Towards Large-scale Generative Ranking

TL;DR:
Why generative ranking works? How to make it production ready?

The auto-regressive architecture, not the training paradigm, is the primary source of effectiveness.

To overcome efficiency bottlenecks, engineers at Xiaohongshu introduce GenRank, an architecture that halves the effective sequence length by treating items as context and generating actions. Combined with efficient, parameter-free position biases (ALiBi), GenRank achieves a 94.8% training speed-up over the baseline with better offline AUC. In online A/B tests on tens of millions of users, it delivered significant engagement lifts with comparable resource costs and a >25% improvement in P99 response time.

The Generative Paradigm

Industrial recommender systems are typically multi-stage cascades, with the ranking stage acting as the final, fine-grained arbiter of what a user sees. While generative models have shown promise, their application in large-scale ranking has been under-explored. The team at Xiaohongshu (creators of RedNote) moved beyond simply proposing a new model to ask two fundamental system design questions:

1. What are the core mechanisms that make the generative paradigm effective for ranking?

2. How can we design a generative architecture that meets the stringent efficiency demands of a system serving hundreds of millions of users?