# OneRec Technical Report

**Authors:** Guorui Zhou, Jiaxin Deng, Jinghao Zhang, Kuo Cai, Lejian Ren, Qiang Luo, Qianqian Wang, Qigen Hu, Rui Huang, Shiyao Wang, Weifeng Ding, Wuchao Li, Xinchen Luo, Xingmei Wang, Zexuan Cheng, Zixing Zhang, Bin Zhang, Boxuan Wang, Chaoyi Ma, Chengru Song  
**arXiv:** [2506.13695](https://arxiv.org/abs/2506.13695)  
**Date:** 2025-06-16  
**Summary page:** https://www.summarizepaper.com/en/paper/2506.13695v1/

## Key points

- Longstanding reliance on multi-stage cascaded architectures in recommender systems
- Limitations due to computational fragmentation and optimization inconsistencies
- Development of OneRec by Guorui Zhou, Jiaxin Deng, and team as a groundbreaking approach
- OneRec reshapes recommendation systems through an end-to-end generative framework
- Enhances computational FLOPs of existing models by 10 times and establishes scaling laws for recommendations
- Leverages reinforcement learning techniques to optimize recommendations effectively
- Achieves impressive Model FLOPs Utilization rates on flagship GPUs during training and inference stages
- Deployment in Kuaishou/Kuaishou Lite APP results in handling a quarter of total queries per second and enhancing App Stay Time significantly
- Substantial increases in metrics like 7-day Lifetime post-OneRec implementation, improving user engagement and satisfaction
- Drastic reduction in operational expenses associated with traditional recommendation pipelines to just 10.6%
- Technical report authored by Zhou et al. provides insights into the development and optimization process behind OneRec, with real-world implications for production-scale recommendation systems

## Summary

In the field of recommender systems, there has been a longstanding reliance on multi-stage cascaded architectures that have not kept pace with the rapid advancements in artificial intelligence. This has led to computational fragmentation and optimization inconsistencies, limiting the effective integration of key AI breakthroughs into recommendation scenarios. To address these challenges, a team of researchers led by Guorui Zhou, Jiaxin Deng, and their colleagues have developed OneRec - a groundbreaking approach that reshapes recommendation systems through an end-to-end generative framework. OneRec represents a significant leap forward in recommendation technology by enhancing the computational FLOPs of existing models by 10 times and establishing scaling laws for recommendations within specific boundaries. By leveraging reinforcement learning techniques previously deemed challenging for optimizing recommendations, OneRec demonstrates substantial potential in improving recommendation accuracy and efficiency. Furthermore, through infrastructure optimizations, the team has achieved impressive Model FLOPs Utilization rates on flagship GPUs during both training and inference stages - aligning closely with leading-edge practices in the AI community. The deployment of OneRec in the Kuaishou/Kuaishou Lite APP has yielded remarkable results - handling a quarter of total queries per second while enhancing overall App Stay Time by significant margins. Notably, metrics such as 7-day Lifetime have shown substantial increases following the implementation of OneRec - underscoring its positive impact on user engagement and satisfaction. Additionally, operational expenses associated with traditional recommendation pipelines have been drastically reduced to just 10.6% through the adoption of this innovative architecture. The technical report authored by Zhou et al. not only presents the development and optimization process behind OneRec but also offers valuable insights derived from maintaining a production-scale recommendation system with tangible real-world implications. The comprehensive approach taken by the research team showcases how cutting-edge technologies can be effectively harnessed to revolutionize recommender systems and enhance user experiences across diverse platforms.