Source-linked AI summary
jina-reranker-v3: Last but Not Late Interaction for Listwise Document Reranking
Feng Wang, Yuqing Li, Han Xiao
TL;DR
Neural retrieval faces a trade-off between cross-encoder effectiveness and efficient separately encoded representations. jina-reranker-v3 introduces listwise causal attention between a query and multiple documents before extracting contextual embeddings. It achieves 61.85 nDCG@10 on BEIR, the highest score among evaluated rerankers, while its multilingual result trails a dedicated multilingual model.
Problem
Neural retrieval must balance cross-encoder interaction quality, which requires separate query-document passes, against efficient embedding models that lose fine-grained interaction signals.
Method
jina-reranker-v3 applies causal attention between a query and all candidate documents in one shared context window, then extracts contextual embeddings for similarity scoring.
Results
61.85 nDCG@10 on BEIR is the highest score among evaluated rerankers, improving 4.79% over jina-reranker-v2.
Takeaways & Limitations
The 0.6B-parameter model combines listwise cross-document interaction with efficient embedding-based similarity computation.
Takeaways & Limitations
On MIRACL, jina-reranker-v3 averages 2.82 points below bge-reranker-v2-m3’s dedicated multilingual optimization.
Abstract
from arXiv · showhide
jina-reranker-v3 is a 0.6B-parameter multilingual listwise reranker that introduces a novel "last but not late" interaction. Unlike late interaction models like ColBERT that encode documents separately before multi-vector matching, our approach applies causal attention between the query and all candidate documents in the same context window, enabling rich interactions before extracting contextual embeddings from each document's final token. The new model achieves state-of-the-art BEIR performance with 61.94 nDCG@10 while being significantly smaller than other models with comparable performance.
1 Introduction
Neural retrieval must balance cross-encoder effectiveness against the efficiency of separately encoded representations. jina-reranker-v3 addresses this tradeoff with listwise causal interaction and reports leading reranking performance.
- Late interaction models separately encode queries and documents before token-level similarity, preserving efficiency but delaying their interaction.
- jina-reranker-v3 applies causal attention between the query and all documents in one context window before extracting document embeddings.
- 61.85 nDCG@10 on BEIR is the highest score among evaluated rerankers and improves 4.79% over jina-reranker-v2.
2 Related Work
Related work spans learning-to-rank objectives, late interaction retrieval, and LLM-powered rerankers. These approaches trade computational efficiency, fine-grained matching, and model-based relevance reasoning in different ways.
- Cross-encoders jointly process each query-document pair for strong interaction but require separate forward passes, creating computational bottlenecks.
- ColBERT independently encodes queries and documents into multi-vector representations, then computes token-level similarity while allowing document pre-computation.
- LLM-powered rerankers include generative approaches that produce ranked lists and fine-tuning methods that adapt pretrained models for relevance scoring.
3 Model Architecture
The architecture processes a query and multiple documents together, enabling causal cross-document interaction before extracting embeddings from designated special-token positions. A projector then converts these contextual states into vectors scored by cosine similarity.
- 3.1 Architecture: jina-reranker-v3 processes the query and multiple documents simultaneously within shared context windows using a Qwen3-0.6B backbone.
- 3.1 Architecture: Causal self-attention lets each document attend to other documents, capturing inter-document relationships unavailable to separate encoding.
- 3.1 Architecture: Special token positions extract contextual query and document embeddings from final-layer hidden states after causal self-attention.
- 3.1 Architecture: A two-layer projector maps 1024-dimensional hidden states to 256-dimensional embeddings, whose cosine similarity produces relevance scores.
- 3.1 Architecture: Documents exceeding the 131K-token limit are processed in batches of up to 64 while maintaining query embeddings across batches.
- 3.2 Prompt Template: The prompt places the query at both sequence ends, sandwiching documents so the final query position can attend to all preceding documents.
4 Training
Training combines ranking, dispersion, dual-matching, and similarity objectives with progressive stages that increase context length, hard negatives, and domain specialization. The final model merges specialized models while retaining architectural efficiency.
- 4.1 Loss Functions: The total objective combines ranking, dispersive, dual-matching, and similarity losses with weights 1, 0.45, 0.85, and 0.85.
- 4.1 Loss Functions: InfoNCE ranking loss supplies the core contrastive signal using hard negative documents.
- 4.1 Loss Functions: Dispersive loss increases embedding diversity, while dual matching enforces consistency between query-to-document and document-to-query scores.
- 4.1 Loss Functions: Similarity loss treats augmented document versions as positive pairs to encourage consistent representations across surface-form changes.
- Training Stages: Training progresses from 16-document foundation batches to longer documents and up to 45 negatives under the 131K-token limit.
- Training Stages: The final stage linearly merges specialized models with weights from 0.25 to 0.65 to combine domain expertise.
5 Evaluation
Across four benchmarks, jina-reranker-v3 achieves strong English, multilingual, cross-lingual, and code-retrieval performance while remaining competitive with larger rerankers. Its results show gains over same-scale and larger alternatives, alongside a multilingual trade-off against a specialized baseline.
- Overall Performance Across Benchmarks: 61.85 nDCG@10 on BEIR is the highest score among evaluated rerankers, improving 4.79% over jina-reranker-v2’s 57.06.Table 2 reports reranking over top-100 candidates from jina-embeddings-v3; BEIR uses nDCG@10.
- Overall Performance Across Benchmarks: 61.85 versus 61.44 gives jina-reranker-v3 higher BEIR performance than 1.5B mxbai-rerank-large-v2 with 2.5× fewer parameters.The model also reaches 70.64 on CoIR, providing specialized code-retrieval coverage.
- Multilingual Performance on MIRACL: 66.83 on MIRACL across 18 languages and 67.92 Recall@10 on MKQA across 26 languages demonstrate broad multilingual and cross-lingual capability.MKQA uses Recall@10, while the other listed benchmark scores use nDCG@10.
- Multilingual Performance on MIRACL: The 66.83 MIRACL score trails the multilingual-specialized bge-reranker-v2-m3 at 69.32 by 2.49 points, exposing a multilingual performance trade-off.The evaluation describes this as a trade-off alongside stronger English performance and architectural efficiency.
- English Retrieval Performance on BEIR: 78.58 on HotpotQA and 94.01 on FEVER mark particularly strong results for multi-hop reasoning and fact verification.These tasks are included among BEIR’s heterogeneous evaluation settings.
- English Retrieval Performance on BEIR: 61.85 versus 56.51 yields a 5.34% improvement over the same-scale 0.6B bge-reranker-v2-m3.The passage attributes this comparison to the model’s architectural design rather than simple parameter scaling.
6 Conclusion
The paper presents jina-reranker-v3 as a compact multilingual listwise reranker built around last but not late interaction. It processes queries and multiple documents together, enabling cross-document interaction during encoding before extracting ranking embeddings.
- Conclusion: jina-reranker-v3 is a 0.6B-parameter multilingual listwise reranker introducing last but not late interaction for efficient document reranking.The method processes queries and multiple documents simultaneously within shared context windows.
- Conclusion: Causal attention across the shared query-document context enables cross-document interactions during encoding before contextual embeddings are extracted from designated token positions.The approach adapts long-context generative LLMs into a discriminative reranker.
A Model Configuration and Training Details
The model configuration and training details are summarized through an architecture table and a multi-stage fine-tuning hyperparameter table spanning 47 configurations.
- Table 5 presents the architecture configuration for jina-reranker-v3.
- Table 6 reports multi-stage supervised fine-tuning hyperparameter ranges across 47 training configurations.