NAVER LABS Europe's RRK Reranker Runs 3x-18x Faster Than Smaller Models
Researchers at NAVER LABS Europe have introduced RRK, a listwise reranker that compresses documents into fixed-size embedding representations. According to the authors, their 8B-parameter model runs 3x-18x faster than smaller rerankers of 0.6-4B parameters while matching or outperforming them in effectiveness.
- NAVER LABS Europe introduces RRK, a listwise reranker that compresses documents into multi-token fixed-size embeddings.
- According to the authors, the 8B-parameter model runs 3x-18x faster than rerankers with 0.6-4B parameters.
- It matches or outperforms those smaller rerankers in effectiveness, the paper says.
- Gains are reported to be larger on long-document benchmarks.
- The work is by Hervé Déjean and Stéphane Clinchant and is listed for EMNLP in Budapest, 24-29 October, 2026.
What RRK does
Reranking is the step that refines the output of a first-stage retriever. According to the paper, it is often considered computationally expensive, especially when Large Language Models (LLMs) are used. The usual ways to reduce that cost are to use smaller LLMs or to control the length of the input.
Hervé Déjean and Stéphane Clinchant, the authors, take a different route. Inspired by recent advances in document compression for retrieval-augmented generation (RAG), they present RRK, which compresses documents into multi-token fixed-size embedding representations and then ranks them as a list.
Training and reported results
The authors say the system is trained in a simple way, through distillation. In their words, the combination of rich compressed representations and listwise reranking yields a highly efficient and effective system.
The headline figure comes from a comparison of model sizes. The 8B-parameter RRK model runs 3x-18x faster than smaller rerankers with 0.6-4B parameters, according to the authors, and it matches or outperforms them in effectiveness. The result is notable because the larger model is the faster one, which reverses the common assumption that a smaller model is the cheaper choice.
The paper also reports that the efficiency gains are even more striking on long-document benchmarks, where RRK widens its advantage further. The source does not give specific benchmark names or scores for those tests.
Where it will be presented and why it matters
The work is listed for the Conference on Empirical Methods in Natural Language Processing (EMNLP), taking place in Budapest, Hungary, from 24 to 29 October, 2026. An arXiv version is linked from the publication page.
For teams building search and RAG pipelines, reranking cost is a practical constraint, and long documents make it worse. If the authors' claims hold up, compressing documents before ranking offers a way to use a larger reranker without paying more in speed. The source does not describe code release or deployment plans, so how widely the approach can be adopted remains to be seen.
NAVER LABS Europe: Efficient listwise reranking with compressed document representations
