Diffusion Beats ARM: Diffusion Large Language Models for Generative Recommendation

Hao Jiang, Zixuan Tang, Haonan Wang, Xiaoyu He, Congcong Liu, Ming Pang, Changping Peng, Zhangang Lin, Stavros Petridis, Björn Schuller, Maja Pantic

Keywords: Large Language Models, Generative Recommendation, Discrete Diffusion Models

Abstract: As a promising new paradigm, generative recommender systems frame recommendation as a process of learning data distributions, enabling content creation and diversity exploration by modeling patterns in user behaviors or item characteristics. A common practice to handle large-scale item catalogs is to quantize different item features into discrete semantic sequences, which are then used to train large language models for item generation. However, we argue that this autoregressive generation approach is fundamentally misaligned with the nature of item features in recommendation. Unlike natural language, item attributes are parallel, intertwined, and mutually defining—lacking the hierarchical and sequential dependency that autoregressive models assume. This misalignment limits the effectiveness of existing generative recommendation methods. To address this issue, we propose a new generative recommendation paradigm called \texttt{GREED} (\textbf{G}enerative \textbf{RE}commendation via \textbf{E}lemental \textbf{D}iffusion over Large Language Models). Instead of relying on sequential generation, \texttt{GREED} leverages diffusion-based generative modeling to capture the joint distribution of item features in a non-autoregressive manner. This design better respects the parallel structure of item attributes, thereby improving both efficiency and ranking performance. Extensive experiments demonstrate that \texttt{GREED} outperforms state-of-the-art methods on multiple benchmark datasets. We also conduct detailed offline analyses to validate the efficiency and effectiveness of our approach.

Primary Area: other topics in machine learning (i.e., none of the above)

Submission Number: 7802