Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Unbiased Multimodal Reranking for Long-Tail Short-Video Search

Wenyi Xu, Feiran Zhu, Songyang Li, Renzhe Zhou, Chao Zhang, Chenglei Dai, Yuren Mao, Yunjun Gao, Yi Zhang

Abstract

Kuaishou serving hundreds of millions of searches daily, the quality of short-video search is paramount. However, it suffers from a severe Matthew effect on long-tail queries: sparse user behavior data causes models to amplify low-quality content such as clickbait and shallow content. The recent advancements in Large Language Models (LLMs) offer a new paradigm, as their inherent world knowledge provides a powerful mechanism to assess content quality, agnostic to sparse user interactions. To this end, we propose a LLM-driven multimodal reranking framework, which estimates user experience without real user behavior. The approach involves a two-stage training process: the first stage uses multimodal evidence to construct high-quality annotations for supervised fine-tuning, while the second stage incorporates pairwise preference optimization to help the model learn partial orderings among candidates. At inference time, the resulting experience scores are used to promote high-quality but underexposed videos in reranking, and further guide page-level optimization through reinforcement learning. Experiments show that the proposed method achieves consistent improvements over strong baselines in offline metrics including AUC, NDCG@K, and human preference judgement. An online A/B test covering 15\% of traffic further demonstrates gains in both user experience and consumption metrics, confirming the practical value of the approach in long-tail video search scenarios.

Unbiased Multimodal Reranking for Long-Tail Short-Video Search

Abstract

Kuaishou serving hundreds of millions of searches daily, the quality of short-video search is paramount. However, it suffers from a severe Matthew effect on long-tail queries: sparse user behavior data causes models to amplify low-quality content such as clickbait and shallow content. The recent advancements in Large Language Models (LLMs) offer a new paradigm, as their inherent world knowledge provides a powerful mechanism to assess content quality, agnostic to sparse user interactions. To this end, we propose a LLM-driven multimodal reranking framework, which estimates user experience without real user behavior. The approach involves a two-stage training process: the first stage uses multimodal evidence to construct high-quality annotations for supervised fine-tuning, while the second stage incorporates pairwise preference optimization to help the model learn partial orderings among candidates. At inference time, the resulting experience scores are used to promote high-quality but underexposed videos in reranking, and further guide page-level optimization through reinforcement learning. Experiments show that the proposed method achieves consistent improvements over strong baselines in offline metrics including AUC, NDCG@K, and human preference judgement. An online A/B test covering 15\% of traffic further demonstrates gains in both user experience and consumption metrics, confirming the practical value of the approach in long-tail video search scenarios.

Paper Structure

This paper contains 35 sections, 10 equations, 3 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related works
  3. Architecture for Reranking
  4. Personalized Ranking with LLMs
  5. Multimodal Alignment and Synthetic Data
  6. Methodology
  7. Framework Overview
  8. Multimodal Quality Alignment
  9. Dataset.
  10. Supervised Fine Tuning
  11. Pairwise Ranking Alignment
  12. Dataset.
  13. Pairwise Preference Fine Tuning.
  14. Integration into the Ranking Pipeline
  15. Stage I: Pre-Training Enhancement
  16. ...and 20 more sections

Figures (3)

  • Figure 1: Data construction and two-stage training of the multimodal experience-scoring model, including supervised fine-tuning with multimodal annotations and pairwise fine-tuning with preference data.
  • Figure 2: Overview of the two-stage reranking integration. Stage I enhances pre-training with experience-score-ordered supervision; Stage II aligns page-level ranking through GRPO-based reward optimization guided by experience-derived nDCG.
  • Figure 3: Ablations within the two-stage optimization in reranking. Bars report NDCG under four label types: click: click-based labels, long = long-play labels, dur = watch-duration labels, human = human-preference labels.