/2506.20639
Prior knowledge:
Fundamentals of discrete diffusion (video by Sasha Rush)
How LLMs learn to reason [GRPO] (video by Jia-Bin Huang)
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, 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 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.
Yours truly,
Subham, Justin, Zhihan
Diffusion LLM
Gentle reminder: See you all at 10 AM ET / 4 PM CET.
Meeting Link: click here
Today's paper: https://arxiv.org/abs/2506.20639