Tag
Meta researchers explore three diffusion-LLM parameterizations for permutation generation — hard-assignment, soft-assignment, and direct ordering-score prediction — evaluating them on recommendation reranking across Amazon domains and showing that design choices must be jointly optimized for corruption, parameterization, objective, and decoding.
This paper empirically studies Jev, a decision-oriented "System One Model," for personalized recommendation reranking, finding it occupies a distinct quality–latency operating regime compared with recommendation-specific models and pointwise/listwise LLM rerankers across multiple domains and candidate-set sizes.