DiffGRM: Diffusion-Based Generative Recommendation Model

A new research paper called 'DiffGRM: Diffusion-based Generative Recommendation Model' just dropped on arXiv, with the identifier 2510.21805v2. It focuses on the limitations of autoregressive models (ARMs) in generative recommendation (GR). The authors point out two main issues with semantic IDs (SIDs): one is intra-item consistency, which is about how the digits represent a single item but are limited by a left-to-right approach. The second issue is inter-digit heterogeneity, where differences in semantic detail create training imbalances. To solve these problems, they propose DiffGRM, which uses a masked discrete diffusion model (MDM) instead of an autoregressive decoder, allowing for better context and digit relationships. You can check it out at https://arxiv.org/abs/2510.21805.