Latent Reasoning with Recurrent Depth for Sequential Recommendation

Enze Liu, Bowen Zheng, Xiaolei Wang, Xin Zhao, Jinpeng Wang, Sheng Chen, Ji-Rong Wen

Keywords: sequential recommendation, latent reasoning

Abstract: Sequential recommender systems play a pivotal role in modern applications by modeling user behavior sequences to predict their preferences. However, current approaches primarily adopt non-reasoning paradigms, which constrain their computational capacity and lead to suboptimal performance. To overcome these limitations, we propose LARES, an innovative and scalable \textbf{LA}tent \textbf{R}easoning framework for \textbf{S}equential Recommendation that unlocks deep thinking with a recurrent depth. Unlike conventional parameter scaling methods, LARES enhances the model's representational power by increasing the computational density of parameters through depth-recurrent latent reasoning. Its recurrent architecture allows flexible expansion of reasoning depth without extra parameters, thereby effectively capturing complex and evolving user interest patterns. To fully exploit the model's reasoning potential, we introduce a two-stage training strategy: (1) Self-supervised pre-training (SPT) with \textit{trajectory-level alignment} and \textit{step-level alignment}, where the model learns latent reasoning patterns tailored for sequential recommendation tasks without annotated data, and (2) Reinforcement post-training (RPT), which leverages reinforcement learning (RL) to encourage exploration of diverse reasoning paths and further refine its reasoning capabilities. Extensive experiments on real-world benchmarks demonstrate LARES's superiority. Notably, the framework exhibits seamless compatibility with existing advanced models, consistently improving their recommendation performance. Our code is available at https://anonymous.4open.science/r/LARES-E458/.

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

Submission Number: 11629