Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Reranker-guided graph search outperforms sequential reranking at the same rerank budget.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 22:44 UTC pith:CROYWAA2

load-bearing objection Solid empirical paper: RGS gives real gains under a fixed reranker budget; the mechanism isn't fully proven, but the paper is worth a serious referee. the 4 major comments →

arxiv 2509.07163 v1 pith:CROYWAA2 submitted 2025-09-08 cs.IR cs.CLcs.LG

Beyond Sequential Reranking: Reranker-Guided Search Improves Reasoning Intensive Retrieval

Haike Xu , Tong Chen This is my paper
classification cs.IR cs.CLcs.LG
keywords reranker-guided searchretrieve-and-rerankproximity graphgreedy searchreranker budgetreasoning-intensive retrievallistwise rerankingclustering hypothesis
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that the standard retrieve-and-rerank pipeline is budget-inefficient: reranking the top-k documents in embedding order spends expensive reranker calls on documents the embedding model already favors. Reranker-Guided-Search (RGS) instead uses the reranker as a navigation signal, greedily expanding a proximity graph around documents the reranker likes and reranking only those new neighborhoods. On three reasoning-intensive benchmarks, with a budget of 100 rerank calls, RGS raises NDCG@10 from 25.3 to 28.8 on BRIGHT, 60.4 to 63.3 on FollowIR, and 25.9 to 31.0 on M-BEIR compared with sequential reranking. The paper's central claim is that, given a fixed embedding and reranker, strategically choosing which documents to rerank materially changes final retrieval accuracy. If correct, this means the reranker budget, not the embedding model, is the main lever for hard retrieval tasks.

Core claim

The central discovery is that reranker preferences can serve as a navigation signal over a document-similarity graph. RGS builds a proximity graph from document embeddings with the DiskANN algorithm, seeds a frontier with the document nearest to the query, then repeatedly takes the most promising unexpanded document in the reranked frontier, appends its graph neighbors, and reranks the newly added tail with a listwise LLM reranker. This greedy expansion rests on the clustering hypothesis—documents close in embedding space tend to be relevant to the same query—so a high reranker score is treated as evidence that the document's neighbors deserve the next rerank calls. The method finds document

What carries the argument

The central mechanism is Reranker-Guided-Search: a greedy expansion on a proximity graph built from document embeddings, driven by a listwise reranker. The reranker maintains an ordered frontier of candidate documents; each step expands the first unexpanded document's graph neighbors into the frontier and reranks the appended tail in a sliding window. The proximity graph is built by the DiskANN approximate-nearest-neighbor algorithm, and the listwise reranker is an LLM-based ranker. The load-bearing assumption is the clustering hypothesis: documents similar in embedding space tend to have similar relevance to a given query, which justifies spending rerank budget on the neighborhood of a rera

Load-bearing premise

The load-bearing premise is that documents close in embedding space are likely relevant to the same query, so a reranker's approval of one document is a trustworthy signal to spend the next rerank calls on its graph neighbors; if relevant documents are instead isolated islands in the embedding graph, RGS squanders its budget and the measured gains disappear.

What would settle it

Construct or identify a retrieval benchmark where each ground-truth document is a distant outlier in the embedding graph—no two relevant documents share an edge, and the query's nearest neighbors are all distractor topics. Run RGS, sequential reranking, and SlideGAR at a budget of 100; if sequential reranking reaches higher NDCG@10 or RGS fails to beat its own random-start baseline, the clustering hypothesis fails and the paper's core mechanism is falsified. A second, cheaper check: measure the clustering coefficient of ground-truth documents in the embedding graph across existing benchmarks;

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • Retrieval accuracy with a fixed rerank budget is not determined by the initial embedding retrieval: RGS beats sequential top-k reranking on BRIGHT, FollowIR, and M-BEIR at budgets 100, 300, and 500.
  • The embedding model's capacity mainly controls how fast accuracy grows with rerank budget; across weak to strong embeddings, RGS converges to similar final accuracy at Reranker@500, placing the accuracy ceiling on reranker quality.
  • Because the query embedding only supplies the starting point, RGS remains accurate even when the query embedding is corrupted or replaced by a fixed start point, and can run without vector similarity search entirely.
  • The Reranker@k evaluation setup—reporting NDCG@10 as a function of number of rerank calls—offers a direct comparison of reranker-budget efficiency that the paper argues future retrieval evaluation should adopt.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the clustering hypothesis holds in general, RGS's gains should transfer to pointwise rerankers (replacing the frontier list with a priority queue); a cheap pointwise reranker paired with graph-guided expansion might approach listwise reranker quality at a fraction of the cost—an extension the paper leaves untested.
  • The same logic implies a measurable predictor of RGS's benefit: datasets where ground-truth documents form tight clusters in the embedding graph should show larger gains; dataset designers could use this clustering coefficient as a warning that sequential reranking will underperform.
  • The paper's error analysis shows a growing class of 'seen but not selected' documents as the budget grows, meaning the reranker's own preference boundary, not the search, is what caps accuracy. A practical consequence is that tuning or fine-tuning the reranker on the target task should raise the ceiling for any guided-search strategy.
  • Since RGS discards the query embedding, it could be extended to federated or privacy-sensitive settings where the query must not be embedded into a shared space, though the greedy traversal would still need a graph index.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes Reranker-Guided-Search (RGS), a retrieval method that replaces sequential top-k reranking with a reranker-guided greedy search over a DiskANN proximity graph built from document embeddings. Given a fixed reranker budget, RGS initializes from vector-similar seed documents, then iteratively expands the graph neighborhood of the documents most favored by the reranker, using a listwise reranker with a sliding window to reorder candidates. The method is evaluated on three benchmarks (BRIGHT, FollowIR, M-BEIR) against retrieve-and-rerank and SlideGAR, reporting NDCG@10 gains under reranker budgets of 100/300/500, e.g., 25.3→28.8 on BRIGHT and 25.9→31.0 on M-BEIR at budget 100. The paper also includes ablations on embedding quality, query/document perturbation, start-point selection, and graph type, plus an error analysis showing that many ground-truth documents are never examined.

Significance. If the claimed gains are robust, the paper makes a useful practical contribution: it shows that under a fixed LLM-reranker budget, accuracy can be improved by replacing the standard sequential scan with a graph-guided exploration, and it introduces a Reranker@k evaluation setup that is better aligned with the cost of LLM-based rerankers. The method is simple, code is provided, and the ablations (query-embedding perturbation, graph type, start-point dependence) are thoughtful and informative. The central comparison is internally consistent and the gains are large and repeat across three benchmarks, which strengthens the case that the effect is real. However, the absence of statistical significance testing, the tuning of Ls on the test benchmarks, and the unverified clustering mechanism leave the generality of the headline claim not fully established.

major comments (4)
  1. [Section 4.1, Tables 1–3] No error bars or significance tests are reported. M-BEIR is evaluated on a random sample of 100 queries, and the claimed 5.1-point gain at Reranker@100 could be within sampling noise. Please report variance (e.g., bootstrap or multiple random subsets), or at least query-level standard errors, especially for M-BEIR and for the per-dataset comparisons in Tables 1–3.
  2. [Section 4.1, Algorithm 1] The search list size Ls is tuned per reranker budget on the same benchmarks where accuracy is measured (Ls=20/30/50 for k=100/300/500), and the number of start points (k/5) is also chosen per budget. No validation split or sensitivity analysis is described. This is a tuning-to-test-set concern that can inflate the reported gains. Please show that the improvements hold across a reasonable range of Ls, or select Ls on a held-out set and report the resulting accuracy.
  3. [Section 3.2, Section 5.4, Figure 6] The clustering hypothesis is load-bearing: RGS can only find relevant documents that are reachable from the initial seeds through short paths in the DiskANN graph. The ablations (perturbed document embeddings, random graphs) show that document-wise similarity is necessary, but they do not establish that the specific DiskANN graph actually connects seeds to relevant documents in the three benchmarks. Figure 6 shows that a large fraction of ground-truth documents are never examined, but it does not distinguish graph unreachability from reranker misalignment. Please measure the graph-distance distribution between the seeds and the relevant documents, or otherwise directly test whether the relevant documents are within the explored neighborhoods; without this, the mechanism underlying the headline gains remains unverified.
  4. [Section 4.1, Tables 1–3] The comparison with SlideGAR uses different sliding window sizes (w=10 for RGS, w=20 for SlideGAR) and a reimplementation of the baseline. This is not necessarily unfair, but it is a confound. Please report a sensitivity analysis of RGS with w=20, and if possible confirm that the reimplemented SlideGAR matches the original authors' reported behavior on a common setting, so that the gain over the graph-based baseline is not an artifact of implementation details.
minor comments (6)
  1. [Section 4.1] The definition of 'Reranker@k' is ambiguous: the text says the reranker scans at most k documents, but listwise rerankers may see the same document multiple times. The token/API-cost analysis in Appendix A partially addresses this, but the main text should explicitly state whether k is unique documents or total reranker calls.
  2. [Algorithm 1] The loop 'for i=|A| to 0 step size −w/2' is underspecified when |A| < w, and the window boundaries are not precise. Please clarify the iteration range and the handling of the initial small-list case.
  3. [Section 5.2] The sentence 'we investigate the role of each by introducing different levels of perturbation to either the query or document embeddings' refers to 'RR, SlideGAR, and A', but 'A' should be 'RGS'.
  4. [Figure 3] The two panels are labeled 'Document' and 'Query' but the legend/placement is confusing; it is hard to tell which set of curves corresponds to which perturbation type. Please label the panels explicitly.
  5. [Section 5.4] There is a typo: 'docuemnt-wise similarity' should be 'document-wise similarity'.
  6. [Section 5.1] Minor wording: 'Biology, Psychology, Sustainable Livings, and TheoremQA-T—from the BRIGHT benchmark' contains an unusual plural 'Livings'; should be 'Sustainable Living' or 'Sustainable Living tasks'.

Circularity Check

0 steps flagged

No significant circularity; the claimed improvements are empirical and not constructionally equivalent to the method's inputs.

full rationale

The paper is an empirical systems paper rather than a derivation. Its central claim — that under a fixed reranker budget, graph-guided selection of documents to rerank improves retrieval accuracy relative to sequential scanning — is established by direct benchmark measurements (Tables 1–3), not derived from an equation. The clustering hypothesis in Section 3.2 is explicitly assumed and then probed by ablations: document-embedding perturbation (Section 5.2), removal of the first-stage query-dependent initialization (Section 5.3), and graph-type comparison (Section 5.4). The self-citations to [6] and [27] supply background and technique (worst-case ANN caveats and the bi-metric search idea) but are not the load-bearing justification for the headline results, which are externally benchmarked against RR and SlideGAR. The only protocol caveat is that the search-list size Ls is "tuned based on the reranker budget" on the same benchmarks where accuracy is reported (Section 4.1). This is a possible selection-bias concern, but it is not a definitional circularity: Ls does not by construction determine NDCG@10, and the baselines are evaluated under the same reranker budget. No equation in the paper reduces the reported result to an input, and no fitted parameter is renamed as a prediction. Therefore no circularity step meets the evidentiary bar, and the mild self-citation and tuning issues are not load-bearing.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The method rests on three unverified premises: the clustering hypothesis, the fidelity of the DiskANN graph, and the local informativeness of the reranker. The only tuned numbers are hyperparameters (Ls, w, start count), none of which are fitted directly to the target metric, but they are tuned per budget on the same benchmarks where accuracy is reported.

free parameters (3)
  • Search list size Ls = 20/30/50 for budgets 100/300/500
    Tuned per reranker budget in Section 4.1; affects the truncation of list A and hence how many documents the greedy search can keep before reranking.
  • Sliding window size w = 10 for RGS, 20 for SlideGAR
    Window width for listwise reranking; chosen from prior work [22], not fitted, but affects all methods asymmetrically.
  • Number of start points = k/5 (20, 60, 100)
    Initial seeds from first-stage vector search; chosen by the authors; sensitivity is not reported.
axioms (3)
  • domain assumption Clustering hypothesis: documents similar in embedding space have similar relevance to the same query
    Invoked in Section 3.2 to justify expanding the neighborhood of reranker-favored documents; if false the greedy expansion misdirects the budget.
  • domain assumption DiskANN proximity graph preserves the local structure of document-wise similarity with enough fidelity that neighbors of a point are truly similar documents
    The paper builds the graph with DiskANN [8] and relies on its heuristic quality; the authors themselves note in [6] that such graphs lack worst-case guarantees.
  • domain assumption The LLM listwise reranker yields scores that are locally consistent, meaning rankings within a sliding window are informative about true relevance
    The algorithm uses the reranker to decide which documents to expand next; Section 5.5 concedes the reranker's preferences do not always align with ground truth, which caps accuracy.

pith-pipeline@v1.3.0-alltime-deepseek · 11797 in / 11812 out tokens · 115531 ms · 2026-08-04T22:44:18.261321+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Beyond Sequential Reranking: Reranker-Guided Search Improves Reasoning Intensive Retrieval." pith.science (2026). https://pith.science/paper/CROYWAA2

@misc{pith2026250907163,
  author       = {Pith},
  title        = {Pith review of: Beyond Sequential Reranking: Reranker-Guided Search Improves Reasoning Intensive Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CROYWAA2}},
  note         = {Machine review of arXiv:2509.07163}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The widely used retrieve-and-rerank pipeline faces two critical limitations: they are constrained by the initial retrieval quality of the top-k documents, and the growing computational demands of LLM-based rerankers restrict the number of documents that can be effectively processed. We introduce Reranker-Guided-Search (RGS), a novel approach that bypasses these limitations by directly retrieving documents according to reranker preferences rather than following the traditional sequential reranking method. Our method uses a greedy search on proximity graphs generated by approximate nearest neighbor algorithms, strategically prioritizing promising documents for reranking based on document similarity. Experimental results demonstrate substantial performance improvements across multiple benchmarks: 3.5 points on BRIGHT, 2.9 on FollowIR, and 5.1 on M-BEIR, all within a constrained reranker budget of 100 documents. Our analysis suggests that, given a fixed pair of embedding and reranker models, strategically selecting documents to rerank can significantly improve retrieval accuracy under limited reranker budget.

Figures

Figures reproduced from arXiv: 2509.07163 by Haike Xu, Tong Chen.

Figure 1
Figure 1. Figure 1: An example from TheoremQA-T illustrating how our [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Average NDCG@10 result over a selected subset of datasets from BRIGHT. The left fig￾ure is our method RGS, the right figure is retrieve-and-rerank. Ex￾periments involve 4 different em￾bedding models with varying em￾bedding quality. 5.2 Different Roles of Document-Document Similarity and Query-Document Relevance in the Retrieval Process We perform a fine-grained ablation study to examine how embedding quali… view at source ↗
Figure 3
Figure 3. Figure 3: For different levels of perturbation added on query (right) / document(left) embed￾dings, we report the retrieval accuracy for different retrieval methods Perturbation on document embeddings Fix a ratio w ∈ [0, 1], we mix each document embedding with another random document embedding according to the ratio 1 − w : w. This simulates the scenario where the embedding model fails to distinguish subtle differen… view at source ↗
Figure 5
Figure 5. Figure 5: We initialize RGS on the 50 documents closest to the query and observe how quickly the greedy search approaches the true answer. RGS still manages to improve on KNN graphs but the process is much slower, and it shows no improvement on random graphs. We hypothesize that in KNN graphs, edges exist only between nearby documents, while connections between distant documents are missing, preventing the search fr… view at source ↗
Figure 5
Figure 5. Figure 5: Impact of running our second-stage search on different graph data structures for re￾trieval accuracy. DiskANN: proximity graph gen￾erated by the DiskANN algorithm. KNN: each document is connected to its k nearest neighbors in the embedding space. Random: a randomly connected graph with degree k. 100 200 300 400 500 43.3% 15.0% 41.8% 48.6% 26.0% 25.5% 48.1% 30.5% 21.4% 52.2% 34.3% 13.5% 53.8% 34.4% 11.8% Re… view at source ↗
Figure 6
Figure 6. Figure 6: For different Reranker budgets, each stacked bar shows the percentage of ground-truth passages that were re￾trieved (orange), examined but not se￾lected (light orange), or never examined (grey). Results are reported for running RGS (left) and RR (right) methods on BRIGHT/Biology dataset. search on KNN graphs. For random graphs, it is expected that RGS shows no improvement because the search is equivalent t… view at source ↗
Figure 7
Figure 7. Figure 7: Average NDCG@10 versus number of tokens or API calls sent to the LLM-based reranker [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. A Survey of Reasoning-Intensive Retrieval: Progress and Challenges

    cs.IR 2026-04 unverdicted novelty 6.0

    A survey that categorizes RIR benchmarks by domain and modality, proposes a taxonomy for integrating reasoning into retrieval pipelines, and outlines key challenges.

Reference graph

Works this paper leans on

29 extracted references · 12 canonical work pages · cited by 1 Pith paper · 1 internal anchor

  1. [1]

    Overview of the trec 2019 deep learning track.arXiv preprint arXiv:2003.07820, 2020

    Nick Craswell, Bhaskar Mitra, Emine Yilmaz, Daniel Campos, and Ellen M V oorhees. Overview of the trec 2019 deep learning track.arXiv preprint arXiv:2003.07820, 2020

    Pith/arXiv arXiv 2019

  2. [2]

    Splade: Sparse lexical and expansion model for first stage ranking

    Thibault Formal, Benjamin Piwowarski, and Stéphane Clinchant. Splade: Sparse lexical and expansion model for first stage ranking. InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2288–2292, 2021

    work page 2021

  3. [3]

    Effective adhoc retrieval through traversal of a query-document graph

    Erlend Frayling, Sean MacAvaney, Craig Macdonald, and Iadh Ounis. Effective adhoc retrieval through traversal of a query-document graph. InEuropean Conference on Information Retrieval, pages 89–104. Springer, 2024

    work page 2024

  4. [4]

    Nsg : Navigating spread-out graph for approximate nearest neighbor search.https://github.com/ZJULearning/nsg, 2019

    Cong Fu, Chao Xiang, Changxu Wang, and Deng Cai. Nsg : Navigating spread-out graph for approximate nearest neighbor search.https://github.com/ZJULearning/nsg, 2019

    work page 2019

  5. [5]

    Simcse: Simple contrastive learning of sentence embeddings.arXiv preprint arXiv:2104.08821, 2021

    Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings.arXiv preprint arXiv:2104.08821, 2021

    Pith/arXiv arXiv 2021

  6. [6]

    Piotr Indyk and Haike Xu. Worst-case performance of popular approximate nearest neigh- bor search implementations: Guarantees and limitations.Advances in Neural Information Processing Systems, 36:66239–66256, 2023

    2023

  7. [7]

    The use of hierarchic clustering in information retrieval.Information storage and retrieval, 7(5):217–240, 1971

    Nick Jardine and Cornelis Joost van Rijsbergen. The use of hierarchic clustering in information retrieval.Information storage and retrieval, 7(5):217–240, 1971

    work page 1971

  8. [8]

    Diskann: Fast accurate billion-point nearest neighbor search on a single node.Advances in neural information processing Systems, 32, 2019

    Suhas Jayaram Subramanya, Fnu Devvrit, Harsha Vardhan Simhadri, Ravishankar Krishnawamy, and Rohan Kadekodi. Diskann: Fast accurate billion-point nearest neighbor search on a single node.Advances in neural information processing Systems, 32, 2019

    2019

  9. [9]

    Diskann.https://github.com/microsoft/DiskANN, 2023

    Suhas Jayaram Subramanya, Fnu Devvrit, Harsha Vardhan Simhadri, Ravishankar Krishnawamy, and Rohan Kadekodi. Diskann.https://github.com/microsoft/DiskANN, 2023

    work page 2023

  10. [10]

    Dense passage retrieval for open-domain question answering

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick SH Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. Dense passage retrieval for open-domain question answering. InEMNLP (1), pages 6769–6781, 2020

    2020

  11. [11]

    Adaptive re-ranking with a corpus graph

    Sean MacAvaney, Nicola Tonellotto, and Craig Macdonald. Adaptive re-ranking with a corpus graph. InProceedings of the 31st ACM International Conference on Information & Knowledge Management, pages 1491–1500, 2022

    work page 2022

  12. [12]

    Yu A Malkov and Dmitry A Yashunin. Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs.IEEE transactions on pattern analysis and machine intelligence, 42(4):824–836, 2018

    2018

  13. [13]

    Sfr-embedding-mistral:enhance text retrieval with transfer learning

    Rui Meng, Ye Liu, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou, and Semih Yavuz. Sfr-embedding-mistral:enhance text retrieval with transfer learning. Salesforce AI Research Blog, 2024

    work page 2024

  14. [14]

    Passage re-ranking with bert.arXiv preprint arXiv:1901.04085, 2019

    Rodrigo Nogueira and Kyunghyun Cho. Passage re-ranking with bert.arXiv preprint arXiv:1901.04085, 2019

    Pith/arXiv arXiv 1901

  15. [15]

    Instructir: A benchmark for instruction following of information retrieval models

    Hanseok Oh, Hyunji Lee, Seonghyeon Ye, Haebin Shin, Hansol Jang, Changwook Jun, and Minjoon Seo. Instructir: A benchmark for instruction following of information retrieval models. arXiv preprint arXiv:2402.14334, 2024

    Pith/arXiv arXiv 2024

  16. [16]

    Rankvicuna: Zero-shot listwise document reranking with open-source large language models.arXiv preprint arXiv:2309.15088, 2023

    Ronak Pradeep, Sahel Sharifymoghaddam, and Jimmy Lin. Rankvicuna: Zero-shot listwise document reranking with open-source large language models.arXiv preprint arXiv:2309.15088, 2023

    Pith/arXiv arXiv 2023

  17. [17]

    Large language models are effective text rankers with pairwise ranking prompting.arXiv preprint arXiv:2306.17563, 2023

    Zhen Qin, Rolf Jagerman, Kai Hui, Honglei Zhuang, Junru Wu, Le Yan, Jiaming Shen, Tianqi Liu, Jialu Liu, Donald Metzler, et al. Large language models are effective text rankers with pairwise ranking prompting.arXiv preprint arXiv:2306.17563, 2023. 10

    Pith/arXiv arXiv 2023

  18. [18]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. InInternational conference on machine learning, pages 8748–8763. PmLR, 2021

    2021

  19. [19]

    Guiding retrieval using llm-based listwise rankers

    Mandeep Rathee, Sean MacAvaney, and Avishek Anand. Guiding retrieval using llm-based listwise rankers. InEuropean Conference on Information Retrieval, pages 230–246. Springer, 2025

    work page 2025

  20. [20]

    Reasonir: Training retrievers for reasoning tasks.arXiv preprint arXiv:2504.20595, 2025

    Rulin Shao, Rui Qiao, Varsha Kishore, Niklas Muennighoff, Xi Victoria Lin, Daniela Rus, Bryan Kian Hsiang Low, Sewon Min, Wen-tau Yih, Pang Wei Koh, et al. Reasonir: Training retrievers for reasoning tasks.arXiv preprint arXiv:2504.20595, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    Bright: A realistic and challenging benchmark for reasoning-intensive retrieval.arXiv preprint arXiv:2407.12883, 2024

    Hongjin Su, Howard Yen, Mengzhou Xia, Weijia Shi, Niklas Muennighoff, Han-yu Wang, Haisu Liu, Quan Shi, Zachary S Siegel, Michael Tang, et al. Bright: A realistic and challenging benchmark for reasoning-intensive retrieval.arXiv preprint arXiv:2407.12883, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Is chatgpt good at search? investigating large language models as re-ranking agents.arXiv preprint arXiv:2304.09542, 2023

    Weiwei Sun, Lingyong Yan, Xinyu Ma, Shuaiqiang Wang, Pengjie Ren, Zhumin Chen, Dawei Yin, and Zhaochun Ren. Is chatgpt good at search? investigating large language models as re-ranking agents.arXiv preprint arXiv:2304.09542, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    Uniir: Training and benchmarking universal multimodal information retrievers

    Cong Wei, Yang Chen, Haonan Chen, Hexiang Hu, Ge Zhang, Jie Fu, Alan Ritter, and Wenhu Chen. Uniir: Training and benchmarking universal multimodal information retrievers. In European Conference on Computer Vision, pages 387–404. Springer, 2024

    work page 2024

  24. [24]

    Followir: Evaluating and teaching information retrieval models to follow instructions.arXiv preprint arXiv:2403.15246, 2024

    Orion Weller, Benjamin Chang, Sean MacAvaney, Kyle Lo, Arman Cohan, Benjamin Van Durme, Dawn Lawrie, and Luca Soldaini. Followir: Evaluating and teaching information retrieval models to follow instructions.arXiv preprint arXiv:2403.15246, 2024

    Pith/arXiv arXiv 2024

  25. [25]

    Rank1: Test-time compute for reranking in information retrieval.arXiv preprint arXiv:2502.18418, 2025

    Orion Weller, Kathryn Ricci, Eugene Yang, Andrew Yates, Dawn Lawrie, and Benjamin Van Durme. Rank1: Test-time compute for reranking in information retrieval.arXiv preprint arXiv:2502.18418, 2025

    Pith/arXiv arXiv 2025

  26. [26]

    C-pack: Packaged resources to advance general chinese embedding, 2023

    Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. C-pack: Packaged resources to advance general chinese embedding, 2023

    2023

  27. [27]

    A Bi-metric Framework for Fast Similarity Search

    Haike Xu, Sandeep Silwal, and Piotr Indyk. A bi-metric framework for fast similarity search. arXiv preprint arXiv:2406.02891, 2024. 11 A Experimental Details in Section 4 Prompt for Gemini-2.0-FlashHere, we use a similar prompt from [ 22] to ask Gemini-2.0-Flash to rerank the documents. System Instruction You are RankGPT, an intelligent assistant that can...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  28. [28]

    Rank the 10 passages above based on their relevance to the query

    {Passage 10} Query: {query}. Rank the 10 passages above based on their relevance to the query. The passages should be listed in descending order using identifiers. The most relevant passages should be listed first. The output format should be like

    work page

  29. [29]

    Gemini-2.0-Flash

    >[2] ... >[10]. Only response the ranking results, do not say any word or explain. Table 4: Prompt for “Gemini-2.0-Flash” to serve as a reranker Computational resource to reproduce our experimentsOur experiments are run on an Intel(R) Xeon(R) Platinum 8481C CPU with 44 cores and we use one NVDIA A100 GPU to generate the embeddings. The estimated cost spen...

    work page