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

2026-06-25 · Source: Takara TLDR - Daily AI Papers · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, medium

Summary

A new framework, MO-DiT+HPPO, addresses the challenge of pattern-preserving attribute retrieval in continuous generative retrieval systems. This task requires finding items that satisfy a target attribute while maintaining a fine-grained pattern from a seed set, a goal where traditional embedding-based methods often fail. MO-DiT+HPPO is a staged framework comprising raw-sequence pretraining, multi-domain metric-ordered continuation pretraining, tail-centroid fine-tuning, and Hybrid-Policy Preference Optimization (HPPO). Metric-ordered training converts sparse online retrieval labels into in-pattern trajectories, teaching the model to improve metrics across domains. HPPO then aligns the generated query distribution with the true online objective using a hybrid candidate pool and reference-anchored preference optimization, incorporating a Pareto pair filter to enhance attribute metrics without compromising pattern purity. Evaluations across four attribute domains demonstrate that MO-DiT improves the intersection metric over existing generative retrievers, with HPPO providing further significant gains on seven of eight domain-split cells.

Key takeaway

For Machine Learning Engineers developing generative retrieval systems, if you are struggling to balance pattern preservation with attribute density, consider implementing the MO-DiT+HPPO framework. This approach offers a structured way to improve intersection metrics by leveraging metric-ordered sequence training and hybrid-policy preference optimization. You should explore integrating its Pareto pair filter to ensure attribute gains do not compromise the desired pattern purity in your retrieval results.

Key insights

A staged generative retrieval framework improves pattern-preserving attribute retrieval by combining metric-ordered training and hybrid preference optimization.

Principles

• Pattern-preserving attribute retrieval balances pattern purity and attribute density.
• Metric-ordered training can teach domain-agnostic metric improvement.
• Hybrid-Policy Preference Optimization aligns generated queries with online objectives.

Method

MO-DiT+HPPO uses raw-sequence pretraining, multi-domain metric-ordered continuation pretraining, tail-centroid fine-tuning, and Hybrid-Policy Preference Optimization (HPPO) with a Pareto pair filter for generative retrieval.

In practice

• Use Pareto pair filtering to balance attribute gain and pattern purity.
• Apply metric-ordered training for cross-domain metric improvement.
• Implement HPPO for online objective alignment in generative retrieval.

Topics

• Generative Retrieval
• Diffusion Transformers
• Preference Optimization
• Metric-Ordered Training
• Pattern Preservation
• Embedding Retrieval

Code references

• Joyies/GDPO
• basiclab/DiffusionDRO

Best for: Research Scientist, AI Scientist, Machine Learning Engineer

Related on AIssential

• Time-Aware Diffusion based on Preference Disentanglement for Generative Recommendation — TDPM enhances generative recommenders by integrating time-aware diffusion and disentangling user preferences into period and point components.
• Evolutionary Generative Optimization: Towards Fully Data-Driven Evolutionary Optimization via Generative Learning — EvoGO leverages generative learning to create a fully data-driven evolutionary optimization framework, eliminating handcrafted heuristics.
• Drifting Preference Optimization for One-Step Generative Models — DrPO enables efficient preference finetuning for one-step generative models using black-box rewards and feature-space updates.
• DiffMath: Symbol- and Graph-Aware Latent Diffusion Transformer for Handwritten Mathematical Expression Generation — DiffMath generates HMEG without explicit spatial supervision by using LaTeX structure and a structured latent diffusion model.
• Metric-agnostic Learning-to-Rank via Boosting and Rank Approximation — A new LTR framework offers metric-agnostic optimization for improved ranking generalization and stability.
• Rethinking Token Reduction for Diffusion Models via Output-Similarity-Awareness — Output-centric token reduction in Diffusion Models, guided by prior-step output similarity, minimizes recovery error and boosts generation quality.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Takara TLDR - Daily AI Papers.