DIRECTOR: Optimal transport speeds up recommendation reranking by 10x

A new paper from researchers (Pu et al.) introduces DIRECTOR, a reranking framework that tackles the combinatorial complexity of selecting and ordering high-utility item slates for recommendation systems. Traditional autoregressive methods build slates position by position, capturing dependencies but suffering from sequential latency and limited search space under greedy decoding. Non-autoregressive alternatives run in parallel but often produce duplicate or conflicting items due to independent position-wise factorization. DIRECTOR bridges this gap by using optimal transport (OT) to guide parallel generation, mapping items into a continuous latent space and creating dynamic retrieval indices for all target positions at once. During training, entropy-regularized OT provides conflict-aware supervision, while inference uses hard matching on a similarity matrix to ensure duplicate-free results. A prefix-anchored credit assignment mechanism also aligns the generator with opaque list-wise evaluators, converting scalar rewards into position-specific signals.

Extensive offline and online experiments demonstrate DIRECTOR's effectiveness in large-scale industrial recommendation scenarios, consistently outperforming strong reranking baselines. The framework achieves significant improvements in utility while maintaining parallel efficiency, making it a promising solution for real-time systems. The paper includes 13 pages and one figure, and is available on arXiv (2607.26418). By combining the structural coordination of autoregressive models with the speed of non-autoregressive approaches, DIRECTOR represents a notable advance in recommendation reranking, with potential applications in search, advertising, and content feeds.