PaperlayerView exact revision on arXiv

Source-linked AI summary

E2Rank: Unifying Text Embedding and Listwise Reranking for Effective and Efficient Search

Qi Liu, Yanzhao Zhang, Mingxin Li, Dingkun Long, Pengjun Xie, Jiaxin Mao

arXiv:2510.22733v3cs.CLcs.AIcs.IR

TL;DR

Embedding retrieval is efficient, but LLM listwise rerankers capture fine-grained interactions at high computational cost. E2RANK turns listwise context into a PRF query embedding and reports strong benchmark performance with lower latency than RankGPT-like rerankers.

  • Problem

    LLM-based listwise rerankers capture fine-grained query–document and document–document interactions but incur high computational costs and inference latency.

    Source: §1 · p. 1

  • Method

    E2RANK treats the listwise prompt as a PRF query and trains a single embedding model to perform retrieval and listwise reranking through cosine similarity.

    Source: §1 · p. 2, §3.3 · p. 4

  • Results

    E2RANK achieves state-of-the-art reranking on BEIR, strong results on BRIGHT, lower latency than RankGPT-like rerankers, and competitive embedding performance on MTEB.

    Source: §1 · p. 2, §5 · p. 9

  • Takeaways & Limitations

    A single embedding model can serve as a unified retrieval-reranking engine and an efficient alternative to complex multi-stage ranking systems.

    Source: §5 · p. 9

    Evaluation focuses primarily on decoder-only LLM-based embedding models; encoder-only architectures remain for future validation, and prompt length grows with candidate count.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Text embedding models deliver competitive retrieval performance with high efficiency, but their ranking fidelity remains limited compared to LLM-based listwise rerankers, which capture fine-grained query-document and document-document interactions at high computational cost. We propose E2Rank (Efficient Embedding-based Ranking), a unified framework that extends a single text embedding model to perform both retrieval and listwise reranking via continued training under a listwise ranking objective. The key insight is to treat the listwise prompt---constructed from the query and its top-K candidates---as a pseudo-relevance feedback (PRF) query, enabling reranking via cosine similarity against precomputed document embeddings without autoregressive decoding. Empirically, E2Rank achieves state-of-the-art results on BEIR, competitive performance on the reasoning-intensive BRIGHT benchmark, significantly lower latency than existing LLM-based rerankers, and improved embedding performance on MTEB---all within a single model.

1 Introduction

LLM listwise rerankers capture fine-grained query–document and document–document interactions but incur high computational costs. E2RANK recasts listwise context as a PRF query embedding, unifying retrieval and reranking through similarity scoring.

1 Introduction

  • The central question is whether listwise interaction signals can be transferred into embedding-based ranking so reranking uses similarity rather than autoregressive decoding.Prior observations suggest that contextual interactions, rather than autoregressive generation itself, are key to listwise ranking effectiveness.Source: §1 · p. 1
  • E2RANK achieves state-of-the-art reranking on BEIR and strong performance on BRIGHT, improves inference efficiency, and maintains competitive embedding performance on MTEB.The paper reports these results for a single model.Source: Figure 1 · p. 2, §1 · p. 2
  • E2RANK treats the query-and-candidate listwise prompt as a PRF query, converting its context into a ranking-oriented query embedding while keeping document embeddings reusable.This unifies retrieval and reranking through cosine similarity against precomputed document embeddings.Source: §1 · p. 2
  • E2RANK is trained first with contrastive learning and then with joint contrastive and RankNet objectives using the listwise prompt as a pseudo-query.At inference, reranking uses cosine similarity between document embeddings and the prompt-augmented query embedding, avoiding autoregressive decoding.Source: §1 · p. 2

2 Related Work

Related work covers LLM-based document reranking, text embedding models, and pseudo-relevance feedback for dense retrieval. E2RANK differs by unifying embedding and listwise reranking in one embedding model.

Text Embedding Models

  • Modern embedding models use shared semantic spaces for retrieval, while GritLM unifies embedding and generation; E2RANK unifies embedding and listwise reranking.Source: §2 · p. 3

LLM-based Document Reranking

  • Listwise methods jointly consider candidate documents; E2RANK instead unifies embedding and listwise reranking within a single embedding model.The related-work discussion contrasts this with prior approaches to improving listwise reranking through prompting or output formats.Source: §2 · p. 3

Pseudo Relevance Feedback for Dense Retrieval

  • Unlike prior PRF adaptations for dense retrieval or pointwise rerankers, E2RANK interprets PRF within LLM-based listwise reranking.Classical PRF treats top-retrieved documents as relevant and uses them to refine the query.Source: §2 · p. 3

3 Methodology

E2RANK turns listwise prompt context into an embedding-based ranking signal, then trains one embedding model to support both retrieval and listwise reranking.

3 Methodology

  • The methodology frames listwise prompts as PRF queries and cosine similarity as a unified ranking function, then details training of E2RANK.Source: §3 · p. 3

3.1 Preliminary

  • The embedding model represents documents and instructed queries as vectors, while conventional listwise rerankers prompt on the candidate set and decode a document permutation.The embedding model measures query–document relevance by cosine similarity.Source: §3.1 · p. 3

3.2 Listwise Prompts as Pseudo Relevance Feedback Query

  • E2RANK embeds the listwise prompt as a PRF query and scores candidates by cosine similarity, retaining contextual interactions with reusable document embeddings.The listwise prompt supplies contextual PRF signals while retrieval and reranking use the shared embedding space.Source: §3.2 · pp. 3–4

3.3 Training the Unified Embedding and Listwise Reranking Model

  • Training first uses contrastive learning for retrieval, then combines InfoNCE and pairwise RankNet training to add listwise ranking ability while maintaining embedding capacity.LLM-generated full rankings supply pairwise relative relevance orders for the RankNet objective.Source: §3.3 · p. 4

4 Experiments

Experiments evaluate E2RANK across reranking, embedding, and end-to-end retrieval benchmarks, alongside ablations and analyses of its PRF-based ranking behavior. The results show strong BEIR and BRIGHT performance, competitive embeddings, and gains across first-stage retrievers, with benefits depending on the training and document-selection choices.

4.1 Experimental Setup

  • Experiments use Qwen3 0.6B, 4B, and 8B models, public E5 Stage-I data, and Qwen3-32B-generated ranking permutations treated as noisy supervision.Reranking covers TREC DL, eight BEIR datasets, and BRIGHT on top-100 candidates using NDCG@10.Source: §4.1 · p. 4, §4.2 · p. 5

4.2 Reranking Performance

  • 71.16 on DL20 and 54.35 BEIR average: E2RANK-8B tops both, surpassing even larger zero-shot RankGPT-4o.Against same-backbone RankQwen3, average effectiveness rises at every scale, though TREC-DL results are mixed and a few BEIR datasets regress significantly.Source: §4.2 · pp. 5–6, Table 3 · p. 6

4.3 Embedding Ability

  • E2RANK-8B is competitive with recent embedding models; ranking supervision improves its retrieval tasks by ↑1.58, with a small clustering trade-off.Gains also extend to pair classification, classification, and STS across model scales; evaluation uses query-only embeddings.Source: Table 4 · p. 7, §4.3 · p. 7

4.4 Unified and End-to-End Retrieval and Reranking

  • E2RANK improves every first-stage retriever; the unified pipeline is strongest on DL20 and BRIGHT, while Qwen3-Embedding retains a small BEIR advantage.Gains across sparse and dense retrievers are especially evident on BRIGHT for larger unified models, and scaling E2RANK strengthens end-to-end performance.Source: Table 6 · p. 7, §4.4 · p. 7

4.5 Ablation Study

  • The full training strategy is best or highly competitive across datasets; removing contrastive learning, RankNet, or listwise PRF prompts degrades results.RankNet removal causes the most severe ranking collapse, while removing listwise prompts harms performance even when RankNet remains.Source: Table 7 · p. 8, §4.5 · p. 8

4.6 Analysis

  • Neither text-based listwise prompting nor Rocchio-style vector fusion approaches E2RANK, and both can hurt ranking quality.These PRF baselines use a model without listwise training, indicating that supervision is needed to learn how to use PRF signals.Source: Table 8 · p. 8, §4.6 · p. 8
  • E2RANK improves all tested first-stage retrievers, stronger retrievers yield better PRF, and Stage-II training transfers to bidirectional GTE-Qwen2-1.5B.The transfer improves reranking while preserving general embedding performance.Source: Table 9 · p. 9, §4.6 · p. 9
  • Top-20 retrieved documents perform best; adding candidates helps below 20 and plateaus around 20, while Random-20 and Last-20 can underperform query-only.Listwise prompts also produce higher scores for top-ranked documents and a steeper decline for lower-ranked ones than query-only prompts.Source: §4.6 · pp. 8–9, Figure 2 · p. 8, Figure 3 · p. 9

5 Conclusion

E2RANK unifies efficient retrieval and listwise reranking in one text embedding model. It achieves strong reranking results across BEIR and BRIGHT, reduces latency, and maintains competitive MTEB embedding capabilities.

5 Conclusion

  • E2RANK uses one text embedding model for efficient retrieval and listwise reranking by treating the listwise prompt as a pseudo-relevance feedback query.The framework achieves state-of-the-art reranking on BEIR, strong results on BRIGHT, significantly lower latency than RankGPT-like rerankers, and competitive MTEB embedding performance.Source: §5 · p. 9

Limitations

The evaluation focuses primarily on decoder-only LLM-based embedding models, leaving broader validation on encoder-only architectures for future work. Prompt length and training-label quality also constrain the approach.

Limitations

  • The study focuses primarily on decoder-only LLM-based embedding models; broader validation on encoder-only architectures remains future work.The listwise prompt grows with the number of candidate documents, which may challenge very long documents or large candidate sets.Source: Limitations · p. 9
  • Training relies on LLM-generated ranking labels without a separate human-label study, and label quality may vary across domains.Source: Limitations · p. 10

The Use of Large Language Models

The authors used large language models for portions of coding and language polishing, and reviewed all outputs.

The Use of Large Language Models

A Training Dataset Details

The appendix describes the datasets used for both training stages and how ranking labels were generated and assessed.

A Training Dataset Details

  • Stage I uses roughly 1.5 million samples from the public E5 training data, with one positive and one negative per query.The mixture includes retrieval, QA, NLI, and duplicate-question datasets.Source: A Training Dataset Details · p. 13
  • Stage II uses an approximately 157k-sample mixture, filters for at least 15 negative examples, and includes two Chinese retrieval datasets.The data combines retrieval datasets from E5 and BGE-M3; documents are restricted to lengths below 500.Source: A Training Dataset Details · p. 13
  • Qwen3-32B generates full ranking labels from instructions in Table 10, and outputs with incorrect formats are filtered.Source: A Training Dataset Details · p. 13, Table 10 · p. 13
  • Qwen3-32B places the gold positive first only slightly above 50% of the time on MS MARCO, so its labels are treated as noisy supervision.The authors report no separate human-label study at this scale.Source: A Training Dataset Details · p. 13
  • The zero-shot teacher reaches 53.37 average NDCG@10 on the BEIR subset and 23.8 on BRIGHT, and is described as comparable to RankGPT-4o on BEIR.Higher-quality label construction is identified as future work.Source: A Training Dataset Details · p. 13

B Implementation Details

The appendix reports the two-stage training setup and evaluation instruction sensitivity, including a small, statistically nonsignificant instruction effect.

B Implementation Details

  • Stage I trains full model parameters for one epoch on eight A100 80G GPUs, using batch size 512, sequence length 512, and learning rate 2×10−5.Training uses BF16, DeepSpeed Zero3, gradient checkpointing, and 300 warm-up steps.Source: B Implementation Details · p. 14
  • Stage II samples at most 10,000 instances per dataset, yielding about 87,000 training instances, and supplies the top 20 candidates in first-stage rank order at inference.Training uses batch size 128, one epoch, and documents up to 1,024 tokens.Source: B Implementation Details · p. 14
  • Different evaluation instructions have very small, statistically nonsignificant effects on performance.The stated finding concerns the evaluation instructions used for reranking tasks.Source: B Implementation Details · p. 14

C Discussion of PRF-like Mechanism

The section discusses E2RANK as a learned PRF-like reranker and examines its ranking behavior across reranking depths and model backbones.

C Discussion of PRF-like Mechanism

  • E2RANK treats top-ranked documents as weak relevance cues and learns their weighting and interpretation through listwise supervision rather than manual PRF mechanisms.Unlike heuristic query expansion, it models query-related and document-relational signals jointly in a listwise context.Source: C Discussion of PRF-like Mechanism · pp. 14–15

D.1 Influence of the Reranking Depth

  • Across reranking depths, E2RANK matches or outperforms RankQwen3 on NDCG@5 and NDCG@10, while RankQwen3 leads on NDCG@1 in several cases.The comparison uses DL20; E2RANK is described as maintaining competitive rank-1 precision while performing more strongly at broader cutoffs.Source: §D.1 · p. 16, Figure 6 · p. 17

D.2 Reranking Behavior

  • Table 16 decomposes NDCG@100 gains into head reordering and tail promotion; both contribute substantially, and E2RANK has higher total gains than RankQwen3 at the same scale.E2RANK tends toward slightly higher within-top-20 gains, while RankQwen3 shows a more balanced pattern with comparable tail gains.Source: §D.2 · p. 18

D.3 Training From Existing Embedding Models

  • Applying Stage-II training to GTE-Qwen2-1.5B improves both reranking benchmarks while maintaining embedding performance, supporting applicability beyond the Qwen3 backbone.The model is an encoder-based embedding model with bidirectional attention; its architecture was not modified.Source: §D.3 · p. 18

D.4 Robustness to Degraded PRF Windows

The robustness experiment progressively degrades the top-20 PRF feedback window while keeping relevant candidates available in the top-100 scoring pool.

D.4 Robustness to Degraded PRF Windows

  • Reranking performance degrades gradually as the quality of the feedback window declines.The test progressively replaces relevant feedback documents with non-relevant candidates.Source: §D.4 · p. 18, Table 18 · p. 19
  • Even with no relevant document in the feedback window, reranking remains above BM25 on all three dense-relevance datasets.The result indicates that feedback quality is a limitation but not the sole source of the gains.Source: §D.4 · p. 18
  • The experiment replaces relevant documents in the top-20 PRF window with non-relevant candidates while retaining relevant documents in the top-100 scoring pool.Table 18 reports NDCG@10 for E2RANK-0.6B under this degradation.Source: §D.4 · p. 18, Table 18 · p. 19

E Additional Experimental Results

The appendix expands the BEIR comparisons and provides detailed benchmark and ablation results.

E Additional Experimental Results

  • The BEIR comparison includes a pointwise cross-encoder trained on E2RANK's data with the same RankNet loss, and the authors report that it demonstrates the effectiveness of listwise reranking.The authors suggest insufficient training data may explain why some reasoning rerankers perform poorly on these general reranking tasks.Source: E Additional Experimental Results · p. 19, Table 19 · p. 20

Full results for the ablation studies of training

Loading 2510.22733v3…