SPLICE: When a Weaker LLM Ranker Helps a Strong Sequential Recommender

Abstract

Can a large language model improve a strong sequential recommender even when it ranks worse on its own? We show that it can with SPLICE (Single-Pass LLM-Informed Candidate Evaluation), which keeps a sequential recommender frozen as the retriever and adds an open-weight LLM decision model (Clef, of the same model class as the hosted Jev) that reads the user's history once and assigns a probability to each of the top-50 candidates in a single forward pass; a seven-weight fusion, fitted on validation interactions, combines the two without fine-tuning either model. Evaluated by full ranking on all test users, without injecting the target into the candidate list, SPLICE improves NDCG@10 by 12.1% on Amazon Beauty and 13.1% on Toys, to values above those published for FENRec and BSARec on the same data. It also significantly improves Sports, MovieLens-1M, a second, SASRec-style retriever and every retriever seed we trained, although the decision model alone ranks worse than the retriever in every setting. Removing the decision model's score from the fusion removes most of the gain (all of it on the Amazon datasets), and the fusion weights transfer across datasets and retrievers without refitting. On the same users and GPU, the single forward pass has a 9.7x lower median latency per request than a similar-size general LLM that generates its ranking, which is also less accurate, and the same general LLM read out in a single pass through its candidate-label probabilities does not improve the retriever at all. A decision model's stand-alone ranking accuracy thus understates its value as a complementary reranker.