RecRec: New RL-free method decouples reasoning, beats SOTA on 4 datasets

Sequential recommender systems typically use a single forward pass to encode user history and predict the next item. A new paper accepted at RecSys 2026 introduces RecRec (Recursive Reasoning for Recommendation), a framework that overcomes the limitations of prior reasoning-enhanced methods. Existing approaches couple reasoning and prediction in a single d-dimensional state, limiting depth and often requiring reinforcement learning. RecRec decouples these processes with two components: a Context Compressor that distills the backbone's hidden states into a compact set of latent interests (enforced by an Interest Diversity Regularizer to capture distinct behavioral aspects), and a Recursive Reasoner that iteratively refines these interests in a separate latent space. The model is trained using simple supervised learning, and deep supervision allows inference-time adjustment of reasoning depth without retraining.

On four real-world datasets, RecRec outperforms state-of-the-art reasoning-enhanced sequential recommenders. Notably, on three of the datasets, performance continues to improve as reasoning depth increases beyond what was seen during training, highlighting the flexibility of the decoupled design. The authors show that structuring reasoning as a multi-vector latent space—rather than a single-state bottleneck—unlocks deeper and more effective inference. For professionals building recommendation systems (e.g., for e-commerce, streaming), RecRec offers a practical, RL-free way to boost accuracy and adapt reasoning compute at runtime, with consistent gains across diverse user behavior datasets.