TLDR: E2RANK is a novel framework that transforms a single text embedding model into both an efficient retriever and a high-quality listwise reranker. It achieves this by reinterpreting the listwise prompt (query + candidate documents) as a pseudo-relevance feedback query and employing continued training with a listwise ranking objective. This approach significantly reduces inference latency (up to 5x faster than LLM-based rerankers) while achieving state-of-the-art reranking performance on benchmarks like BEIR and competitive results on reasoning-intensive tasks, all while maintaining strong general embedding capabilities.
In the world of search and information retrieval, getting accurate and relevant results quickly is paramount. Traditionally, this process often involves a two-stage system: a fast but less precise initial retrieval using text embedding models, followed by a more powerful but slower reranking step, often powered by large language models (LLMs). While LLM-based rerankers excel at understanding complex relationships between queries and documents, their computational cost and latency can be a significant bottleneck for real-time applications.
A new research paper introduces a novel framework called E2RANK, which stands for Efficient Embedding-based Ranking (or Embedding-to-Rank). This innovative approach aims to bridge the gap between efficiency and effectiveness by enabling a single text embedding model to perform both high-quality retrieval and sophisticated listwise reranking. The core idea is surprisingly elegant: it reinterprets the ‘listwise prompt’ – a combination of the original query and its candidate documents – as a ‘pseudo-relevance feedback’ (PRF) query. This enriched query provides the embedding model with additional context, allowing it to capture the fine-grained interactions typically handled by more complex rerankers.
E2RANK achieves this unification through a two-stage training process. First, a standard LLM-based decoder-only text embedding model is trained using contrastive learning to establish strong semantic representations. In the second stage, this model undergoes continued training under a multi-task learning framework. This involves not only maintaining its embedding capabilities but also incorporating a learning-to-rank objective, specifically the RankNet loss. This combination teaches the model to understand relative relevance orders within a list of documents.
What makes E2RANK particularly compelling is its efficiency. At inference time for reranking, it simply computes the cosine similarity between the PRF-enhanced query embedding and the precomputed document embeddings. This avoids the computationally intensive autoregressive decoding or full-sequence encoding required by traditional LLM-based rerankers. The result is a significant reduction in inference latency, with E2RANK demonstrating up to a 5x speedup compared to RankGPT-like listwise rerankers, even with larger models. For instance, the E2RANK-8B model can be faster than a RankQwen3-0.6B model.
The empirical results are impressive. E2RANK achieves state-of-the-art reranking performance on the BEIR benchmark and shows competitive results on the reasoning-intensive BRIGHT benchmark. Furthermore, the ranking training process also improves the model’s general embedding performance on the MTEB benchmark. This indicates that the model doesn’t sacrifice its foundational retrieval abilities while gaining advanced reranking capabilities.
The paper also highlights the potential for E2RANK to serve as a unified, end-to-end search engine. A single E2RANK model can first retrieve a set of candidate documents and then rerank them, leading to consistent performance improvements across different model sizes and datasets. This unified design simplifies system architecture and reduces overall latency, offering a practical and accurate alternative to more complex multi-stage ranking systems.
Further analysis revealed that incorporating more documents into the listwise prompt generally improves ranking performance, up to a certain point (around 20 documents), after which the benefits plateau. The listwise prompts also lead to sharper discrimination between relevant and irrelevant documents, yielding higher similarity scores for top-ranked items and a steeper decline for lower-ranked ones. The robustness of E2RANK was also demonstrated across various first-stage retrievers, consistently improving performance regardless of the initial retrieval quality.
Also Read:
- RaCoT: Enhancing LLM Reasoning Reliability with Pre-Retrieval Contrastive Thinking
- Boosting Language Model Efficiency with Encoder-Decoder Diffusion
In conclusion, E2RANK presents a powerful and efficient solution for modern search applications, unifying retrieval and listwise reranking within a single text embedding model. For more details, you can read the full research paper here.
