Source-linked AI summary
OneRec-Think: In-Text Reasoning for Generative Recommendation
Zhanyu Liu, Shiyao Wang, Xingmei Wang, Rongzhou Zhang, Jiaxin Deng, Honghui Bao, Jinghao Zhang, Wuchao Li, Pengfei Zheng, Xiangyu Wu, Yifei Hu, Qigen Hu, Xinchen Luo, Lejian Ren, Zixing Zhang, Qianqian Wang, Kuo Cai, Yunfan Wu, Hongtao Cheng, Zexuan Cheng, Lu Ren, Huanjie Wang, Yi Su, Ruiming Tang, Kun Gai, Guorui Zhou
TL;DR
Existing generative recommenders are effective implicit predictors but lack explicit, controllable reasoning. OneRec-Think integrates itemic alignment, reasoning activation, and recommendation-specific reinforcement learning, achieving state-of-the-art benchmark performance and a 0.159% APP Stay Time gain in industrial deployment.
Problem
Existing generative recommendation models lack explicit, verifiable reasoning pathways, limiting the reasoning capability that distinguishes modern LLMs.
Method
OneRec-Think unifies dialogue, personalized recommendation, and reasoning through itemic alignment, CoT-based reasoning activation, and reinforcement-based reasoning refinement.
Results
OneRec-Think achieves state-of-the-art performance across public benchmarks and a 0.159% gain in APP Stay Time in industrial deployment.
Takeaways & Limitations
The framework provides interpretable reasoning-aware recommendations and demonstrates practical industrial impact through the Think-Ahead architecture.
Takeaways & Limitations
Public datasets have limited behavior-sequence lengths and restricted item spaces, constraining the quality of reasoning learned by the proposed modules.
Abstract
from arXiv · showhide
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.
1 Introduction
Generative recommendation models replace query-candidate matching with autoregressive item generation, but existing systems lack explicit, verifiable reasoning. OneRec-Think addresses this gap by unifying dialogue, reasoning, and personalized recommendation through itemic alignment, reasoning activation, and recommendation-specific enhancement.
- Generative Recommendation: Generative retrieval autoregressively decodes target-item identifiers with Transformer-based sequence-to-sequence models.This paradigm shifts recommendation away from traditional query-candidate matching.
- Generative Recommendation: Unified generative frameworks replace separate retrieval and ranking stages with end-to-end optimization toward the final recommendation objective.The cited frameworks include OneRec, OneLoc, OneSug, and OneSearch.
- Research Gap: Existing generative recommenders lack the explicit, verifiable reasoning pathways associated with text-based Chain-of-Thought.OneRec-Think is proposed to bridge this gap.
- OneRec-Think: OneRec-Think unifies dialogue, reasoning, and personalized recommendation while generating textual reasoning paths and tailoring suggestions to user constraints.Its three stages are Itemic Alignment, Reasoning Activation, and Reasoning Enhancement with a reward capturing multiple valid user choices.
- Contributions: OneRec-Think achieves state-of-the-art results on public benchmarks and delivers a 0.159% gain in APP Stay Time industrially.The deployment-friendly Think-Ahead architecture supports the reported industrial impact.
2 Related Work
Related work develops LLM reasoning through prompting and reinforcement learning, while reasoning-based recommendation remains divided between explicit rationale generation and implicit deduction. Generative recommenders have demonstrated effectiveness but generally lack reasoning capabilities.
- LLM Reasoning: Chain-of-Thought prompting decomposes problems into intermediate reasoning steps and motivates extensions such as zero-shot, self-consistency, and tree-of-thoughts.These prompting approaches support test-time scaling through additional inference computation.
- LLM Reasoning: Recent research shifts from prompting toward posttraining enhancement of reasoning capabilities using techniques such as reinforcement learning.
- Reasoning-Based Recommendation: Existing reasoning-based recommendation methods either generate human-readable rationales for discriminative tasks or perform implicit multi-step deduction.Generative recommenders such as TIGER, HSTU, and OneRec are described as effective but lacking reasoning capabilities.
3 Preliminary
The paper represents items as semantic-rich itemic token sequences and formulates sequential recommendation as next-item prediction from chronological user histories. OneRec-Think extends this formulation by generating a reasoning sequence before the next itemic tokens in one autoregressive pass.
- Itemic Tokens: An itemic token is a discrete, semantic-rich representation unit analogous to a word token, generated from an item’s multimodal and collaborative content.Each item v is mapped to a sequence of such tokens.
- Problem Definition: Users are represented by chronological interaction histories over the item set, with each history containing an ordered sequence of interacted items.
- Problem Definition: The motivating example predicts the next click after a user searches for Arctic, clicks one itemic representation, and collects another.
- Framework Overview: The framework overview comprises item-level semantic alignment, explicit reasoning activation, and reinforcement-learning refinement of reasoning paths.These stages use multi-task pretraining, preference-rationale generation, and a recommendation-tailored reward.
- Unified Formulation: The unified formulation generates a reasoning sequence first and concludes with the next item’s itemic tokens in a single autoregressive pass.The next itemic tokens are sampled conditionally on the prompted history and reasoning sequence.
4 Methodolody
OneRec-Think aligns itemic tokens with language, activates reasoning from user histories, and enhances recommendation through reward-based training. Its Think-Ahead architecture separates offline reasoning from online item finalization to meet real-time latency requirements.
- Itemic Alignment: Multi-task pretraining aligns itemic and text tokens through persona grounding, sequential preference modeling, dense captioning, and general language modeling.A two-substage strategy first warms up itemic embeddings with the base LLM frozen, then jointly optimizes all parameters.
- Reasoning Activation: Reasoning activation distills rationales from pruned contexts and uses them to supervise rationale and target-item generation from noisy behavioral sequences.The objective jointly models rationale tokens and target item tokens, helping bridge user history to target interactions.
- Reasoning Activation: Optimizing LRA teaches the model to extract relevant context from noisy sequences and generate coherent rationales connecting user history with target interactions.The target item is represented by its item tokens alongside the rationale sequence.
- Reasoning Enhancement: Beam Candidate Reward Maximization uses top-K beam candidates to provide denser signals for GRPO under the multi-validity nature of user preferences.Aligning reward computation with beam-search inference establishes training-inference consistency through multi-path evaluation.
- Industrial Deployment: Think-Ahead decouples inference by generating reasoning and initial item tokens offline, then using them as a constrained prefix for real-time final itemization.The design addresses the latency challenge of multi-step reasoning while incorporating current contextual data during online finalization.
5 Experiments
Experiments show that OneRec-Think achieves the best performance across three public recommendation benchmarks, with ablations supporting the contributions of Itemic Alignment and enhanced reasoning. Industrial evaluations and case studies further examine deployment gains, semantic understanding, contextual adaptation, and interpretable reasoning.
- 5.2 Overall Performance: OneRec-Think achieves the best performance across all three public benchmarks, outperforming traditional sequential and generative recommender baselines.The evaluation uses Recall and NDCG at K=5 and 10 on Beauty, Toys, and Sports.
- 5.3 Ablation Study: Itemic Alignment and the enhanced reasoning mechanism each provide additional gains over the base model on Beauty.The ablation compares Base, Base+IA, and Base+IA+R configurations.
- 5.4.2 Results: 0.159% gain in APP Stay Time is reported in a Kuaishou online A/B test, alongside positive trends for Video View and Forward.The experiment used 1.29% of traffic for one week, and repeated experiments consistently improved stay time and related interaction metrics.
- 5.4.2 Results: The model adapts recommendations toward relaxing content after negative emotional signals and user commands in conversational settings.This case study illustrates context-aware recommendation adaptation through interaction with the user.
- 5.4.2 Results: Fine-grained reasoning paths connect user behavior analysis to interpretable recommendations, while itemic and textual tokens support semantic consistency and content anchoring.The case studies describe multi-step preference inference and interleaved reasoning across recommendation scenarios.
6 Conclusion
OneRec-Think unifies generative recommendation with explicit reasoning through itemic token alignment, CoT-based reasoning activation, and reinforcement-based refinement. The paper reports state-of-the-art benchmark performance and industrial gains, while identifying long-sequence modeling and denser reinforcement-learning rewards as future directions.
- 6 Conclusion: OneRec-Think bridges reasoning and generative recommendation through hierarchical itemic token alignment, CoT supervised fine-tuning, and reinforcement-based reasoning refinement.The framework generates interpretable rationales alongside recommendations.
- 6 Conclusion: The model achieves state-of-the-art performance across multiple benchmarks and reports a 0.15% gain in APP Stay Time in industrial deployment.The conclusion presents these results as evidence of both benchmark performance and concrete industrial impact.
- 6 Conclusion: Future work targets user long-sequence modeling and dense reinforcement-learning rewards for finer-grained preference modeling.These directions aim to further connect LLM reasoning with industrial recommendation systems.
Limitations
Public datasets constrain reasoning development because they contain short behavior sequences and restricted item spaces. The paper therefore simplifies its reasoning approach for stable performance on these datasets while developing a larger benchmark.
- Public datasets have limited behavior sequence lengths and restricted item spaces.
- These dataset constraints hinder Reasoning Activation and Reasoning Enhancement from learning industrial-scale reasoning capabilities.
- The authors simplify and adapt the approach to maintain stable, simplified reasoning capacity on public datasets.
Ethics Statement
The experiments use public or platform-collected interaction data under stated privacy and anonymity conditions. Open-source datasets remain publicly available, while industrial data usage follows the platform’s privacy policy.
- Open-source benchmark datasets come from previous works or public APIs and maintain anonymity.
- The industrial experiment uses user interaction data collected from the platform to train the recommendation model.
- All industrial data collection and usage comply with the platform’s privacy policy.
A.1 Experiment settings
The experiment compares OneRec-Think with traditional and generative recommendation baselines using a Qwen3-1.7B backbone. Evaluation uses hierarchical item tokens, beam search, and manually constructed category-based CoT for stable reasoning guidance.
- Baselines: The baselines include traditional models BERT4Rec, HGN, GRU4Rec, and SASRec, alongside generative models TIGER and HSTU.These models represent language-, graph-, sequential-, codebook-, and generative-transduction-based comparison approaches as described.
- Model configuration: The open-source experiments use Qwen3-1.7B as the backbone model.
- Model configuration: The model vocabulary adds 1,024 tokens for four-level hierarchical semantic IDs plus two item boundary tokens.
- Evaluation and reasoning setup: Top-K recommendations are generated with beam search using a beam width of 10.
- Evaluation and reasoning setup: Manually constructed category-based CoT supplies pruned Reasoning Activation content because public benchmark sequences are short and sparse.
A.2 Model Demonstration
The demonstrations show OneRec-Think explaining item tokens and inferring latent user needs from multi-step behavioral patterns. Its examples connect observed interests to recommendations addressing performance optimization, parenting challenges, and item semantics.
- Reasoning examples: Figure 7 presents two generated reasoning paths for short-video recommendation.The cases illustrate multi-step interest inference and identification of underlying user needs.
- Reasoning examples: One case connects gaming preferences and hardware comparisons to performance optimization, recommending monitor-analysis videos.
- Reasoning examples: A second case associates sports and military viewing with adolescent-rebellion searches to identify parenting challenges as the core concern.
- Semantic comprehension: Semantic alignment enables the model to generate coherent item descriptions without explicit training on the caption-generation task.The demonstrations include itemic-token explanations for makeup and skincare products.
A.3 Implementation Details
The implementation uses multi-task data and a two-stage curriculum to align itemic tokens with textual meaning before developing recommendation and reasoning capabilities. It also describes offline reasoning-path generation for deployment-oriented inference.
- Training Data: The training data combines user-persona grounding, sequential preference modeling, itemic dense captioning, and general language modeling.These tasks connect itemic tokens with user attributes, predict next items, generate item descriptions, and preserve general language ability.
- Training Curriculum: The two-stage curriculum first performs Token Warm-up for semantic alignment, then uses Multi-Task Integration to develop recommendation capabilities without losing that alignment.Token Warm-up trains itemic-token understanding, while Multi-Task Integration prevents itemic tokens from becoming non-semantic identifiers during recommendation training.
- Token Warm-up: During Token Warm-up, the LLM backbone is frozen and only newly introduced itemic-token embeddings are trained.The implementation uses a higher learning rate of 5 × 10^-4 for the new embeddings.
- Reasoning Activation: Reasoning Activation progressively moves from pruned, relevant user contexts to reasoning over noisy industrial behavior sequences.Top-k relevant items are selected using semantic similarity, and the semantically aligned model generates step-by-step trajectories explaining target-item alignment.
- System Deployment: Think-Ahead generates reasoning paths and initial item tokens offline, then uses them as constrained prefixes for online finalization.This decoupling is designed to reconcile multi-step reasoning with real-time latency requirements.