OneRec-Think: In-Text Reasoning for Generative Recommendation

TL;DR

OneRec-Think addresses the gap between generative recommendation and explicit reasoning by introducing an integrated framework that grounds item semantics in text, activates step-by-step reasoning within recommendations, and refines reasoning through reinforcement learning with a reasoning-aware reward. The three-stage Itemic Alignment, Reasoning Activation, and Reasoning Enhancement, paired with a Think-Ahead deployment architecture, yield state-of-the-art results on public benchmarks and measurable industrial gains (e.g., a 0.159% increase in APP Stay Time on Kuaishou). The approach produces interpretable rationales alongside high-quality recommendations and demonstrates robust transfer from closed benchmarks to real-world deployment, while outlining limitations of public datasets and future directions toward larger-scale, long-sequence evaluation. Overall, OneRec-Think makes explicit reasoning a core, deployable asset in generative recommendations, enabling more transparent and effective user experiences.

Abstract

The powerful generative capacity of Large Language Models (LLMs) has instigated a paradigm shift in recommendation. However, existing generative models (e.g., OneRec) operate as implicit predictors, critically lacking the capacity for explicit and controllable reasoning-a key advantage of LLMs. To bridge this gap, we propose OneRec-Think, a unified framework that seamlessly integrates dialogue, reasoning, and personalized recommendation. OneRec-Think incorporates: (1) Itemic Alignment: cross-modal Item-Textual Alignment for semantic grounding; (2) Reasoning Activation: Reasoning Scaffolding to activate LLM reasoning within the recommendation context; and (3) Reasoning Enhancement, where we design a recommendation-specific reward function that accounts for the multi-validity nature of user preferences. Experiments across public benchmarks show state-of-the-art performance. Moreover, our proposed "Think-Ahead" architecture enables effective industrial deployment on Kuaishou, achieving a 0.159\% gain in APP Stay Time and validating the practical efficacy of the model's explicit reasoning capability.

Figures (7)

• Figure 1: Examples of OneRec-Think's Unified Dialogue, Reasoning and Recommendation Framework.
• Figure 2: The framework of the OneRec-Think. In the first stage, we achieve item-level semantic alignment through multi-task pre-training. In the second stage, we activate explicit reasoning by prompting the model to generate preference rationales. In the third stage, we refine the reasoning paths through RL based on a reward tailored for recommendations.
• Figure 3: Demonstration of context-aware recommendation adaptation: our model dynamically shifts recommendations to relaxing content based on the user's command.
• Figure 4: Demonstration of fine-grained interest reasoning, which shows the end-to-end process from user behavior analysis to interpretable recommendations.
• Figure 5: The model's reasoning process evolves from broad interest matching (left) to fine-grained theme specification (middle), with recommendations (right) showing semantic consistency with each reasoning step.