DenoiseRank: Learning to Rank by Diffusion Models

Researchers Ying Wang, Preslav Nakov, and Shangsong Liang have introduced DenoiseRank, a novel framework that reframes the classic learning to rank (LTR) problem through the lens of generative diffusion models. Unlike nearly all existing LTR methods, which operate from a discriminative perspective (e.g., classifying or scoring documents relative to queries), DenoiseRank treats ranking as a denoising process. It first adds controlled noise to relevance labels during a forward diffusion phase, then learns to reverse this process conditioned on query-document pairs to recover accurate relevance distributions. This marks the first time diffusion models have been applied to traditional LTR tasks, opening a new generative paradigm for information retrieval.

Extensive experiments on standard benchmark datasets demonstrate that DenoiseRank achieves competitive performance, establishing a strong baseline for generative LTR. The method's key advantage lies in its ability to model uncertainty and complex relevance patterns more naturally than discriminative approaches. By leveraging the iterative refinement capabilities of diffusion, the model can potentially handle noisy or ambiguous relevance judgments better. This work bridges the gap between recent advances in generative AI and core information retrieval tasks, suggesting that search engines, recommendation systems, and any application relying on ranking could benefit from this probabilistic, noise-aware approach. The paper is available on arXiv (2604.20852).