Next-Scale Generative Reranking: A Tree-based Generative Rerank Method at Meituan

2026-04-06

AI Summary

Existing generative re-ranking methods suffer from limited performance due to an imbalance between global and local information and misaligned training objectives between generators and evaluators. To address these issues, this work proposes NSGR, a tree-structured generative re-ranking framework that models user interests through a coarse-to-fine hierarchical generation mechanism. The approach introduces a novel multi-scale neighbor loss function and a co-training strategy to align the objectives of generators and evaluators across all levels of the hierarchy. By effectively integrating the strengths of both non-autoregressive and autoregressive generation paradigms, NSGR achieves substantial improvements over state-of-the-art baselines on both public and industrial datasets. The method has been successfully deployed in Meituan’s food delivery recommendation system, demonstrating its practical efficacy and scalability.

Abstract

In modern multi-stage recommendation systems, reranking plays a critical role by modeling contextual information. Due to inherent challenges such as the combinatorial space complexity, an increasing number of methods adopt the generative paradigm: the generator produces the optimal list during inference, while an evaluator guides the generator's optimization during the training phase. However, these methods still face two problems. Firstly, these generators fail to produce optimal generation results due to the lack of both local and global perspectives, regardless of whether the generation strategy is autoregressive or non-autoregressive. Secondly, the goal inconsistency problem between the generator and the evaluator during training complicates the guidance signal and leading to suboptimal performance. To address these issues, we propose the \textbf{N}ext-\textbf{S}cale \textbf{G}eneration \textbf{R}eranking (NSGR), a tree-based generative framework. Specifically, we introduce a next-scale generator (NSG) that progressively expands a recommendation list from user interests in a coarse-to-fine manner, balancing global and local perspectives. Furthermore, we design a multi-scale neighbor loss, which leverages a tree-based multi-scale evaluator (MSE) to provide scale-specific guidance to the NSG at each scale. Extensive experiments on public and industrial datasets validate the effectiveness of NSGR. And NSGR has been successfully deployed on the Meituan food delivery platform.

Problem

• generative reranking
• recommendation systems
• goal inconsistency
• combinatorial optimization
• contextual modeling

Innovation

• generative reranking
• tree-based framework
• multi-scale evaluation
• next-scale generation
• recommendation systems