Neural Rerankers for Retrieval
Cross-encoder or LLM-based models that score each query and candidate document together to reorder a first-stage retrieval list.
Tier Grades
Tier A: Frontier AI labs
feasible
Alibaba, BAAI, NVIDIA and Cohere release rerankers from 0.6B to 8B (Qwen3 Embedding, C-Pack).
Tier B: Indonesian AI labs
feasible
Fine-tuning a 0.6B reranker on Indonesian query and passage pairs follows the published Qwen3 and BGE recipes; Indonesian lab capacity is not in the sources.
Tier C: Institution units such as ITB
feasible
Reranker training needs pairs and a benchmark of the C-MTEB kind, which a 6-author team built for Chinese (C-Pack).
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
A reranker added after vector search is a cheap quality step in RAG proofs of concept with open 0.6B models (Qwen3 Embedding).
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeQwen3 Embedding series ships rerankers of 0.6B, 4B and 8B parameters on Qwen3 backbones; training compute is not given in the abstract.
- DataBGE (C-Pack) rerankers are trained alongside the C-MTP data and evaluated on C-MTEB (35 datasets); Qwen3 rerankers use the same multi-stage pipeline with LLM-synthesised data.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 12 authors on the Qwen3 Embedding paper; 6 on C-Pack.
- Time horizonMonths.
- Shelf-life signalsRerankers moving from small cross-encoders (MiniLM) to LLM backbones (Qwen3, Llama Nemotron, mxbai on Qwen-2.5), Multimodal rerankers that score page images (Qwen3-VL-Reranker, Llama Nemotron Rerank VL)
- Pipeline roleTier A releases rerankers; Tier C builds Indonesian relevance sets and fine-tunes small rerankers; Tier D adds them to retrieval pipelines in proofs of value.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Linked models
Qwen3-Reranker-0.6Brerankerexemplar
bge-reranker-v2-m3rerankerexemplar
Rerank v3.5rerankerexemplar
ms-marco-MiniLM-L6-v2rerankerexemplar
jina-reranker-v3.5rerankerproduced by
R3-rerank-0.6brerankerproduced byQwen3-VL-Rerankerrerankerproduced byQwen3-VL-Reranker-2Brerankerproduced byQwen3-VL-Reranker-8Brerankerproduced by
Llama Nemotron Rerank VL 1B v2rerankerproduced byLlama Nemotron Rerank 1B v2rerankerproduced by
Granite Embedding Reranker English R2rerankerproduced byQwen3-Reranker-4Brerankerproduced byQwen3-Reranker-8Brerankerproduced by
mxbai-rerank-large-v2rerankerproduced bygte-reranker-modernbert-basererankerproduced bygte-multilingual-reranker-basererankerproduced byjina-reranker-v2-base-multilingualrerankerproduced by
MedCPT Cross-Encoderrerankerproduced bybge-reranker-largererankerproduced byBGE Reranker (v1)rerankerproduced bybge-reranker-basererankerproduced byCohere Rerankrerankerproduced byRerank v4.0 Fastrerankerproduced byLinks
R&D TopicNeural Rerankers for Retrieval
ModelsQwen3-Reranker-0.6BAlibaba (Qwen), rerankerbge-reranker-v2-m3BAAI, rerankerRerank v3.5Cohere, rerankerms-marco-MiniLM-L6-v2Sentence Transformers, reranker+25 more
Qwen3-Reranker-0.6BAlibaba (Qwen), rerankerbge-reranker-v2-m3BAAI, rerankerRerank v3.5Cohere, rerankerms-marco-MiniLM-L6-v2Sentence Transformers, reranker+25 moreUse Casesnone recorded
Who Runs Itnone recorded
Models 29
Cohere RerankCohere, rerankerproduced byQwen3-VL-RerankerAlibaba (Qwen), rerankerproduced byBGE Reranker (v1)BAAI, rerankerproduced byMS MARCO cross-encoders (MiniLM)Sentence Transformers, rerankerproduced byQwen3-VL-Reranker-2BAlibaba (Qwen), rerankerproduced byQwen3-VL-Reranker-8BAlibaba (Qwen), rerankerproduced bybge-reranker-baseBAAI, rerankerproduced bybge-reranker-largeBAAI, rerankerproduced byGranite Embedding Reranker English R2IBM, rerankerproduced byLlama Nemotron Rerank VL 1B v2NVIDIA, rerankerproduced byms-marco-MiniLM-L12-v2Sentence Transformers, rerankerproduced byms-marco-MiniLM-L6-v2Sentence Transformers, rerankerexemplar+17
Qwen3-Reranker-0.6BAlibaba (Qwen), rerankerexemplarQwen3-Reranker-4BAlibaba (Qwen), rerankerproduced byQwen3-Reranker-8BAlibaba (Qwen), rerankerproduced byRerank v3.5Cohere, rerankerexemplarRerank v4.0 FastCohere, rerankerproduced byRerank v4.0 ProCohere, rerankerproduced by
rerank-2.5Voyage AI, rerankerproduced byVoyage rerankersVoyage AI, rerankerproduced bybge-reranker-v2-m3BAAI, rerankerexemplargte-multilingual-reranker-baseAlibaba (Qwen), rerankerproduced bygte-reranker-modernbert-baseAlibaba (Qwen), rerankerproduced byjina-reranker-v2-base-multilingualJina AI, rerankerproduced byjina-reranker-v3.5Jina AI, rerankerproduced byLlama Nemotron Rerank 1B v2NVIDIA, rerankerproduced byMedCPT Cross-EncoderNIH National Library of Medicine, rerankerproduced bymxbai-rerank-large-v2Mixedbread, rerankerproduced byR3-rerank-0.6bTencent, rerankerproduced byPapers
- Passage Re-ranking with BERTRodrigo Nogueira et al., arXiv preprint, 2019 founding
- Is ChatGPT Good at Search? Investigating Large Language Models as Re-Ranking AgentsWeiwei Sun et al., EMNLP 2023 (per arXiv comment), 2023 state of the art
- RankZephyr: Effective and Robust Zero-Shot Listwise Reranking is a Breeze!Ronak Pradeep et al., arXiv preprint, 2023 state of the art
- ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERTOmar Khattab et al., SIGIR 2020 (per arXiv comment), 2020 state of the art
- Pretrained Transformers for Text Ranking: BERT and BeyondJimmy Lin et al., arXiv preprint, 2020 survey
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/2506.05176 https://arxiv.org/abs/2506.05176
- arxiv.org/abs/2309.07597 https://arxiv.org/abs/2309.07597
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Neural Rerankers for Retrieval", https://aurion.semai.id/topics/neural-rerankers-for-retrieval/, records of October 2026.