MO-DiT+HPPO Framework Targets Pattern-Preserving Attribute Retrieval in Production

MO-DiT+HPPO addresses pattern-preserving attribute retrieval — finding items matching both a fine-grained seed pattern and a target attribute — using a diffusion transformer with metric-ordered sequence pretraining and hybrid-policy preference optimization. The staged training framework is motivated by production retrieval settings where simple embedding similarity fails.