PaperlayerView exact revision on arXiv

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

arXiv:2510.11639v2cs.IR

TL;DR

Generative recommendation models lack explicit, verifiable reasoning pathways. OneRec-Think unifies dialogue, reasoning, and personalized recommendation, reporting state-of-the-art results across public benchmarks and a 0.159% gain in APP Stay Time in industrial deployment.

  • Problem

    Existing generative recommendation models lack explicit, verifiable reasoning pathways.

    Source: §1 · p. 1

  • Method

    OneRec-Think unifies dialogue and personalized recommendation through itemic alignment, CoT-based reasoning activation, and reinforcement-based reasoning refinement.

    Source: §1 · p. 2, §6 · p. 8

  • Results

    0.159% gain in APP Stay Time accompanied state-of-the-art performance across multiple public benchmarks.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The reported Kuaishou deployment provides evidence of practical efficacy for explicit reasoning in industrial recommendation.

    Source: §1 · p. 2, §5.4.2 · p. 7

    Public datasets’ short behavior sequences and restricted item spaces limit reasoning-module training and evaluation relative to industrial-scale data.

    Source: Limitations · p. 9

Abstract

from arXiv · show

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

OneRec-Think unifies dialogue, explicit reasoning, and personalized generative recommendation in a single framework. It combines item-semantic alignment, reasoning activation, and preference-aware reasoning enhancement.

Introduction

  • Generative retrieval shifts from query-candidate matching to Transformer sequence-to-sequence models that autoregressively decode target-candidate identifiers (Peng et al., 2025; Zhang et al., 2025b; Li et al., 2023; Wang et al., 2023).Source: §1 · p. 1
  • End-to-end generative frameworks replace separate retrieval and ranking stages, enabling holistic optimization toward the final objective and concentrating computation for industrial scaling (Deng et al., 2025; Zhou et al., 2025a,b; Wei et al., 2025; Guo et al., 2025b; Chen et al., 2025).Source: §1 · p. 1
  • Existing generative models lack explicit, verifiable reasoning pathways such as text-based Chain-of-Thought (CoT) (Rajput et al., 2023; Zheng et al., 2024; Wang et al., 2024b).Source: §1 · pp. 1–2
  • OneRec-Think integrates dialogue, reasoning, and personalized generative recommendation through Itemic Alignment, Reasoning Activation, and multi-validity-aware Reasoning Enhancement.Itemic Alignment maps item semantics into the LLM’s textual embedding space; Reasoning Activation induces reasoning in recommendation contexts; Reasoning Enhancement accounts for multiple valid user choices.Source: §1 · p. 2
  • The framework generates interpretable reasoning paths, tailors suggestions to user constraints, and supports efficient, real-time industrial serving with its inference architecture.The paper describes the generated paths as high quality and says they enhance recommendation accuracy and user trustworthiness.Source: §1 · p. 2
  • OneRec-Think reports state-of-the-art results on multiple public benchmarks, while Think-Ahead achieves a 0.159% gain in APP Stay Time.Source: §1 · p. 2

2 Related Work

Prior work develops LLM reasoning through prompting and posttraining, while recommendation approaches separate human-readable rationales from latent reasoning; this work brings explicit reasoning to generative recommendation.

2.1 Reasoning in Large Language Models

2.2 Reasoning-Based Recommendation

  • Explicit reasoning methods produce human-readable rationales but are limited to discriminative tasks, whereas implicit methods perform latent reasoning without textual interpretability.Generative recommendation models including TIGER, HSTU, and OneRec are described as lacking reasoning capabilities; this work introduces explicit reasoning to generative recommendation.Source: §2.2 · p. 2

3 Preliminary

The preliminaries define items as sequences of semantic itemic tokens and frame recommendation as generating reasoning followed by the next item’s tokens in one autoregressive pass.

Itemic Token

  • Each item v is represented by semantic-rich itemic tokens generated from its multimodal and collaborative content.The representation is analogous to a word token in natural language.Source: §3 · p. 2

Problem Definition

  • Each user u has a chronological interaction history V_u of length n_u over the item set V.Source: §3 · p. 2
  • Conditioned on a prompted user history, one autoregressive pass generates a reasoning sequence τ before the next item’s itemic tokens.The prompt is defined as a valid prompt constructed for recommendation.Source: §3 · p. 3

4 Methodolody

OneRec-Think aligns item tokens with language, activates reasoning through rationale supervision, enhances it with beam-based rewards, and uses Think-Ahead inference for industrial deployment.

4.1 Itemic Alignment through Multi-Task Pre-training

  • Multi-task pretraining aligns itemic tokens with language through interleaved persona grounding, sequential preference prediction, itemic captioning, and general language modeling.Persona data interleaves item and text tokens across static attributes, search behaviors, interaction sequences, and summarized interests.Source: §4.1 · pp. 3–4
  • Training proceeds from a frozen-LLM token warm-up on persona grounding to joint optimization of all parameters across the combined tasks.The multi-task integration stage uses a designed task ratio.Source: §4.1 · p. 4

4.2 Reasoning Activation

  • Reasoning Activation retrieves relevant items from pruned histories to elicit target-aligned rationales, then trains on raw sequences to generate both rationales and target items.This distillation teaches the model to extract relevant context from noisy sequences and connect user history to target interactions.Source: §4.2 · p. 4

4.3 Reasoning Enhancement

  • Beam-based rollout rewards address sparse recommendation rewards by scoring the best candidate in a constrained beam and support GRPO training aligned with beam-search inference.The reward uses multi-path evaluation to provide denser signals that account for the multi-validity of user preferences (Shao et al., 2024).Source: §4.3 · pp. 4–5

4.4 Industrial Deployment: A "Think-Ahead" Architecture

  • Think-Ahead generates reasoning and initial item tokens offline, then uses a real-time updated OneRec model to finalize the item from those tokens as a constrained prefix.The initial tokens capture broad intent or preference context, while online finalization leverages current contextual data.Source: §4.4 · p. 5

5 Experiments

The experiments report that OneRec-Think achieves the best performance across benchmarks and show gains in online APP Stay Time. Ablations and case studies examine its components and reasoning behavior.

5.1 Experimental Settings

  • Experiments compare OneRec-Think with classic sequential and generative recommenders on Amazon Beauty, Toys, and Sports using R@5/10 and N@5/10.The baselines include BERT4Rec, HGN, GRU4Rec, SASRec, TIGER, HSTU, and ReaRec.Source: §5.1 · p. 5

5.2 Overall Performance

  • OneRec-Think achieves the best performance across all benchmarks, while reasoning-based ReaRec and OneRec-Think outperform traditional sequential and generative recommenders.The overall comparison covers Beauty, Toys, and Sports.Source: §5.2 · p. 5, §5.1 · p. 5, Table 1 · p. 6

5.3 Ablation Study

  • On Beauty, Itemic Alignment boosts the base model, and adding enhanced reasoning yields a further significant gain.The study compares the Base model, Base+IA, and Base+IA+R; the passages attribute the gains to coherent itemic-token semantics and the reasoning mechanism.Source: §5.3 · p. 5, Table 2 · p. 6

5.4.1 Training Settings

  • Industrial training uses pretrained Qwen-8B with 24,576 hierarchical itemic tokens and two boundary tokens; production training updates daily on 80 GPUs, processing about 20B tokens per day.The vocabulary includes 8,192 itemic tokens per level across three levels.Source: §5.4.1 · p. 6

5.4.2 Results

  • 0.159% gain in APP Stay Time was reported in a one-week Kuaishou A/B test using 1.29% of traffic; Video View and Forward also showed positive trends.Multiple experiments at different times consistently improved stay time and related interaction metrics.Source: §5.4.2 · pp. 6–7, Table 3 · p. 7
  • On industrial benchmarks, Token Warm-up provides limited gain for text-heavy User Understanding, while Multi-Task Integration boosts it; both stages progressively improve Short Video Understanding.The evaluation uses BertScore and attributes the Short Video Understanding gains to interpreting non-textual item information.Source: Table 4 · p. 7, §5.4.2 · pp. 7–8

5.5 Case Study

  • Case studies show affect-aware shifts to relaxing content, fine-grained preference reasoning, semantic consistency between reasoning and recommendations, and interleaved itemic-textual reasoning.The examples include reasoning about gameplay mechanics and narrative patterns, with itemic tokens anchoring content and textual tokens articulating causal links.Source: Figure 3 · p. 6, Figure 4 · p. 7, §5.5 · p. 8

6 Conclusion

OneRec-Think integrates itemic alignment, reasoning activation, and reinforcement-based refinement to bring explicit reasoning to generative recommendation. The conclusion reports benchmark and industrial gains and identifies long-sequence modeling and denser RL rewards as future directions.

6 Conclusion

  • OneRec-Think achieves state-of-the-art performance across multiple benchmarks and a 0.15% gain in primary industrial metrics like APP Stay Time.The conclusion describes these results as evidence of concrete industrial impact.Source: §6 · p. 8
  • OneRec-Think bridges reasoning and generative recommendation through hierarchical itemic token alignment, CoT supervised fine-tuning, and reinforcement-based reasoning refinement.It transforms recommendation from item prediction into reasoning-aware output that pairs recommendations with interpretable rationales.Source: §6 · p. 8
  • Future work will explore user long-sequence modeling and dense RL rewards for finer-grained preference modeling and stronger links between LLM reasoning and industrial recommendation.Source: §6 · p. 9

Limitations

Public datasets’ short behavior sequences and restricted item spaces constrain reasoning quality; the approach is simplified for robust performance on these datasets, and a larger benchmark is being constructed.

Limitations

  • Short behavior sequences and restricted item spaces in public datasets limit reasoning quality, prompting a simplified robust adaptation and construction of a larger benchmark with longer trajectories and diverse catalogs.The planned benchmark is intended to support more comprehensive evaluation of reasoning-based recommendation models.Source: Limitations · p. 9

Ethics Statement

The study uses publicly available benchmark datasets and consented platform interaction data; industrial training uses aggregated information and does not access or process personally identifiable information.

Ethics Statement

  • The open-source benchmark datasets are publicly available from prior work or public APIs while maintaining anonymity.The industrial scenario uses platform interaction data under the platform’s privacy policy and terms of service, with users’ explicit consent.Source: Ethics Statement · p. 9
  • Industrial training uses aggregated behavioral sequences, textual content, and user base information without processing personally identifiable information.Data collection and usage comply with the platform’s privacy policy and terms of service, to which users have provided explicit consent.Source: Ethics Statement · p. 9

A Appendix · A.1 Experiment settings

Appendix A.1 specifies the experimental settings: competitive traditional and generative recommender baselines, ranking metrics, model configuration, and evaluation and training procedures.

A.1 Experiment settings

  • BERT4Rec (Sun et al., 2019), HGN (Ma et al., 2019), and GRU4Rec (Hidasi et al., 2016) are included as traditional recommender baselines.They respectively use pretrained language representations, graph neural networks, and graph convolution to model user-item relationships or interactions.Source: §A.1 · p. 11
  • SASRec, TIGER (Rajput et al., 2023), HSTU (Zhai et al., 2024), and ReaRec (Tang et al., 2025) are included as additional baselines.They represent self-attention, codebook-based generative recommendation, sequential transduction, and implicit multi-step reasoning approaches, respectively.Source: §A.1 · p. 12
  • Evaluation uses top-K Recall (R@K) and NDCG (N@K) at K = 5 and 10, following (Rajput et al., 2023).Source: §A.1 · p. 12
  • Open-source dataset experiments use Qwen3-1.7B (Yang et al., 2025) with 1,024 hierarchical semantic-ID tokens and two item-boundary tokens.The 1,024 added tokens represent four levels with 256 tokens per level; the boundary tokens are <|item_begin|> and <|item_end|>.Source: §A.1 · p. 12
  • Evaluation generates top-K recommendations using beam search with beam width 10, while Reasoning Activation uses manually constructed category-based CoT for stable semantic guidance.The manual CoT addresses the challenge of deriving robust reasoning paths from short, sparse public-benchmark item sequences.Source: §A.1 · p. 12
  • Training removes users and items with fewer than 5 interactions, uses sequential leave-one-out splits, and caps user histories at 50 items.Preprocessing follows (Rajput et al., 2023; Wang et al., 2024a), dataset splitting follows (Rajput et al., 2023; Zheng et al., 2024), and the history cap follows.Source: §A.1 · p. 12

A.2 Model Demonstration

The demonstrations show OneRec-Think inferring latent user needs for short-video recommendations and interpreting itemic tokens through natural-language semantics. Together, they illustrate reasoning beyond topic matching and semantic alignment that supports token-to-text comprehension.

Reasoning Cases for Short-video Recommendation

  • One path infers a need for performance optimization from gaming and hardware-comparison interests; the other identifies parenting challenges from sports, military, and rebellion-related viewing and searches.Source: §A.2 · p. 12

Semantic Comprehension Validation

  • Without explicit training for captioning, the aligned model generates coherent descriptions of itemic tokens, indicating it can translate between discrete tokens and natural-language semantics.Source: §A.2 · p. 12

Reasoning Cases for Short-video Recommendation

  • Figure 7 presents two reasoning paths that demonstrate multi-step interest inference and identification of underlying user needs.Source: §A.2 · p. 12, Figure 7 · p. 13

A.3 Implementation Details … A.3.4 System Deployment: A "Think-Ahead" Architecture

The implementation grounds itemic tokens in text, develops recommendation reasoning through staged training, and uses Think-Ahead to cache reasoning-derived prefixes offline for latency-sensitive online recommendation.

A.3.1 Itemic Alignment

  • Itemic Alignment trains persona grounding, next-item prediction, itemic-to-text description, and general language modeling tasks.Sequential prediction uses histories of up to 256 items; its loss is applied only to target itemic tokens, while captioning loss applies to target descriptions.Source: §A.3.1 · pp. 13–14
  • Token Warm-up aligns itemic tokens to textual meanings before Multi-Task Integration builds recommendation capabilities while reinforcing semantic understanding.In industry, warm-up uses 6 billion tokens of persona-grounding data and freezes the LLM backbone while training new item-token embeddings.Source: §A.3.1 · pp. 14–15

A.3.2 Reasoning Activation

  • Reasoning Activation first generates rationales from the top 10 semantically similar historical items, then trains on complete histories to jointly predict rationale and target-item tokens.The joint objective weights rationale generation and recommendation prediction equally, transferring reasoning from pruned contexts to noisy sequences.Source: §A.3.2 · pp. 15–16

A.3.3 Reasoning Enhancement

  • Reinforcement Learning with VERL’s GRPO and Rollout-Beam reward further refines reasoning coherence and recommendation accuracy.Training samples 16 CoT paths, uses beam width 32, and runs for 2 epochs with learning rate 1e-5 and KL coefficient β = 0.001.Source: §A.3.3 · p. 17

A.3.4 System Deployment: A "Think-Ahead" Architecture

  • Offline, Think-Ahead samples diverse reasoning paths, generates candidate item prefixes, and caches their union as a personalized candidate space.In the industrial setting, it uses L = 3 hierarchical item identifiers and caches T × m high-potential prefixes in distributed storage.Source: §A.3.4 · p. 17
  • Online, Think-Ahead retrieves the cached candidates and constrains real-time final-token decoding to their prefixes before serving top-K items.The online stage uses a real-time updated OneRec for finalization, reconciling reasoning with strict latency requirements.Source: §A.3.4 · p. 17

Related papers

Loading 2510.11639v2…