Pith. sign in

REVIEW 3 major objections 4 minor 5 cited by

REARANK: Reasoning Re-ranking Agent via Reinforcement Learning

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read REARANK shows that a 7-billion-parameter reranking agent, trained by reinforcement learning on only 179 labeled queries, matches GPT-4 on standard retrieval benchmarks and outperforms it on reasoning-intensive ones.

desk verdict Worth a referee, but the GPT-4 parity claim rests on an unmatched baseline and a dataset-size arithmetic error. read the letter →

arxiv 2505.20046 v1 pith:B2NITSQF submitted 2025-05-26 cs.IR cs.CL

Le Zhang , Bo Wang , Xipeng Qiu , Siva Reddy , Aishwarya Agrawal This is my paper
classification cs.IRcs.CL
keywords listwisererankingreinforcementlearningreasoningagentGRPOdataaugmentationpassageretrievalBM25small-datatraining
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

REARANK sets out to show that a compact 7-billion-parameter language model can become a top-tier passage reranker through reinforcement learning alone, with no supervised fine-tuning and only 179 annotated queries. The method treats listwise reranking—reordering a whole window of passages at once—as a decision problem, rewards explicit reasoning before ranking with a normalized nDCG@10 signal, and expands the tiny labeled seed into 12,000 training instances by sampling diverse 20-passage windows from BM25 results. The resulting model matches GPT-4 on TREC-DL and BEIR and surpasses it on the reasoning-intensive BRIGHT benchmark. A careful reader would take away that data scarcity need not block high-quality reranking, and that smaller models with visible reasoning can compete with much larger proprietary ones.

What carries the argument

The load-bearing mechanism is GRPO (grouped policy optimization) applied to listwise reranking with a rule-based composite reward. The policy is prompted to reason inside <think> tags and then output the permutation inside <answer> tags; the reward is $r = 0.8 r_{rank} + 0.1 r_{format1} + 0.1 r_{format2}$, where $r_{rank} = (r_{rerank} - r_{init})/(r^* - r_{init})$ is the min-max normalized relative nDCG@10 improvement over the initial BM25 ranking. This normalization removes variance from candidate sets that differ in how many relevant passages they contain. A multi-sampling augmentation pipeline converts each of 179 queries into 50 random 20-passage windows drawn from BM25 top-100 results, yielding 12,000 listwise training instances. The sliding-window permutation function then lets the trained model rerank all 100 candidates in 10 LLM calls per query.

What would settle it

Retrain REARANK exactly as described but with the 179 seed queries replaced by a different 179-query sample from MS MARCO-V2, then evaluate on BRIGHT and BEIR; if the nDCG@10 parity with GPT-4, and especially the BRIGHT edge of 17.7 vs 16.8, does not reproduce, the result is tied to the specific seed rather than to the method.

Watch

Extended reading notes

Core claim

The central discovery is that reinforcement learning, not more labels, is what unlocks strong reranking from small data. Starting from Qwen2.5-7B-Instruct and applying GRPO directly, without an initial SFT phase, REARANK learns to emit a <think> block that evaluates each passage's relevance and then a ranked list inside <answer> tags. The reward is a min-max normalized relative improvement in nDCG@10 over the initial BM25 ranking, scaled by the best achievable score for that candidate set, plus format rewards; this low-variance signal is what makes training stable. With 179 queries from MS MARCO-V2, each expanded into 50 random 20-passage sets, the model sees 12,000 listwise ranking tasks and, at inference, reranks BM25 top-100 passages with a sliding window of 20. The paper reports nDCG@10 comparable to GPT-4 on TREC-DL19/20 and BEIR, and a win on BRIGHT (17.7 vs 16.8), while using roughly 0.2% of the training data of a concurrent setwise RL reranker.

Load-bearing premise

The load-bearing premise is that 179 queries from MS MARCO-V2, randomly expanded by sampling 20-passage windows from BM25 top-100 results, form a training distribution rich enough to teach a ranking policy that generalizes to BEIR and BRIGHT.

Editorial extensions

If this is right

  • Listwise ranking rewards carry much more learning signal than binary setwise ones: REARANK needs 179 queries where a concurrent setwise RL reranker uses 72,000, and still wins on BEIR and BRIGHT.
  • RL without an SFT cold start can beat large-scale distillation: REARANK-7B outperforms RankZephyr-7B, which was trained on 105,000 synthetic rankings.
  • Explicit reasoning learned for reranking transfers to other reasoning tasks, as shown by improved pass@1 scores on AIME 2024 and AMC after reranking training.
  • A single reranking pass is enough; multiple passes do not reliably help, which keeps inference at only 10 LLM calls per 100 passages.
  • Compact open-weight models can serve as practical stand-ins for proprietary rankers, lowering cost and latency while retaining competitive accuracy.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The multi-sampling augmentation recipe should transfer to other ranking tasks with small annotated seeds, such as ranking candidate answers, tool calls, or retrieved evidence in RAG pipelines, wherever fine-grained preference labels exist for a few queries.
  • The normalized relative NDCG reward is a general design principle: score progress toward the best achievable ranking for a given candidate set, not raw quality, to keep RL gradients stable when candidate difficulty varies.
  • Because REARANK shows no correlation between reasoning length and reranking performance, its reasoning trace may be a decision scaffold rather than a search process; a concise-reasoning distillation that preserves nDCG would test this directly.
  • The method's dependence on BM25 candidates is untested; switching the initial retriever to a dense or hybrid one would reveal whether the learned policy transfers across candidate distributions or only to BM25-shaped inputs.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces REARANK, a listwise LLM reranking agent built on Qwen2.5-7B and trained with GRPO reinforcement learning. To reduce annotation cost, the authors sample multiple 20-passage candidate sets from BM25 top-100 results for each of 179 annotated MS MARCO-V2 queries, filter low-quality sets, and train with a reward that combines a min-max normalized NDCG@10 term with format rewards. At inference, REARANK produces explicit reasoning before outputting a passage ranking, using a sliding-window listwise strategy. Experiments report nDCG@10 on TREC-DL19/DL20, BEIR, and BRIGHT, with claims that REARANK is comparable to GPT-4 in-domain and out-of-domain, and surpasses GPT-4 on BRIGHT, while using far fewer annotated queries than the Rank-R1 baseline.

Significance. If the results hold, the paper makes a useful contribution: it shows that explicit reasoning can be trained into a compact reranker with RL from a very small annotation budget, and it provides a concrete data-augmentation recipe plus a public code release. The paper also includes thoughtful ablations (reward normalization, data filtering, reasoning on/off) and a transfer experiment to math reasoning, which strengthen the evidence that the RL-trained reasoning is not merely a prompt artifact. The comparison with Rank-R1 is valuable because it isolates listwise versus setwise RL training on the same base model. These strengths make the central idea worth pursuing.

major comments (3)
  1. [§4.2, Table 1 footnote; §4.3, Table 2] The GPT-4 comparison is not like-for-like and this undermines the headline claims. Table 1's footnote states that RankGPT4 reranks the top 30 passages from RankGPT3.5 while all other models rerank the BM25 top 100; Table 2's caption says that all other models rerank the top 100 BM25 results but does not state what RankGPT4 does. If the same top-30 cascade was used in Table 2, the BRIGHT result (REARANK 17.7 vs. RankGPT4 16.8) is compared against a candidate pool truncated by a lossy RankGPT3.5 pass, which can suppress recall and make the 'surpasses GPT-4' claim an artifact of the cascade. If a different protocol was used in Table 2, the two tables are inconsistent. The abstract and conclusion do not mention this asymmetry. Please rerun GPT-4 on the same BM25 top-100 candidate set as REARANK, or clearly state the protocol and qualify the claims accordingly.
  2. [§4.1, Training Details] The reported training data size is internally inconsistent. The text says training instances are generated by sampling 20 candidate passages per query, repeated 50 times for each of the 179 queries, which gives 179 × 50 = 8,950 instances before filtering, and the stated filter (removing samples with no relevant passages or initial nDCG@10 < 0.1) can only decrease this count. The paper nevertheless reports 'resulting in 12k training instances.' This number also appears in the direct-SFT ablation in Table 3, so the low-data claim and the SFT comparison depend on it. Please correct the sampling counts or the reported training size.
  3. [§4.3, Table 2] The claim that REARANK 'surpasses GPT-4' on BRIGHT rests on a single average (17.7 vs. 16.8) that is not accompanied by any variance estimate or significance test, and the per-domain pattern is mixed: GPT-4 is better on 7 of the 13 BRIGHT subsets while REARANK is better on 6, with the average gap driven by large margins on a few domains (e.g., Sustainability, Pony, AoPS). Given the protocol concern above, this superiority claim is not currently established. Please report standard errors or paired tests, or at least discuss the per-domain variability before making the 'surpasses' claim.
minor comments (4)
  1. [§4.1, Baselines] Table 1 includes RankGPT3.5, but Section 4.1's baseline enumeration does not describe how RankGPT3.5 was prompted and evaluated; please add a sentence to include it.
  2. [§4.4, Table 3] The column labeled 'TREC-DL' in Table 3 appears to be the average of DL19 and DL20, but the label is ambiguous; please label it 'TREC-DL (avg)' or report the two collections separately.
  3. [Figure 1] The top panel of Figure 1 has no axis labels or legend, making it impossible to identify which benchmarks and models are plotted; please add these details.
  4. [§3.4 and Appendix B] There are typos: 'intial ranking' should be 'initial ranking' in Section 3.4, and 'TREL-DL' should be 'TREC-DL' in Appendix B.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the RL training signal and evaluations are anchored in independent human relevance judgments, and no predicted quantity or self-citation chain defines the reported result.

full rationale

The derivation chain is self-contained with respect to the paper's central claim. The training reward (Eq. 6) is a weighted combination of normalized NDCG@10 computed from human-annotated relevance judgments (Eq. 5) and format rewards; the evaluation metric is also NDCG@10, but this is standard supervision rather than circularity because the relevance labels come from an external human source and no model output is used to define the target. The 179-query seed and 50x passage-window sampling generate 12k training instances, but the same human judgments are reused only as ground truth, not as fitted predictions. The ablations isolate RL from SFT and from prompting, supporting the claim that RL training, not the augmentation alone, drives gains. Self-citations (e.g., Zhang et al. 2024a,c in the negative-sampling analogy) are contextual and not load-bearing; no uniqueness theorem or ansatz is imported from the authors' prior work. The protocol asymmetry for RankGPT4 on BRIGHT is a baseline-fairness concern, not a circularity, because it does not make REARANK's result equivalent to its inputs by construction.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the choice of NDCG@10 as both reward and metric, on the representativeness of the 179-query seed after augmentation, and on the stability of GRPO for this task. No entirely new entities (particles, mediators, dimensions) are introduced; the method combines existing techniques with hand-chosen parameters (reward weights, filter threshold, window size, sampling count).

free parameters (4)
  • reward weight for rrank = 0.8
    Equation (6) sets the total reward as 0.8*rrank + 0.1*rformat1 + 0.1*rformat2; the weights are hand-chosen without a sensitivity study.
  • initial nDCG@10 filter threshold = 0.1
    Section 4.1 filters out training instances with initial nDCG@10 < 0.1; this threshold is chosen by the authors and affects the training distribution.
  • sliding window size k = 20
    Section 4.1 sets window size to 20 passages per LLM call; this is a design choice that influences both quality and number of calls.
  • multi-sampling count per query = 50
    Section 3.4 samples 50 passage sets per query to build training instances; the resulting dataset size depends on this number.
assumptions (3)
  • domain assumption NDCG@10 is a valid measure of ranking quality for training and evaluation.
    The reward and all reported results use NDCG@10; if this metric does not capture what users care about, the optimization target is questionable.
  • domain assumption The 179-query MS MARCO-V2 seed set, after augmentation, is representative enough that learned reranking transfers to other domains.
    The low-data claim and OOD results depend on this; the paper tests on BEIR and BRIGHT but does not show that the 179 queries span the difficulty range of those benchmarks.
  • domain assumption GRPO with a rule-based reward effectively optimizes the LLM policy for ranking without catastrophic forgetting or collapse.
    The paper adopts GRPO from DeepSeek-R1 without modifications; it assumes the algorithm's stability guarantees hold for this task.

how reviews work

0 comments
Cite this review

Pith. "Pith review of REARANK: Reasoning Re-ranking Agent via Reinforcement Learning." pith.science (2026). https://pith.science/paper/B2NITSQF

@misc{pith2026250520046,
  author       = {Pith},
  title        = {Pith review of: REARANK: Reasoning Re-ranking Agent via Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B2NITSQF}},
  note         = {Machine review of arXiv:2505.20046}
}
read the original abstract

We present REARANK, a large language model (LLM)-based listwise reasoning reranking agent. REARANK explicitly reasons before reranking, significantly improving both performance and interpretability. Leveraging reinforcement learning and data augmentation, REARANK achieves substantial improvements over baseline models across popular information retrieval benchmarks, notably requiring only 179 annotated samples. Built on top of Qwen2.5-7B, our REARANK-7B demonstrates performance comparable to GPT-4 on both in-domain and out-of-domain benchmarks and even surpasses GPT-4 on reasoning-intensive BRIGHT benchmarks. These results underscore the effectiveness of our approach and highlight how reinforcement learning can enhance LLM reasoning capabilities in reranking.

Figures

Figures reproduced from arXiv: 2505.20046 by the authors.

Figure 1
Figure 1. (Top) Average rerank results on popular bench [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Listwise vs. Setwise Reranking. Setwise reranking yields binary scores (0 or 1); listwise rerank￾ing offers richer, continuous scores between 0 and 1. 3 Method 3.1 Listwise Re-ranking Agent Given a query q and an initial set of n retrieved pas￾sages P = (p1, . . . , pn), the objective of reranking is to find the optimal permutation (ranking) of these passages. This can be formally expressed as maxi￾mizing a ranking … view at source ↗
Figure 3
Figure 3. Pipeline of the proposed GRPO-based RL framework for listwise passage reranking. Training utilizes data [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: (Top) Reward evolving curve (Bottom) Re [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Reasoning patterns: Before- vs. After-RL training under identical prompt and query. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Impact of reasoning length. Data points are [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Multiple reranking results of REARANK. Multiple Rerank Pass Analyzing the im￾pact of multiple reranking passes with REAR￾ANK(evaluating performance with and without rea￾soning) reveals mixed results across benchmarks (fig. 7). Multiple passes improve performance on TER…
Figure 8
Figure 8. Figure 8: Examples from biology split of BRIGHT, only 3 passages selected for better visibility. Relevant sentences [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 5 Pith papers

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

  1. Optimizing RAG Rerankers with LLM Feedback via Reinforcement Learning

    cs.CL 2026-04 conditional novelty 6.0 of 10

    RRPO formulates document reranking as a sequential MDP and optimizes a pointwise reranker with PPO using LLM generation rewards and a reference-anchored deterministic baseline.

  2. MemSifter: Offloading LLM Memory Retrieval via Outcome-Driven Proxy Reasoning

    cs.IR 2026-03 conditional novelty 6.0 of 10

    MemSifter trains a 4B proxy with an outcome-driven, rank-sensitive RL reward to sift LLM memory, and on eight benchmarks it matches or beats embedding, graph, and long-context baselines.

  3. ERank: Fusing Supervised Fine-Tuning and Reinforcement Learning for Effective and Efficient Text Reranking

    cs.IR 2025-08 conditional novelty 6.0 of 10

    ERank trains a pointwise reasoning-LLM reranker with fine-grained integer scoring and a listwise reinforcement-learning reward, reporting nDCG@10 of 40.2 on BRIGHT for its 32B variant with BM25 hybrid scoring.

  4. A Vision for Geo-Temporal Deep Research Systems: Towards Comprehensive, Transparent, and Reproducible Geo-Temporal Information Synthesis

    cs.CL 2025-06 conditional novelty 4.0 of 10

    A research agenda calling for geo-temporal reasoning in deep research systems, with no experiments or system implementation.

  5. CogniSQL-R1-Zero: Lightweight Reinforced Reasoning for Efficient SQL Generation

    cs.AI 2025-07 reject novelty 3.0 of 10

    A 7B model trained with GRPO and a sparse execution-correctness reward reaches 59.97% execution accuracy on BIRD dev, though the evaluation protocol and baseline numbers contain inconsistencies.

Reference graph

Works this paper leans on

44 extracted references · 6 canonical work pages · cited by 5 Pith papers

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  2. [2]

    Parishad BehnamGhader, Vaibhav Adlakha, Marius Mosbach, Dzmitry Bahdanau, Nicolas Chapados, and Siva Reddy. 2024. Llm2vec: Large language models are secretly powerful text encoders. arXiv preprint arXiv:2404.05961

    arXiv 2024

  3. [3]

    Sebastian Borgeaud, Arthur Mensch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George Bm Van Den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, and 1 others. 2022. Improving language models by retrieving from trillions of tokens. In International conference on machine learning, pages 2206--2240. PMLR

    2022

  4. [5]

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

    arXiv 2020

  5. [6]

    Thibault Formal, Benjamin Piwowarski, and St\' e phane Clinchant. 2021. https://doi.org/10.1145/3404835.3463098 SPLADE: Sparse Lexical and Expansion Model for First Stage Ranking , page 2288–2292. Association for Computing Machinery, New York, NY, USA

    arXiv 2021

  6. [7]

    Evan Greensmith, Peter L Bartlett, and Jonathan Baxter. 2004. Variance reduction techniques for gradient estimates in reinforcement learning. Journal of Machine Learning Research, 5(Nov):1471--1530

    work page 2004

  7. [8]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

    arXiv 2025

  8. [9]

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2022. Large language models are zero-shot reasoners. Advances in neural information processing systems, 35:22199--22213

    2022

Show all 44 references
  1. [10]

    Chankyu Lee, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. 2024. Nv-embed: Improved techniques for training llms as generalist embedding models. arXiv preprint arXiv:2405.17428

    2024 arXiv

  2. [11]

    u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K \"u ttler, Mike Lewis, Wen-tau Yih, Tim Rockt \"a schel, and 1 others. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in neural informati...

    2020

  3. [12]

    Percy Liang, Rishi Bommasani, Tony Lee, Dimitris Tsipras, Dilara Soylu, Michihiro Yasunaga, Yian Zhang, Deepak Narayanan, Yuhuai Wu, Ananya Kumar, and 1 others. 2022. Holistic evaluation of language models. arXiv preprint arXiv:2211.09110

    2022 arXiv

  4. [13]

    Xueguang Ma, Liang Wang, Nan Yang, Furu Wei, and Jimmy Lin. 2024 a . Fine-tuning llama for multi-stage text retrieval. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2421--2425

    2024

  5. [14]

    Xueguang Ma, Liang Wang, Nan Yang, Furu Wei, and Jimmy Lin. 2024 b . Fine-tuning llama for multi-stage text retrieval. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2421--2425

    2024

  6. [15]

    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 arXiv

  7. [16]

    Sara Vera Marjanovi \'c , Arkil Patel, Vaibhav Adlakha, Milad Aghajohari, Parishad BehnamGhader, Mehar Bhatia, Aditi Khandelwal, Austin Kraft, Benno Krojer, Xing Han L \`u , and 1 others. 2025. Deepseek-r1 thoughtology: Let's< think> about llm reasoning. arXiv preprint arXiv:2...

    2025

  8. [17]

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

    2023 arXiv

  9. [18]

    Ronak Pradeep, Sahel Sharifymoghaddam, and Jimmy Lin. 2023 b . Rankzephyr: Effective and robust zero-shot listwise reranking is a breeze! arXiv preprint arXiv:2312.02724

    2023 arXiv

  10. [19]

    Zhen Qin, Rolf Jagerman, Kai Hui, Honglei Zhuang, Junru Wu, Le Yan, Jiaming Shen, Tianqi Liu, Jialu Liu, Donald Metzler, and 1 others. 2023. Large language models are effective text rankers with pairwise ranking prompting. arXiv preprint arXiv:2306.17563

    2023 arXiv

  11. [20]

    Nils Reimers and Iryna Gurevych. 2019. https://arxiv.org/abs/1908.10084 Sentence-bert: Sentence embeddings using siamese bert-networks . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics

    2019 arXiv

  12. [21]

    Nils Reimers and Iryna Gurevych. 2020. https://arxiv.org/abs/2004.09813 Making monolingual sentence embeddings multilingual using knowledge distillation . In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing. Association for Computational L...

    2020 arXiv

  13. [22]

    Stephen Robertson, Hugo Zaragoza, and 1 others. 2009. The probabilistic relevance framework: Bm25 and beyond. Foundations and Trends in Information Retrieval , 3(4):333--389

    2009

  14. [23]

    Devendra Singh Sachan, Mike Lewis, Mandar Joshi, Armen Aghajanyan, Wen-tau Yih, Joelle Pineau, and Luke Zettlemoyer. 2022. Improving passage retrieval with zero-shot question generation. arXiv preprint arXiv:2204.07496

    2022 arXiv

  15. [24]

    John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. 2015. High-dimensional continuous control using generalized advantage estimation. arXiv preprint arXiv:1506.02438

    2015 arXiv

  16. [25]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, and 1 others. 2024. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300

    2024 arXiv

  17. [26]

    Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. 2024. Hybridflow: A flexible and efficient rlhf framework. arXiv preprint arXiv:2409.19256

    2024 arXiv

  18. [27]

    Hongjin Su, Howard Yen, Mengzhou Xia, Weijia Shi, Niklas Muennighoff, Han-yu Wang, Haisu Liu, Quan Shi, Zachary S Siegel, Michael Tang, and 1 others. 2024. Bright: A realistic and challenging benchmark for reasoning-intensive retrieval. arXiv preprint arXiv:2407.12883

    2024 arXiv

  19. [28]

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

    2023 arXiv

  20. [29]

    Nandan Thakur, Nils Reimers, Andreas R \"u ckl \'e , Abhishek Srivastava, and Iryna Gurevych. 2021. Beir: A heterogenous benchmark for zero-shot evaluation of information retrieval models. arXiv preprint arXiv:2104.08663

    2021 arXiv

  21. [30]

    Shuai Wang, Shengyao Zhuang, and Guido Zuccon. 2021. Bert-based dense retrievers require interpolation with bm25 for effective passage retrieval. In Proceedings of the 2021 ACM SIGIR international conference on theory of information retrieval, pages 317--324

    2021

  22. [31]

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171

    2022 arXiv

  23. [32]

    Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, and 1 others. 2022 a . Emergent abilities of large language models. arXiv preprint arXiv:2206.07682

    2022 arXiv

  24. [33]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, and 1 others. 2022 b . Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824--24837

    2022

  25. [34]

    Lee Xiong, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Paul Bennett, Junaid Ahmed, and Arnold Overwijk. 2020. Approximate nearest neighbor negative contrastive learning for dense text retrieval. arXiv preprint arXiv:2007.00808

    2020 arXiv

  26. [35]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, and 1 others. 2025. Qwen3 technical report. arXiv preprint arXiv:2505.09388

    2025 arXiv

  27. [36]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, and 1 others. 2024. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115

    2024 arXiv

  28. [37]

    Yixin Ye, Zhen Huang, Yang Xiao, Ethan Chern, Shijie Xia, and Pengfei Liu. 2025. Limo: Less is more for reasoning. arXiv preprint arXiv:2502.03387

    2025 arXiv

  29. [38]

    Le Zhang, Rabiul Awal, and Aishwarya Agrawal. 2024 a . Contrasting intra-modal and ranking cross-modal hard negatives to enhance visio-linguistic compositional understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 137...

    2024

  30. [39]

    Le Zhang, Yihong Wu, Fengran Mo, Jian-Yun Nie, and Aishwarya Agrawal. 2023. Moqagpt: Zero-shot multi-modal open-domain question answering with large language model. arXiv preprint arXiv:2310.13265

    2023 arXiv

  31. [40]

    Le Zhang, Yihong Wu, Qian Yang, and Jian-Yun Nie. 2024 b . Exploring the best practices of query expansion with large language models. arXiv preprint arXiv:2401.06311

    2024 arXiv

  32. [41]

    Le Zhang, Qian Yang, and Aishwarya Agrawal. 2024 c . Assessing and learning alignment of unimodal vision and language models. arXiv preprint arXiv:2412.04616

    2024 arXiv

  33. [42]

    Shengyao Zhuang, Xueguang Ma, Bevan Koopman, Jimmy Lin, and Guido Zuccon. 2025. Rank-r1: Enhancing reasoning in llm-based document rerankers via reinforcement learning. arXiv preprint arXiv:2503.06034

    2025 arXiv

  34. [43]

    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. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information ...

    2024

  35. [44]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  36. [45]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.