This post describes a reranking method where an LLM generates yes/no relevance questions for a query, and Jev scores each question's probability of yes on a document; the summed probabilities form the final rerank score.

On Wayfair WANDS and Amazon ESCI datasets, this bag-of-decisions approach outperformed BM25 and a single-question Jev reranker in mean NDCG. The technique reframes query evaluation as per-query feature generation rather than bag-of-terms matching.