Full-Sequence Masked Diffusion for Generative Recommendation

Autoregressive large language models have dominated generative recommendation by sequentially predicting the next item. However, this unidirectional paradigm suffers from error propagation, and limited ability to jointly optimize an entire recommendation list. In this paper, we introduce Full-Sequence Masked Diffusion for generative recommendation (FSMD), which extends masked diffusion models to enable global joint modeling over the entire user history and Top-$K$ recommendation list. We formulate recommendation as a unified sequence diffusion process, where the forward process randomly masks tokens across the full sequence and the reverse process simultaneously denoises all masked positions using bidirectional context. This formulation enables parallel generation of the entire recommendation list and naturally captures inter-item dependencies within a coherent global structure. To better align generation with ranking objectives, we further introduce a unified training objective that combines masked diffusion likelihood with a listwise ranking loss. Extensive experiments on multiple public benchmarks demonstrate that FSMD consistently outperforms state-of-the-art autoregressive and item-level diffusion-based generative recommendation methods across accuracy and list quality metrics.