RankT5

RankT5 asked a pointed question: if ranking is not classification, why train rankers with classification losses? Honglei Zhuang and colleagues at Google Research kept the T5 backbone from monoT5 but replaced the true/false head with direct ranking scores, and trained with pairwise and listwise losses that optimize the actual ordering. They studied both an encoder-decoder and an encoder-only variant.

The payoff was substantial gains over monoT5 on MS MARCO and, notably, better zero-shot performance on out-of-domain datasets when trained listwise — a property that mattered once BEIR-style evaluation became standard. Google published the paper (arXiv 2210.10634) and pointed to reference code under googleresearch, but it never became a supported product; it is a research artifact that improved the field's training recipe.

ListT5 later beat it by about 1.3 points of average NDCG@10 on BEIR, and closed models have since moved past it. It remains the cleanest demonstration that ranking losses beat classification losses for neural rerankers.

• Paper-reported (arXiv 2210.10634, Oct 2022): substantial ranking gains over monoT5-style classification baselines on MS MARCO and BEIR-family datasets, with listwise training improving zero-shot out-of-domain transfer
• Independent comparison (E2Rank paper table, 2026): RankT5-3B averages 55.40 NDCG@10 across 8 BEIR rerank sets vs MonoT5-3B 53.93 and ListT5-3B 55.74