# OneRec: Unifying Retrieve and Rank with Generative Recommender and Iterative Preference Alignment

**Authors:** Jiaxin Deng, Shiyao Wang, Kuo Cai, Lejian Ren, Qigen Hu, Weifeng Ding, Qiang Luo, Guorui Zhou  
**arXiv:** [2502.18965](https://arxiv.org/abs/2502.18965)  
**Date:** 2025-02-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2502.18965v1/

## Key points

- OneRec is a unified end-to-end generative framework for single-stage recommendation
- Incorporates encoder-decoder architecture and scales up model parameters based on sparse Mixture-of-Experts (MoE) structure
- Adopts session-wise list generation approach considering content and order of items within each session
- Explores preference learning through direct preference optimization (DPO) using self-hard rejected samples from beam search results
- Implements Iterative Preference Alignment (IPA) strategy to rank sampled responses based on pre-trained reward model scores
- Extensive experiments demonstrate superiority in generating high-quality recommendations
- Successfully deployed in Kuaishou, resulting in a significant 1.6% increase in watch-time

## Summary

, , , , 

This paper introduces OneRec, a unified end-to-end generative framework for single-stage recommendation, designed to address the challenges faced by traditional recommendation systems. Inspired by the scaling laws observed in training large language models, OneRec incorporates an encoder-decoder architecture and scales up its model parameters based on a sparse Mixture-of-Experts (MoE) structure. This allows OneRec to effectively capture user interests. Unlike conventional point-by-point prediction methods, OneRec adopts a session-wise list generation approach that considers both the relative content and order of items within each session. This eliminates the need for hand-crafted strategies and allows the model to autonomously learn optimal session structures from input data. To further improve the quality of generated recommendations, OneRec explores preference learning through direct preference optimization (DPO). To construct preference pairs, self-hard rejected samples are created from beam search results instead of random sampling. An Iterative Preference Alignment (IPA) strategy ranks sampled responses based on scores provided by a pre-trained reward model (RM), identifying the best-chosen and worst-rejected samples. Extensive experiments conducted on large-scale industry datasets demonstrate the superiority of OneRec in generating high-quality recommendations. The framework was successfully deployed in Kuaishou, a popular short video recommendation platform with millions of daily active users, resulting in a significant 1.6% increase in watch-time. Overall, OneRec represents a novel approach to recommendation systems that combines advanced generative modeling techniques with innovative preference alignment strategies to deliver superior performance in real-world scenarios.