OneRec-V2 Technical Report

2025-08-28

AI Summary

OneRec-V1 faces two critical bottlenecks in production deployment: severe computational resource imbalance—97.66% of computation is consumed by sequence encoding—and limited generalization due to pure reward-model-driven reinforcement learning (RL). To address these, we reformulate recommendation as an autoregressive generation task and propose Lazy Decoder-Only, a novel architecture featuring lazy decoding that reduces total computation by 94%. We further design duration-aware reward shaping grounded in real user dwell time and incorporate adaptive ratio clipping to improve preference alignment. Our approach unifies generative AI, decoder-only modeling, and online RL. Evaluated on Kuaishou’s production system, it increases average app dwell time by 0.467% and 0.741% across two major traffic segments, achieves more balanced multi-objective recommendations, and scales successfully to an 8B-parameter model—significantly enhancing scalability and online efficacy.

Technology Category

• Search and Optimization: Learning to Search
• Machine Learning: Deep Generative Models & Autoencoders
• Natural Language Processing: Generation

Application Category

• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

Recent breakthroughs in generative AI have transformed recommender systems through end-to-end generation. OneRec reformulates recommendation as an autoregressive generation task, achieving high Model FLOPs Utilization. While OneRec-V1 has shown significant empirical success in real-world deployment, two critical challenges hinder its scalability and performance: (1) inefficient computational allocation where 97.66% of resources are consumed by sequence encoding rather than generation, and (2) limitations in reinforcement learning relying solely on reward models. To address these challenges, we propose OneRec-V2, featuring: (1) Lazy Decoder-Only Architecture: Eliminates encoder bottlenecks, reducing total computation by 94% and training resources by 90%, enabling successful scaling to 8B parameters. (2) Preference Alignment with Real-World User Interactions: Incorporates Duration-Aware Reward Shaping and Adaptive Ratio Clipping to better align with user preferences using real-world feedback. Extensive A/B tests on Kuaishou demonstrate OneRec-V2's effectiveness, improving App Stay Time by 0.467%/0.741% while balancing multi-objective recommendations. This work advances generative recommendation scalability and alignment with real-world feedback, representing a step forward in the development of end-to-end recommender systems.

Problem

• Optimize computational efficiency in generative recommender systems
• Address reinforcement learning limitations with reward models
• Enhance alignment with real-world user interaction preferences

Innovation

• Lazy Decoder-Only Architecture reducing computation
• Duration-Aware Reward Shaping for preference alignment
• Adaptive Ratio Clipping with real-world feedback