Reproduce “UniRank: Unified List-wise Reranking via Confidence-Ordered Denoising” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Pengyue Jia, Hailan Yang, Shuchang Liu, Xiaobei Wang, Wanyu Wang, Xiang Li, Yongqi Liu, Kaiqiao Zhan and 2 more

UniRank introduces a unified list-wise reranking framework that employs confidence-ordered iterative denoising to enable bidirectional inter-item interaction modeling while mitigating error propagation. This approach, which adapts diffusion models for item-level and candidate-constrained slate generation, consistently surpassed state-of-the-art baselines by over 5% on various metrics and yielded improvements in user engagement during large-scale online A/B tests.

Key insights

  • A unified framework can reconcile the strengths of AR and NAR paradigms by allowing flexible, confidence-ordered commitment during iterative denoising, effectively addressing both error propagation and comprehensive inter-item dependency modeling.
  • Adapting diffusion models for reranking tasks critically benefits from task-specific interfaces like the Semantic Fusion Layer for item-level denoising and Latent Pool Selection for candidate-constrained generation.
  • Performance gains are consistently observed with iterative denoising, where prioritizing high-confidence decisions early provides a more robust and complete contextual signal for refining subsequent, more uncertain item placements.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.