# Towards Large-scale Generative Ranking

**Authors:** Yanhua Huang, Yuqi Chen, Xiong Cao, Rui Yang, Mingliang Qi, Yinghao Zhu, Qingchang Han, Yaowei Liu, Zhaoyu Liu, Xuefeng Yao, Yuting Jia, Leilei Ma, Yinqi Zhang, Taoyu Zhu, Liujie Zhang, Lei Chen, Weihang Chen, Min Zhu, Ruiwen Xu, Lei Zhang  
**arXiv:** [2505.04180](https://arxiv.org/abs/2505.04180)  
**Date:** 2025-05-07  
**Summary page:** https://www.summarizepaper.com/en/paper/2505.04180v1/

## Key points

- Generative recommendation is a promising paradigm in information retrieval
- Generative ranking systems are particularly understudied in large-scale industrial settings
- Generative ranking outperforms current industrial recommenders primarily due to its architecture
- RankGPT is proposed as a novel architecture tailored for large-scale training and inference
- Main contributions of the paper include identifying sources of effectiveness in generative recommendation, proposing an efficient generative architecture for industrial scenarios, and conducting large-scale online A/B testing
- The study utilizes hundreds of billions of item exposure logs from Xiaohongshu's Explore Feed to construct datasets for offline experiments
- Generative ranking effectiveness is explored through experiments focusing on distinct underlying mechanisms that differentiate it from conventional approaches
- Scaling laws in recommendation systems are discussed, drawing parallels with principles in natural language processing and computer vision
- Generative ranking methods have the potential to enhance user satisfaction and performance metrics in industrial recommender systems

## Summary

Generative recommendation has emerged as a promising paradigm in information retrieval, with generative ranking systems being particularly understudied in large-scale industrial settings. This paper delves into the topic at the ranking stage of Xiaohongshu's Explore Feed, a recommender system serving millions of users. The study reveals that generative ranking outperforms current industrial recommenders primarily due to its architecture rather than the training paradigm. To address inefficiencies in current generative architectures, the authors propose RankGPT, a novel architecture tailored for large-scale training and inference. The main contributions of this paper include identifying the sources of effectiveness in generative recommendation, proposing an efficient generative architecture for industrial scenarios, and conducting large-scale online A/B testing to demonstrate the feasibility of generative ranking in recommender systems. The study utilizes hundreds of billions of item exposure logs from Xiaohongshu's Explore Feed to construct datasets for offline experiments, incorporating categorical features, numerical features discretized into categorical ones, and frozen embeddings from pre-trained models. The effectiveness of generative ranking is explored through experiments focusing on distinct underlying mechanisms that differentiate it from conventional approaches. The study also discusses scaling laws in recommendation systems, drawing parallels with established principles in natural language processing and computer vision. Overall, the findings highlight the potential of generative ranking methods in enhancing user satisfaction and performance metrics in industrial recommender systems.