Generative Retrieval via Diffusion Transformer with Metric-Ordered Sequence Training and Hybrid-Policy Preference Optimization

2026-06-25

AI Summary

This work addresses the challenge of simultaneously preserving pattern fidelity and enhancing target attributes in pattern-preserving attribute retrieval. To this end, the authors propose the MO-DiT+HPPO framework, which leverages a diffusion Transformer (DiT) architecture and employs a continuous generative retrieval mechanism to transform seed item embeddings into query embeddings for nearest-neighbor search. The approach incorporates a multi-stage training strategy comprising metric-ordered sequence pretraining, tail-centroid fine-tuning, and hybrid-policy preference optimization (HPPO), augmented with Pareto pair filtering and reference-anchored preference learning. This design effectively maintains fine-grained pattern purity while substantially boosting target attributes. Evaluated across four attribute domains under item- and pattern-holdout protocols, the method achieves significant performance gains in seven out of eight domain-split units and matches baseline performance in the remaining one, consistently outperforming existing generative retrieval approaches.

Abstract

Embedding-based retrieval ranks items by their similarity to a query in a shared vector space and usually aims to return the highest-scoring items. In many production settings this is not what is wanted: given a seed set that expresses a fine-grained pattern, one needs more items that both satisfy a target attribute and stay within that pattern. We formalize this as pattern-preserving attribute retrieval. The two goals pull against each other: averaging the seeds preserves the pattern but stays in a low-attribute region, while global attribute retrieval drifts to unrelated patterns. We approach the task with continuous generative retrieval, where a model reads a sequence of item embeddings and generates query embeddings for nearest-neighbor search. We propose MO-DiT+HPPO, a staged framework with raw-sequence pretraining, multi-domain metric-ordered continuation pretraining, tail-centroid fine-tuning, and HPPO. Metric-ordered training turns sparse online retrieval labels into in-pattern trajectories ordered from low to high predicted attribute density, teaching one model the metric-improvement direction across domains. HPPO aligns the generated query distribution with the true online objective by labeling a hybrid candidate pool with the online intersection metric and applying reference-anchored preference optimization. A Pareto pair filter keeps only winner pairs that do not lower same-pattern purity, raising the attribute metric without sacrificing the pattern. Across four attribute domains under item- and pattern-holdout protocols, metric-ordered DiT improves the intersection metric over a pretrained generative retriever, and HPPO improves it further, with significant gains on seven of eight domain-split cells and a marginal tie on the hardest split. Metric-predictor validation, order ablations, CPT/SFT comparisons, and a candidate-policy ablation show where the gains come from.

Problem

• pattern-preserving retrieval
• attribute retrieval
• generative retrieval
• embedding-based retrieval
• retrieval with constraints

Innovation

• Generative Retrieval
• Metric-Ordered Training
• Hybrid-Policy Preference Optimization
• Pattern-Preserving Retrieval
• Diffusion Transformer