OneRec-V2 Technical Report

TL;DR

This work reengineers end-to-end generative recommender systems by replacing the encoder-heavy architecture with a lazy decoder-only design, enabling scalable deployment up to 8B parameters under fixed compute. It couples post-training preference alignment with real-world user feedback signals through duration-aware reward shaping and Gradient-Bounded Policy Optimization to mitigate reward-hacking and improve sample efficiency. Extensive online A/B tests on the Kuaishou platform demonstrate meaningful gains in engagement metrics and app stay time, validating both architectural efficiency and alignment with user behavior. Overall, the approach advances scalable, end-to-end generative recommendations by aligning model optimization with authentic user feedback signals and robust training dynamics.

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.

Figures (12)

• Figure 1: Left: Scaling curves for various model architectures from 0.1B to 8B parameters, among which Lazy Decoder-only models demonstrate best scaling efficiency. Right: OneRec-V1 v.s. OneRec-V2 at 1B parameters.
• Figure 2: The overall architecture and post-training framework of OneRec-V2. The left panel illustrates the Lazy Decoder-Only Architecture, The right panel depicts the post-training preference alignment process
• Figure 3: Naive Impression Organization: The pattern A$\rightarrow$B is redundantly trained across multiple impressions. User-Centric Organization: When training on User-2's data at time $t_3$, the model has already learned the pattern B$\rightarrow$C from User-1's future interactions at $t_4$. New Impression Only Organization: It trains only on the newest impression.
• Figure 4: Architecture of the proposed lazy decoder-only generative recommender. The Context Processor transforms heterogeneous user feature pathways into unified context representations, which are then normalized to produce layer-shared key-value pairs for cross-attention. The Lazy Decoder processes BOS token and tokenized semantic IDs of the target item through stacked transformer blocks. Each block comprises: (1) lazy cross-attention without key-value projections, enabling Grouped Query Attention (GQA); (2) causal self-attention; and (3) a feed-forward network. The final representations are projected to predict semantic IDs for next-item recommendation.
• Figure 5: Training curves for different architectures across three model scales. Despite achieving similar loss, Lazy Decoder-Only architecture requires 10× fewer FLOPs than classic architectures. E1D1 and E1D2 denote encoder-decoder parameter ratios of 1:1 and 1:2, respectively.