In a team of two, I studied how candidate grouping and tournament-style inference affect ListT5, a listwise reranker based on Fusion-in-Decoder T5. We kept the pretrained checkpoints fixed so the experiments examined inference design rather than changes from model training.
The implementation work included strategy-aware first-round caching to avoid repeated neural forward calls and fixes for duplicate-index edge cases in candidate handling.
We evaluated the approaches across five BEIR benchmark datasets, measuring NDCG@10, forward-call counts, and query latency. The study compared retrieval quality with the amount of model computation each grouping strategy required.