ListT5
Listwise reranking without a giant LLM — a Fusion-in-Decoder T5 that sorts candidate passages with a tournament tree.
About
ListT5 is the answer to a practical annoyance: listwise reranking worked, but it needed either quadratic pairwise comparisons or a full-size LLM with a "lost in the middle" problem. Soyoung Yoon and collaborators at Seoul National University and LG AI Research took a different route — a T5 model built on Fusion-in-Decoder that looks at five candidate passages at once, each paired with the query, and emits their identifiers sorted by increasing relevance.
At inference time a tournament-tree structure turns those per-group orderings into a full ranking with O(n + k log n) cost, roughly pointwise efficiency at listwise quality. The paper (ACL 2024, arXiv 2402.15838) reports a +1.3 average NDCG@10 gain over RankT5 on BEIR — a real, held-out improvement, not a training-set artifact.
Unusually for a paper from this space, the code, checkpoints, and the evaluation framework are fully open-sourced on GitHub, so the results are reproducible end to end.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper-reported (ACL 2024, arXiv 2402.15838): +1.3 average NDCG@10 over RankT5 baseline across BEIR; BEIR average 53.1 (r=2 variant) on COCO-DR top-100 first-stage retrieval
- Independent comparison (E2Rank paper table, 2026): ListT5-3B averages 55.74 NDCG@10 across 8 BEIR rerank sets vs RankT5-3B 55.40 and MonoT5-3B 53.93 — the highest of the three in that comparison
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (TIER verified, metrics-v1). Evidence: peer-reviewed ACL 2024 paper (arXiv 2402.15838) from SNU/LG AI Research with a concrete +1.3 NDCG@10 BEIR claim; authors state code, checkpoints and eval framework are fully open-sourced at github.com/soyoung97/ListT5; corroborated by the E2Rank independent comparison. Strengths: reproducible open release, elegant tournament-tree efficiency story, genuine gain over a strong baseline. Docked for: (1) headline numbers are author-reported; (2) most experiments used T5-base with few 3B runs — scaling behavior is thin. License: open-source release per the authors.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · ACL 2024 paper (arXiv 2402.15838) from SNU/LG AI Research: +1.3 average NDCG@10 over RankT5 on BEIR via Fusion-in-Decoder listwise ranking with tournament-tree inference.
- · E2Rank independent comparison: ListT5-3B 55.74 avg NDCG@10 — highest of the T5 reranker trio in that table.
- · Most experiments used T5-base with few 3B runs; scaling behavior is thin. Headline numbers are author-reported.