Research map / Image generation / Autoregressive and latent generation

UniRank: Unified List-wise Reranking via Confidence-Ordered Denoising

NeurIPS 2026 · Image generation › Autoregressive and latent generation · approach: Accelerating autoregressive image generation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes UniRank, a masked discrete diffusion reranker that fills the most confident slot-item pair per step, recovering AR and NAR decoding as special cases, with a Task Grounded Diffusion interface for item-level, candidate-constrained prediction.

Problem

List-wise reranking must model inter-item dependencies in an exposure slate while avoiding error propagation of autoregressive decoding and the slot-independence limitation of non-autoregressive decoding.

Approach

Masked discrete diffusion over slate slots with bidirectional attention, committing the highest-confidence slot-candidate pair each step; item-level embeddings (SFL) and dot-product scoring over the request's candidate pool (LPS) make the diffusion interface match reranking.

Key results

• On Amazon Books, UniRank reaches P@6 88.30 and NDCG@6 94.00 vs. GloRank 83.75 and 90.08 (+5.43% P, +4.35% NDCG relative).
• Online A/B test: +0.159% user average app-time and +1.016% share-rate (significant at p<0.05).
• Simulator reward on MovieLens-1M: 3.0551 for UniRank vs. 2.9730 for GloRank.

Surprising findings

• Removing iterative decoding (single parallel step) only slightly degrades performance, e.g., NDCG@6 93.89 vs 94.00 on Amazon Books.
• Refinement stage gives only marginal changes (e.g., Precision@6 0.7668 to 0.7671), and more refinement steps do not always help.

Limitations

• Final slate quality is bounded by upstream retrieval and ranking candidate pools.
• Gains over strongest baselines are modest on MovieLens-1M and Industry (about 1-2% relative).
• Refinement stage yields only small and non-uniform gains.

Experimental setup

Datasets: Amazon Books MovieLens-1M Industry short-video log

Benchmarks: Amazon Books MovieLens-1M Industry

Metrics: Precision@6 NDCG@6 MAP@6 F1@6 average reward (simulator)

Baselines: DNN DLCM PRM GoalRank GloRank PIER NAR4Rec G-n DCDR

Builds on: masked discrete diffusion GoalRank GloRank DCDR NAR4Rec VQ-VAE

Closest papers by meaning

• Very Efficient Listwise Multimodal Reranking for Long Documents (ICML 2026 · similarity 0.87)
• AcuRank: Uncertainty-Aware Adaptive Computation for Listwise Reranking (NeurIPS 2025 · similarity 0.87)
• Contrastive Retrieval Heads for Improved Attention-Based Reranking (NeurIPS 2026 · similarity 0.87)
• Attention in Large Language Models Yields Efficient Zero-Shot Re-Rankers (ICLR 2025 · similarity 0.86)
• Language Ranker: A Lightweight Ranking framework for LLM Decoding (NeurIPS 2025 · similarity 0.85)
• DynamicRAG: Leveraging Outputs of Large Language Model as Feedback for Dynamic Reranking in Retrieval-Augmented Generation (NeurIPS 2025 · similarity 0.84)