pith. machine review for the scientific record. sign in

arxiv: 2602.22591 · v2 · submitted 2026-02-26 · 💻 cs.IR

Recognition: 2 theorem links

· Lean Theorem

Where Relevance Emerges: A Layer-Wise Study of Internal Attention for Zero-Shot Re-Ranking

Haodong Chen , Shengyao Zhuang , Zheng Yao , Guido Zuccon , Teerapong Leelanupab
Authors on Pith no claims yet

Pith reviewed 2026-05-15 19:23 UTC · model grok-4.3

classification 💻 cs.IR
keywords zero-shot re-rankinginternal attentiontransformer layersLLM re-rankingin-context learningdocument rankingattention signalsefficiency optimization
0
0 comments X

The pith

Selective extraction of internal attention from peak layers allows zero-shot LLMs to match larger reinforced re-rankers.

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

This paper investigates internal attention signals within large language models for zero-shot document re-ranking as an alternative to generative scoring. It discovers that relevance signals form a bell-curve distribution across transformer layers, consistent across architectures. The Selective-ICR method uses only the high-signal layers to extract ranking information directly, avoiding text generation. This yields 30-50% lower latency and strong performance on the BRIGHT benchmark, where an 8B zero-shot model equals 14B RL-trained re-rankers and a 0.6B model beats generation methods. Readers should care as it points to simpler, smaller models sufficing for effective re-ranking.

Core claim

The central discovery is a universal bell-curve distribution of relevance signals across the layers of transformer models when performing in-context re-ranking. By selectively attending to the peak layers identified in this distribution, the proposed Selective-ICR extracts high-quality signals for ranking. This results in reduced inference latency of 30-50% with no effectiveness loss, and demonstrates that an 8B parameter zero-shot model can match the performance of 14B models trained via reinforcement learning on reasoning-intensive tasks.

What carries the argument

The bell-curve distribution of relevance signals across transformer layers, which enables the Selective-ICR strategy to pick only peak layers for efficient signal extraction.

Load-bearing premise

The relevance signal distribution follows a bell-curve pattern that is the same for all transformer architectures and does not require per-task tuning to select the right layers.

What would settle it

Running the method on a new model architecture where the peak layers do not yield better results than averaging all layers would show the assumption does not hold.

Figures

Figures reproduced from arXiv: 2602.22591 by Guido Zuccon, Haodong Chen, Shengyao Zhuang, Teerapong Leelanupab, Zheng Yao.

Figure 1
Figure 1. Figure 1: Layer-wise performance analysis (nDCG@10) on TREC-DL 2019 and 2020. Stars (★) indicate the peak per￾formance achieved by a single layer. Curly brackets {min, max} denote the performance range across all layers. Dashed horizontal lines and values in parentheses (·) represent the nDCG@10 obtained from the full-layer aggregation strat￾egy [2]. Square brackets [𝑙1−𝑙2] indicate selected layers for the aggregati… view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of Selective-ICR with Center-Biased Interval Selection, exemplified by Llama 3.1 8B with 32-layer [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Efficiency gains of the Selective-ICR strategy, mea [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Detailed layer-wise nDCG@10 trends across diverse BEIR tasks. Each sub-figure represents a specific model’s perfor￾mance trajectory across multiple datasets. Despite the varying semantic domains (e.g., medical, financial, or general knowledge), each model exhibits a highly consistent bell-shaped relevance distribution, peaking within the identified middle-layer intervals. controlled decoupling study using … view at source ↗
Figure 5
Figure 5. Figure 5: Overall distribution of nDCG@10 scores across the 12 sub-domains of the reasoning-intensive BRIGHT benchmark for Qwen3 variants. The plots reveal a consistent “bell curve” trend across scientific/mathematical domains, with peak performance concentrated in intermediate layers, while the programming syntactic primitive retrieval task (Pony) exhibits a distinct early￾layer signal concentration. jointly select… view at source ↗
read the original abstract

Zero-shot document re-ranking with Large Language Models (LLMs) has evolved from Pointwise methods to Listwise and Setwise approaches that optimize computational efficiency. Despite their success, these methods predominantly rely on generative scoring or output logits, which face bottlenecks in inference latency and result consistency. In-Context Re-ranking (ICR) has recently been proposed as an O(1) alternative method. ICR extracts internal attention signals directly, avoiding the overhead of text generation. However, existing ICR methods simply aggregate signals across all layers; layer-wise contributions and their consistency across architectures have been left unexplored. Furthermore, no unified study has compared internal attention with traditional generative and likelihood-based mechanisms across diverse ranking frameworks under consistent conditions. In this paper, we conduct an orthogonal evaluation of generation, likelihood, and internal attention mechanisms across multiple ranking frameworks. We further identify a universal "bell-curve" distribution of relevance signals across transformer layers, which motivates the proposed Selective-ICR strategy that reduces inference latency by 30%-50% without compromising effectiveness. Finally, evaluation on the reasoning-intensive BRIGHT benchmark shows that precisely capturing high-quality in-context attention signals fundamentally reduces the need for model scaling and reinforcement learning: a zero-shot 8B model matches the performance of 14B reinforcement-learned re-rankers, while even a 0.6B model outperforms state-of-the-art generation-based approaches. These findings redefine the efficiency-effectiveness frontier for LLM-based re-ranking and highlight the latent potential of internal signals for complex reasoning ranking tasks. Our code and results are publicly available at https://github.com/ielab/Selective-ICR.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper conducts a layer-wise analysis of internal attention signals in LLMs for zero-shot document re-ranking. It identifies a consistent 'bell-curve' distribution of relevance signals across transformer layers, proposes Selective-ICR to extract signals only from peak layers (reducing latency 30-50%), and reports that on the BRIGHT benchmark a zero-shot 8B model using this approach matches 14B RL-trained re-rankers while a 0.6B model outperforms generation-based SOTA methods. Code is released publicly.

Significance. If the bell-curve shape and peak-layer selection prove stable without test-set leakage or per-dataset fitting, the work would meaningfully shift the efficiency frontier for LLM re-ranking by showing that targeted internal signals can substitute for generation and large-scale RL. Public code strengthens reproducibility.

major comments (3)
  1. [Abstract] Abstract: The claim that 'precisely capturing high-quality in-context attention signals fundamentally reduces the need for model scaling and reinforcement learning' rests on Selective-ICR performance; the manuscript must explicitly state whether the layer indices for each model size were determined solely from training/validation splits or by inspecting attention statistics on the BRIGHT test split itself.
  2. [§4 and §5] §4 (Experimental Setup) and §5 (Results): The universality of the bell-curve distribution is asserted but no cross-architecture ablation (e.g., on non-Llama or non-Transformer variants) or frozen-selection protocol is described; without evidence that peak locations remain stable when the selection rule is fixed in advance, the zero-shot framing and latency claims are at risk.
  3. [§5.2] §5.2 (BRIGHT results): The reported equivalence between the 8B zero-shot Selective-ICR and 14B RL re-rankers requires statistical significance tests and full disclosure of data splits; if layer selection was tuned on the same test instances used for final evaluation, the comparison is invalid.
minor comments (2)
  1. Add a table or figure explicitly listing the selected layer indices for each model size and dataset to allow direct replication.
  2. [§3] Clarify the precise aggregation formula for 'internal attention signals' (e.g., mean, max, or weighted) in the methods section rather than leaving it implicit.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the thoughtful and detailed comments, which help clarify key aspects of our experimental protocol and strengthen the zero-shot claims. We address each major point below and will incorporate revisions accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The claim that 'precisely capturing high-quality in-context attention signals fundamentally reduces the need for model scaling and reinforcement learning' rests on Selective-ICR performance; the manuscript must explicitly state whether the layer indices for each model size were determined solely from training/validation splits or by inspecting attention statistics on the BRIGHT test split itself.

    Authors: We agree this clarification is essential for the zero-shot framing. Layer indices were selected by inspecting attention statistics exclusively on the official BRIGHT validation split for each model size; the test split was never used for any selection or tuning. We will explicitly state this protocol in the revised abstract and add a dedicated paragraph in §4 (Experimental Setup) describing the split usage and confirming no test leakage. revision: yes

  2. Referee: [§4 and §5] §4 (Experimental Setup) and §5 (Results): The universality of the bell-curve distribution is asserted but no cross-architecture ablation (e.g., on non-Llama or non-Transformer variants) or frozen-selection protocol is described; without evidence that peak locations remain stable when the selection rule is fixed in advance, the zero-shot framing and latency claims are at risk.

    Authors: Our study is limited to Llama-family models due to computational resources, but the bell-curve pattern held consistently across 0.6B–8B scales. We will add a new subsection in the revision describing a frozen-selection protocol: peak layers identified on the 0.6B model are fixed and applied without re-tuning to the 8B model, preserving the reported latency gains and effectiveness. While a full cross-architecture study (e.g., Mistral or non-Transformer) is beyond the current scope, this frozen protocol directly addresses stability of the selection rule. revision: partial

  3. Referee: [§5.2] §5.2 (BRIGHT results): The reported equivalence between the 8B zero-shot Selective-ICR and 14B RL re-rankers requires statistical significance tests and full disclosure of data splits; if layer selection was tuned on the same test instances used for final evaluation, the comparison is invalid.

    Authors: We will add statistical significance tests (paired t-test and Wilcoxon signed-rank) with p-values to §5.2 comparing all methods. We will also expand §4 to fully disclose the BRIGHT official splits and reiterate that layer selection used only the validation portion. With these additions the equivalence claim remains valid under the zero-shot protocol; no test instances influenced layer choice. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical observation of layer-wise attention patterns drives Selective-ICR without self-referential derivation

full rationale

The paper's core contribution is an empirical layer-wise analysis of internal attention signals on benchmarks including BRIGHT, identifying a bell-curve distribution that motivates Selective-ICR. This is a data-driven measurement and strategy proposal, not a derivation that reduces by construction to fitted parameters, self-definitions, or self-citation chains. Performance comparisons (e.g., 8B zero-shot matching 14B RL) are presented as experimental outcomes rather than tautological predictions. No equations or steps in the abstract or described chain exhibit the enumerated circularity patterns; the universality claim is an observation open to cross-validation, not a load-bearing self-reference.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claims rest on the empirical observation of attention patterns rather than new mathematical axioms or invented entities; the main unstated premise is that internal attention values reliably encode relevance without additional training.

axioms (1)
  • domain assumption Transformer attention layers encode task-relevant information that can be read out directly without generation.
    Invoked when the paper treats raw attention signals as sufficient for zero-shot re-ranking.

pith-pipeline@v0.9.0 · 5614 in / 1267 out tokens · 17957 ms · 2026-05-15T19:23:13.924763+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages · 7 internal anchors

  1. [1]

    Haodong Chen, Guido Zuccon, and Teerapong Leelanupab. 2025. Beyond GeneGPT: A Multi-Agent Architecture with Open-Source LLMs for Enhanced Ge- nomic Question Answering. InProceedings of the 2025 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region (SIGIR-AP ’25). Xi’an, China, 143–152. do...

    work page doi:10.1145/3767695.3769488 2025

  2. [2]

    Shijie Chen, Bernal Jiménez Gutiérrez, and Yu Su. 2025. Attention in Large Language Models Yields Efficient Zero-Shot Re-Rankers. InProceedings of the Thirteenth International Conference on Learning Representations (ICLR ’25). Singa- pore. doi:10.48550/arXiv.2410.02642

    work page doi:10.48550/arxiv.2410.02642 2025

  3. [3]

    Nick Craswell, Bhaskar Mitra, Emine Yilmaz, and Daniel Campos. 2021. Overview of the TREC 2020 Deep Learning Track.arXiv preprint arXiv:2102.07662(2021). doi:10.48550/arXiv.2102.07662

    work page doi:10.48550/arxiv.2102.07662 2021

  4. [4]

    Voorhees

    Nick Craswell, Bhaskar Mitra, Emine Yilmaz, Daniel Campos, and Ellen M. Voorhees. 2020. Overview of the TREC 2019 Deep Learning Track.arXiv preprint arXiv:2003.07820(2020). doi:10.48550/arXiv.2003.07820

    work page doi:10.48550/arxiv.2003.07820 2020

  5. [5]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, et al

    work page

  6. [6]

    The Llama 3 Herd of Models

    The Llama 3 Herd of Models.arXiv preprint arXiv:2407.21783(2024). doi:10.48550/arXiv.2407.21783

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2407.21783 2024

  7. [7]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Singh Chaplot, Devendra, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guil- laume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. 2023. Mistral 7B.arXiv preprint...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2310.06825 2023

  8. [8]

    Vedang Lad, Jin Hwa Lee, Wes Gurnee, and Max Tegmark. 2025. Remarkable Robustness of LLMs: Stages of Inference?. InProceedings of the 39th Annual Conference on Neural Information Processing Systems (NIPS ’25). San Diego, USA. https://openreview.net/forum?id=Wxh5Xz7NpJ Where Relevance Emerges: A Layer-Wise Study of Internal Attention for Zero-Shot Re-Ranki...

    work page 2025

  9. [9]

    Percy Liang, Rishi Bommasani, Tony Lee, Dimitris Tsipras, Dilara Soylu, Michi- hiro Yasunaga, Yian Zhang, Deepak Narayanan, Yuhuai Wu, Ananya Kumar, Benjamin Newman, Binhang Yuan, Bobby Yan, Ce Zhang, et al . 2023. Holis- tic Evaluation of Language Models.arXiv preprint arXiv:2211.09110(2023). doi:10.48550/arXiv.2211.09110

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2211.09110 2023

  10. [10]

    Jimmy Lin, Rodrigo Nogueira, and Andrew Yates. 2021. Pretrained Transformers for Text Ranking: BERT and Beyond. InProceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies: Tutorials (NAACL ’21). doi:10.18653/v1/2021.naacl- tutorials.1

    work page doi:10.18653/v1/2021.naacl- 2021

  11. [11]

    Xueguang Ma, Xinyu Zhang, Ronak Pradeep, and Jimmy Lin. 2023. Zero-Shot Listwise Document Reranking with a Large Language Model.arXiv preprint arXiv:2305.02156(2023). doi:10.48550/arXiv.2305.02156

    work page doi:10.48550/arxiv.2305.02156 2023

  12. [12]

    OpenAI, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Alt- man, Shyamal Anadkat, Red Avila, Igor Babuschkin, et al. 2024. GPT-4 Technical Report.arXiv preprint arXiv:2303.08774(2024). doi:10.48550/arXiv.2303.08774

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2303.08774 2024

  13. [13]

    Ronak Pradeep, Sahel Sharifymoghaddam, and Jimmy Lin. 2023. RankVicuna: Zero-Shot Listwise Document Reranking with Open-Source Large Language Models.arXiv preprint arXiv:2309.15088(2023). doi:10.48550/arXiv.2309.15088

    work page doi:10.48550/arxiv.2309.15088 2023

  14. [14]

    Ronak Pradeep, Sahel Sharifymoghaddam, and Jimmy Lin. 2023. RankZephyr: Effective and Robust Zero-Shot Listwise Reranking Is a Breeze!arXiv preprint arXiv:2312.02724(2023). doi:10.48550/arXiv.2312.02724

    work page internal anchor Pith review doi:10.48550/arxiv.2312.02724 2023

  15. [15]

    Zhen Qin, Rolf Jagerman, Kai Hui, Honglei Zhuang, Junru Wu, Le Yan, Jiaming Shen, Tianqi Liu, Jialu Liu, Donald Metzler, et al. 2024. Large Language Models Are Effective Text Rankers with Pairwise Ranking Prompting. InProceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics (NAACL ’24). Mexico City,...

    work page doi:10.18653/v1/2024.f 2024

  16. [16]

    Revanth Gangi Reddy, JaeHyeok Doo, Yifei Xu, Md Arafat Sultan, Deevya Swain, Avirup Sil, and Heng Ji. 2024. FIRST: Faster Improved Listwise Reranking with Single Token Decoding. InProceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (EMNLP ’24). Miami, USA, 8642–8652. doi:10.186 53/v1/2024.emnlp-main.491

    work page 2024

  17. [17]

    Stephen Robertson and Hugo Zaragoza. 2009. The Probabilistic Relevance Frame- work: BM25 and Beyond. (2009), 333–389. doi:10.1561/1500000019

    work page doi:10.1561/1500000019 2009

  18. [18]

    Devendra Sachan, Mike Lewis, Mandar Joshi, Armen Aghajanyan, Wen-tau Yih, Joelle Pineau, and Luke Zettlemoyer. 2023. Improving Passage Retrieval with Zero-Shot Question Generation. InProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing (EMNLP ’22). Abu Dhabi, United Arab Emirates, 3781–3797. doi:10.18653/v1/2022.emnlp-main.249

    work page doi:10.18653/v1/2022.emnlp-main.249 2023

  19. [19]

    Xinyuan Song, Keyu Wang, PengXiang Li, Lu Yin, and Shiwei Liu. 2025. De- mystifying the roles of llm layers in retrieval, knowledge, and reasoning.arXiv preprint arXiv:2510.02091(2025). doi:10.48550/arXiv.2510.02091

    work page doi:10.48550/arxiv.2510.02091 2025

  20. [20]

    Hongjin Su, Shuyang Jiang, Yuhang Lai, Haoyuan Wu, Boao Shi, Che Liu, Qian Liu, and Tao Yu. 2024. ARKS: Active Retrieval in Knowledge Soup for Code Generation.arXiv preprint arXiv:2402.12317(2024). doi:10.48550/arXiv.2402.12317

    work page doi:10.48550/arxiv.2402.12317 2024

  21. [21]

    Siegel, Michael Tang, Ruoxi Sun, Jinsung Yoon, Sercan O

    Hongjin Su, Howard Yen, Mengzhou Xia, Weijia Shi, Niklas Muennighoff, Han-yu Wang, Haisu Liu, Quan Shi, Zachary S. Siegel, Michael Tang, Ruoxi Sun, Jinsung Yoon, Sercan O. Arik, Danqi Chen, and Tao Yu. 2025. BRIGHT: A Realistic and Challenging Benchmark for Reasoning-Intensive Retrieval. InProceedings of the Thirteenth International Conference on Learning...

    work page doi:10.48550/arxiv.2407.12883 2025

  22. [22]

    Weiwei Sun, Lingyong Yan, Xinyu Ma, Shuaiqiang Wang, Pengjie Ren, Zhumin Chen, Dawei Yin, and Zhaochun Ren. 2023. Is ChatGPT Good at Search? Investi- gating Large Language Models as Re-Ranking Agents. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP ’23). Singapore, 14918–14937. doi:10.18653/v1/2023.emnlp-main.923

    work page doi:10.18653/v1/2023.emnlp-main.923 2023

  23. [23]

    Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, and Iryna Gurevych. 2021. BEIR: A Heterogenous Benchmark for Zero-shot Evaluation of Information Retrieval Models. InProceedings of the 35th Annual Conference on Neural Information Processing Systems ((NIPS ’21). doi:10.48550/arXiv.2104.08663

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2104.08663 2021

  24. [24]

    Jason Wei, Maarten Paul Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew Mingbo Dai, and Quoc V. Le. 2022. Finetuned Language Models are Zero-Shot Learners. InProceedings of the Tenth International Conference on Learning Representations (ICLR ’22). https://openreview.net/for um?id=gEZrGCozdqR

    work page 2022

  25. [25]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, et al. 2025. Qwen3 Technical Report. arXiv preprint arXiv:2505.09388(2025). doi:10.48550/arXiv.2505.09388

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2505.09388 2025

  26. [26]

    Zhe Yin, Xiaodong Gu, and Beijun Shen. 2025. Neuron-Guided Interpretation of Code LLMs: Where, Why, and How?arXiv preprint arXiv:2512.19980(2025). doi:10.48550/arXiv.2512.19980

    work page doi:10.48550/arxiv.2512.19980 2025

  27. [27]

    Wuwei Zhang, Fangcong Yin, Howard Yen, Danqi Chen, and Xi Ye. 2025. Query- Focused Retrieval Heads Improve Long-Context Reasoning and Re-ranking. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (EMNLP ’25). Suzhou, China, 23802–23816. doi:10.18653/v1/2025.emn lp-main.1214

    work page doi:10.18653/v1/2025.emn 2025

  28. [28]

    Yang Zhang, Yanfei Dong, and Kenji Kawaguchi. 2024. Investigating layer impor- tance in large language models. InProceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP (BlackboxNLP ’24). Miami, USA, 469–479. doi:10.18653/v1/2024.blackboxnlp-1.29

    work page doi:10.18653/v1/2024.blackboxnlp-1.29 2024

  29. [29]

    Shengyao Zhuang, Hang Li, and Guido Zuccon. 2021. Deep query likelihood model for information retrieval. InProceedings of the 43rd European Conference on Information Retrieval (ECIR ’21). LUCCA, Italy, 463–470. doi:10.1007/978-3- 030-72240-1_49

    work page doi:10.1007/978-3- 2021

  30. [30]

    Shengyao Zhuang, Bing Liu, Bevan Koopman, and Guido Zuccon. 2023. Open- Source Large Language Models Are Strong Zero-shot Query Likelihood Models for Document Ranking. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP ’23). Singapore, 8807–8817. doi:10.18653 /v1/2023.findings-emnlp.590

    work page 2023

  31. [31]

    Shengyao Zhuang, Xueguang Ma, Bevan Koopman, Jimmy Lin, and Guido Zuccon

    work page

  32. [32]

    doi:10.48550/a rXiv.2503.06034

    Rank-R1: Enhancing Reasoning in LLM-based Document Rerankers via Reinforcement Learning.arXiv preprint arXiv:2503.06034(2025). doi:10.48550/a rXiv.2503.06034

    work page doi:10.48550/a 2025

  33. [33]

    Shengyao Zhuang, Honglei Zhuang, Bevan Koopman, and Guido Zuccon. 2024. A Setwise Approach for Effective and Highly Efficient Zero-shot Ranking with Large Language Models. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’24). Washington, USA. doi:10.1145/3626772.3657813

    work page doi:10.1145/3626772.3657813 2024

  34. [34]

    Shengyao Zhuang and Guido Zuccon. 2021. TILDE: Term Independent Likelihood moDEl for Passage Re-ranking. InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’21). 1483–

    work page 2021

  35. [35]

    doi:10.1145/3404835.3462922

    work page doi:10.1145/3404835.3462922