This research project explored how ListT5-style reranking behaves when the inference process is reorganized rather than retrained. The work focused on grouping candidates, reducing repeated neural forward calls with caching, handling duplicate-index edge cases, and comparing search-quality outcomes across BEIR datasets under realistic compute constraints.
Information Retrieval / 2026
A Preliminary Study of Grouping Strategies for ListT5 Reranking
Extended the inference procedure of ListT5, a listwise reranker based on Fusion-in-Decoder T5, by studying grouping strategies, tournament-style reranking, caching behavior, and BEIR benchmark trade-offs without retraining the underlying model.