Pith. sign in

REVIEW 2 major objections 7 minor 4 cited by

RaDeR: Reasoning-aware Dense Retrieval Models

T0 review · 2 major / 7 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 A dense retriever trained on synthetic math-reasoning trajectories beats BM25 on chain-of-thought queries.

desk verdict Solid method paper for reasoning-aware first-stage retrieval; the ProofWiki overlap and unnamed generator LLM need fixes, but the core transfer story holds. read the letter →

arxiv 2505.18405 v2 pith:ZYDF427S submitted 2025-05-23 cs.CL cs.IR

Debrup Das , Sam O' Nuallain , Razieh Rahimi This is my paper
classification cs.CLcs.IR
keywords denseretrievalreasoning-awaresynthetictrainingdataMonteCarloTreeSearchchain-of-thoughtmathematicalreasoningself-reflectionretrieval-augmentedgeneration
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

RaDeR sets out to make first-stage dense retrieval reason about relevance instead of just matching terms. It builds training data automatically by having an LLM solve math problems with retrieval-augmented Monte Carlo Tree Search: theorems retrieved on trajectories that reach the gold answer are treated as relevant, and theorems the LLM self-reflection rejects become hard negatives. Trained on only 43,120 synthesized samples, the resulting retriever outperforms sparse and dense baselines on the BRIGHT reasoning-retrieval benchmark, generalizes to the RAR-b Math and Coding splits, and is the first dense retriever to outperform BM25 when queries are chain-of-thought reasoning steps. The paper argues this shows reasoning-aware first-stage retrieval can be produced cheaply from synthetic math data and materially improves retrieval-augmented reasoning.

What carries the argument

The load-bearing object is the retrieval-augmented MCTS trajectory with self-reflection labels. In this tree search, a node is a partial solution, and one action generates a retrieval query while the next retrieves the top-k theorems from a theorem corpus; a self-reflection step labels each retrieved theorem relevant or non-relevant before it can enter the tree. The reason this works is that it converts solution search into a data factory: the gold-answer reward selects trajectories, the retrieved theorem on a successful trajectory becomes the positive training document, and the self-reflection rejects become hard negatives. Four query types are generated from each trajectory so the retriever sees both reasoning-intensive and term-matching examples. This is what carries the claim that reasoning-aware relevance can be learned without human relevance judgments.

What would settle it

Retrain RaDeR exactly as described but score trajectories with a process reward that verifies each reasoning step, not just the final answer, and compare nDCG@10 on BRIGHT; if the two retrievers tie, the final-answer proxy is sufficient, while if the process-reward retriever is better, the conceded noise in the training data is load-bearing. A cheaper check is to have human judges label a random sample of the synthesized (query, positive theorem) pairs as genuinely necessary for the correct solution and see whether the fraction of unnecessary positives predicts the benchmark gap.

Watch

Extended reading notes

Core claim

The central claim is that a first-stage bi-encoder dense retriever can learn a reasoning-aware relevance function from synthetic data alone, provided the data is generated by a verifiable reasoning process. The pipeline uses an MCTS solver whose actions include query generation and theorem retrieval; only retrieval-augmented trajectories that end in the correct answer are mined, and each retrieved theorem is paired with four query types (LLM-generated reasoning queries, CoT reasoning queries, the original question, and lexical queries). The relevant theorem is the positive document and rejected theorems are hard negatives. On evaluation, the best RaDeR model improves average nDCG@10 on BRIGHT by at least 2 points over baselines in both question-query and CoT-query settings, and beats BM25 on CoT queries, which previous dense retrievers did not. The authors also report that the same retriever remains competitive on MS MARCO, suggesting the reasoning training does not destroy term-matching ability.

Load-bearing premise

The entire training signal rests on two linked proxies: a theorem retrieved along a trajectory that reaches the gold answer is treated as relevant, and a theorem that self-reflection rejects is treated as irrelevant; if a wrong reasoning path can still lead to a correct answer, the labels are noisy and the retriever learns the LLM's biases rather than a robust relevance function.

Editorial extensions

If this is right

  • First-stage retrievers no longer need to be lexical or semantic matchers: the embedding can encode the reasoning step that connects a query to a relevant document, so retrieval and reasoning are no longer separated.
  • Because math-trained retrieval transfers to coding and theorem splits of BRIGHT, a single reasoning-aware retriever can serve heterogeneous reasoning tasks without task-specific training.
  • Retrieval-augmented reasoning LLMs can use CoT steps as queries instead of the original question, and get better results than with BM25, which matters for agentic and interleaved reasoning systems.
  • The 43,120-sample synthetic set is enough to reach the reported performance, so the bottleneck for reasoning-aware retrieval is trajectory quality rather than data volume.

Reading between the lines

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

  • If the trajectory-label proxy is as reliable as the paper suggests, the same recipe transfers to any domain with verifiable answers and a retrievable knowledge base, such as interactive code synthesis or formal proof search, without needing human relevance judgments.
  • The retriever likely inherits the reasoning biases of the LLM that generates the trajectories; a testable consequence is that swapping the generating LLM would change which wrong theorems are retrieved, even if benchmark scores stay similar.
  • Because relevance is judged per document in isolation, multi-hop questions where relevance is defined jointly across documents probably sit beyond what this training signal captures; combining RaDeR with a set-wise scorer or iterative retrieval would be the natural next step.
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

2 major / 7 minor

Summary. The paper proposes RaDeR, a method for training first-stage dense retrieval models and rerankers on synthetic data generated by LLM-guided Monte Carlo Tree Search (MCTS) over mathematical problem-solving trajectories. The pipeline produces several query types (chain-of-thought reasoning steps, LLM-generated reasoning queries, original questions, and lexical queries) and hard negatives drawn from self-reflection labels. The authors evaluate RaDeR on BRIGHT, the Math and Coding splits of RAR-b, MS MARCO, and a retrieval-augmented QA task, reporting substantial gains over strong baselines, particularly on BRIGHT's theorem-based splits. They further claim that RaDeR is the first dense retriever to outperform BM25 when queries are chain-of-thought reasoning steps, and that it matches ReasonIR's performance with only 2.5% of its training samples.

Significance. If the claims hold, this is a valuable practical contribution: it shows that reasoning-aware first-stage retrievers can be trained cheaply from synthetic MCTS data and that they transfer to several reasoning-intensive benchmarks. The manuscript is comprehensive in coverage, includes ablations on query types and model scales, and provides public code, data, and models, which facilitates reproducibility. The main risk to the central generalization claim is the overlap between the ProofWiki corpus used in MCTS data generation and the BRIGHT theorem splits used for evaluation; the largest reported gains (37-40% relative on TheoQ/TheoT) come from this overlapping distribution. The label-generation proxies, based on final-answer correctness and LLM self-reflection, are acknowledged as noisy in the Limitations section but are not quantitatively validated. These issues do not undermine the Coding, RAR-b, or MS MARCO results, but they need to be addressed before the broader generalization claim can be accepted at face value.

major comments (2)
  1. [§5 and §6 (Tables 1–2)] Section 5 states that the retrieval corpus in the MCTS pipeline is 'formal mathematical theorems from ProofWiki, which is also used in the BRIGHT benchmark.' The largest reported improvements (12.1 and 11.3 nDCG@10 points on TheoQ in Tables 1 and 2, and comparable gains on TheoT) are on BRIGHT splits whose document distribution therefore matches the training distribution. This means the theorem-split results could reflect memorization of ProofWiki-specific surface patterns (LaTeX formatting, theorem titles, tag structure) rather than a generalizable reasoning-aware relevance function. Please either evaluate on a disjoint math-theorem corpus that was not available to MCTS during data generation, or provide an analysis showing that the theorem-split gains do not come from corpus-specific formatting or terminology. The Coding and RAR-b results provide independent evidence of transfer, but the manuscript currently presents the theorem-split gains as part of the overall generalization claim.
  2. [§4.2 and §8] The training labels for positive and hard-negative pairs come from two linked proxies: (i) the assumption that theorems retrieved on a trajectory that reaches the gold answer are relevant to the original question, and (ii) LLM self-reflection labels for relevance. The Limitations section concedes that incorrect CoT paths can produce correct final answers, making the training data noisy. However, the paper does not quantify this noise, does not validate the self-reflection labels against any ground truth, and does not examine how label quality varies across trajectory types. Since the method's core claim is that these synthetic labels teach a robust reasoning-aware relevance function, please add a label-quality analysis (e.g., human judgments on a random sample of positive and hard-negative pairs, or an ablation that filters training samples by self-reflection confidence or trajectory consistency) and report how retrieval performance changes. This would directly test whether the training signal is the reasoning-aware relevance the method intends to teach.
minor comments (7)
  1. [Tables 1–6] All reported numbers are single runs without variance or significance tests; because several per-split differences between RaDeR variants and baselines are small (e.g., Table 1, Qwen2.5-7B vs. gte-Qwen2 on Bio, Earth, and Psy), reporting standard deviations or paired tests would make the comparison more reliable.
  2. [Abstract and §6.6] The 2.5% data-efficiency comparison against ReasonIR uses training-sample count only; the base model, data distribution, and training recipes differ, so this figure is not a controlled comparison. Please state this caveat explicitly where the claim is made.
  3. [§6.4, Table 6 caption] The table caption says reranking is performed 'using questions as retrieval queries,' but the text describes reranking BM25 results obtained with GPT-4o CoT as the query; the caption should be made consistent with the actual setup.
  4. [Appendix P, Figure 12] The prompt text says 'summatization' where 'summarization' is intended; please fix the typo.
  5. [Appendix C] The dataset name appears both as 'NuminaMath' and 'NuminaMATH' in the text; please use a single consistent spelling.
  6. [Figure 1] The example would be clearer if the CoT query text were visually separated from the original question text; the current layout makes it difficult to see the claim that the CoT steps do not simplify retrieval.
  7. [§1 and Abstract] The claim of being the 'first dense retriever that outperforms BM25 when queries are Chain-of-Thought reasoning steps' is made in a zero-shot setting; please state explicitly that this is a zero-shot claim, i.e., without task-specific training on the retrieval corpus, in the abstract and introduction.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RaDeR's claims are benchmark-tested against external labels; noted corpus-overlap and label-proxy issues are validity concerns, not construction-level circularity.

full rationale

RaDeR's derivation chain is empirical rather than definitional. Training data is synthesized from MCTS trajectories over ProofWiki, with relevance labels coming from LLM self-reflection and gold-answer rewards; the trained retriever is then evaluated on BRIGHT, RAR-b, and MS MARCO, all of which use externally defined gold relevance judgments. No equation in the paper defines an evaluation quantity in terms of a fitted parameter; the contrastive InfoNCE loss in Appendix F is standard, and the trained model is not used to define the benchmark labels. The reported nDCG gains are measured against held-out labels, so they do not reduce to the training signal by construction. The two caveats often raised — the ProofWiki overlap between the MCTS retrieval corpus and BRIGHT theorem splits, and the use of the same LLM self-reflection to define both positives and hard negatives — are validity/noise concerns about generalization and label quality (the paper itself acknowledges label noise in Section 8), not instances of a predicted quantity being defined as its input. They would affect interpretation of the generalization claim, not circularity. There are no load-bearing self-citations; the authors do not cite their own prior work for any premise. Therefore no circular step can be exhibited, and the paper is self-contained against external benchmarks.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical or mathematical entities. Its load-bearing postulates are that LLM-generated trajectory labels are reliable relevance labels, that final-answer correctness proxies retrieval utility, and that gains on BRIGHT theorem splits are not an artifact of sharing the ProofWiki corpus with training. Training and MCTS hyperparameters are fixed by hand (Table 8) rather than fitted to the evaluation.

free parameters (5)
  • MCTS top-k retrieved theorems per retrieval action = 5
    Controls how many theorem nodes enter the search tree; fixed in Table 8, not swept against evaluation.
  • MCTS rollouts per question = 16
    Determines trajectory diversity and synthesized data size; fixed in Table 8.
  • MCTS exploration weight C = 2
    Balances exploration and exploitation in the UCT formula; fixed in Table 8.
  • LLM generation temperature = 0.8
    Sets randomness of trajectory sampling; fixed in Table 8.
  • Hard negatives per query in retriever training = 12
    Chosen training hyperparameter; no ablation reported in the paper.
assumptions (5)
  • domain assumption Information retrieved during intermediate steps of the LLM's search process is relevant to the original question.
    Section 4 (Synthesizing Training Data) states this as the basis for using MCTS retrieval nodes as positives. It is a proxy that can inject false positives when a trajectory reaches the right answer for the wrong reason, acknowledged in Limitations.
  • domain assumption Correct final answer in the gold-labeled math datasets is a valid reward signal for the utility of retrieved theorems.
    Section 4 defines R(t)=1 only if the trajectory reaches the gold answer. Limitations notes that correct answers can follow incorrect CoT paths, making labels noisy.
  • domain assumption LLM self-reflection labels of 'relevant' and 'non-relevant' for retrieved theorems are accurate enough to define positives and hard negatives.
    Section 4.2 uses self-reflection rejects as hard negatives; there is no human verification or second model, so false negatives can enter training.
  • domain assumption Training on mathematical theorem retrieval transfers to other reasoning-aware retrieval tasks such as code, StackExchange, and RAR-b.
    Section 6 interprets cross-split gains as evidence; this is an empirical assumption about task transfer, not derived from the method.
  • ad hoc to paper The BRIGHT benchmark is a valid held-out test for generalization even though the retrieval corpus in MCTS and the BRIGHT theorem splits both derive from ProofWiki.
    Section 5 states the corpus is ProofWiki, 'which is also used in the BRIGHT benchmark.' This overlap can inflate theorem-split results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RaDeR: Reasoning-aware Dense Retrieval Models." pith.science (2026). https://pith.science/paper/ZYDF427S

@misc{pith2026250518405,
  author       = {Pith},
  title        = {Pith review of: RaDeR: Reasoning-aware Dense Retrieval Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZYDF427S}},
  note         = {Machine review of arXiv:2505.18405}
}
read the original abstract

We propose RaDeR, a set of reasoning-based dense retrieval models trained with data derived from mathematical problem solving using large language models (LLMs). Our method leverages retrieval-augmented reasoning trajectories of an LLM and self-reflective relevance evaluation, enabling the creation of both diverse and hard-negative samples for reasoning-intensive relevance. RaDeR retrievers, trained for mathematical reasoning, effectively generalize to diverse reasoning tasks in the BRIGHT and RAR-b benchmarks, consistently outperforming strong baselines in overall performance. Notably, RaDeR achieves significantly higher performance than baselines on the Math and Coding splits. In addition, RaDeR presents the first dense retriever that outperforms BM25 when queries are Chain-of-Thought reasoning steps, underscoring the critical role of reasoning-based retrieval to augment reasoning language models. Furthermore, RaDeR achieves comparable or superior performance while using only 2.5% of the training data used by the concurrent work REASONIR, highlighting the quality of our synthesized training data.

Figures

Figures reproduced from arXiv: 2505.18405 by the authors.

Figure 1
Figure 1. An example based on sample ‘Theo￾remQA_jianyuxu/pigeonhole3’ of BRIGHT, where term matching retrievers face challenges in retrieving the rel￾evant theorem w.r.t. both questions and CoT reasoning. Recent works have tried to address the reason￾ing limitation of existing models for relevance prediction. Two main approaches have emerged: (1) interleaved reasoning and retrieval (Hu et al., 2025; Jin et al., 2025; Song et… view at source ↗
Figure 2
Figure 2. An overview of the RaDeR data generation pipeline. The OST action stands for one step thought generation, and CRS stands for complete remaining so￾lution steps action. . We perform a comprehensive evaluation of RaDeR models, including evaluation of retrieval performance on reasoning-intensive benchmarks, traditional benchmarks mainly requiring term matching, as well as evaluation of target QA perfor￾mance using retr… view at source ↗
Figure 3
Figure 3. Examples of different query types from our retrieval training dataset built for the given math question. [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Example of RaDeR success case compared to Qwen2, from TheoremQA theorems of B [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: Example of RaDeR Failure cases from TheoremQA theorems BRIGHT [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]
Figure 6
Figure 6. Figure 6: Analysis of RaDeR Llama3.1-8B-instruct success case for Coding Example from [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Prompt for MCTS One Step Thought Action [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]
Figure 8
Figure 8. Figure 8: Prompt for Complete Remaining Steps (CRS) Action [PITH_FULL_IMAGE:figures/full_fig_p022_8.png]
Figure 9
Figure 9. Figure 9: Instruction for MCTS Query Generation Action [PITH_FULL_IMAGE:figures/full_fig_p022_9.png]
Figure 10
Figure 10. Figure 10: Few shot examples for MCTS Query Generation Action [PITH_FULL_IMAGE:figures/full_fig_p023_10.png]
Figure 11
Figure 11. Figure 11: Prompt for Self Reflection MCTS [PITH_FULL_IMAGE:figures/full_fig_p024_11.png]
Figure 12
Figure 12. Figure 12: Prompt for Self Summarization of Retrieved theorems [PITH_FULL_IMAGE:figures/full_fig_p025_12.png]
Figure 13
Figure 13. Figure 13: Prompt for LLM generated query generation [PITH_FULL_IMAGE:figures/full_fig_p026_13.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 4 Pith papers

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

  1. Antares: Foundation Models for Agentic Vulnerability Localization

    cs.CR 2026-08 conditional novelty 6.0 of 10

    Antares-3B, a 3B model trained with SFT plus GRPO, matches GPT-5.5 on repository-scale vulnerability localization at roughly 1/100th the inference cost.

  2. LLM-guided Hierarchical Search for End-to-end Reasoning Intensive Retrieval

    cs.IR 2025-10 conditional novelty 6.0 of 10

    An LLM directly traverses a hierarchical semantic index of a corpus, using calibrated path-relevance scores to retrieve documents for reasoning-intensive queries.

  3. SHIFT: Self-reconstruction Harnesses Implicit Fine-grained Thinking for Retrieval

    cs.IR 2026-07 conditional novelty 5.0 of 10

    SHIFT trains LLM retrievers to reason through latent continuous tokens and reconstruct explicit reasoning traces, improving reasoning-intensive retrieval.

  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.

Reference graph

Works this paper leans on

54 extracted references · 16 canonical work pages · cited by 4 Pith papers

  1. [1]

    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 INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Akiko Aizawa, Michael Kohlhase, and Iadh Ounis. 2013. https://api.semanticscholar.org/CorpusID:18544455 Ntcir-10 math pilot task overview . In NTCIR Conference on Evaluation of Information Access Technologies

    work page 2013

  4. [4]

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2023. https://arxiv.org/abs/2310.11511 Self-rag: Learning to retrieve, generate, and critique through self-reflection . Preprint, arXiv:2310.11511

    arXiv 2023

  5. [5]

    Akari Asai, Zexuan Zhong, Danqi Chen, Pang Wei Koh, Luke Zettlemoyer, Hannaneh Hajishirzi, and Wen tau Yih. 2024. https://doi.org/10.48550/arXiv.2403.03187 Reliable, adaptable, and attributable language models with retrieval . CoRR, abs/2403.03187

    arXiv doi:10.48550/arxiv.2403.03187 2024

  6. [6]

    Payal Bajaj, Daniel Campos, Nick Craswell, Li Deng, Jianfeng Gao, Xiaodong Liu, Rangan Majumder, Andrew McNamara, Bhaskar Mitra, Tri Nguyen, Mir Rosenberg, Xia Song, Alina Stoica, Saurabh Tiwary, and Tong Wang. 2018. https://arxiv.org/abs/1611.09268 Ms marco: A human generated machine reading comprehension dataset . Preprint, arXiv:1611.09268

    arXiv 2018

  7. [7]

    Luiz Bonifacio, Hugo Abonizio, Marzieh Fadaee, and Rodrigo Nogueira. 2022. https://arxiv.org/abs/2202.05144 Inpars: Data augmentation for information retrieval using large language models . Preprint, arXiv:2202.05144

    arXiv 2022

  8. [8]

    Nick Craswell, Bhaskar Mitra, Emine Yilmaz, and Daniel Campos. 2021. Overview of the trec 2020 deep learning track. corr abs/2102.07662 (2021). TREC

    arXiv 2021

Show all 54 references
  1. [9]

    Nick Craswell, Bhaskar Mitra, Emine Yilmaz, Daniel Campos, and Ellen M Voorhees. 2020. Overview of the trec 2019 deep learning track. TREC

    2020

  2. [10]

    Zhuyun Dai, Vincent Y Zhao, Ji Ma, Yi Luan, Jianmo Ni, Jing Lu, Anton Bakalov, Kelvin Guu, Keith Hall, and Ming-Wei Chang. 2023. https://openreview.net/forum?id=gmL46YMpu2J Promptagator: Few-shot dense retrieval from 8 examples . In The Eleventh International Conference on Lea...

    2023

  3. [11]

    O mer Veysel C a g atan, Akash Kundu, Martin Bernstorff, Shitao Xiao, Akshita Sukhlecha, Bhavish Pahwa, Rafa Po \'s wiata, Kranthi Kiran GV, Shawon Ashraf, Daniel Auras, Bj \

    Kenneth Enevoldsen, Isaac Chung, Imene Kerboua, M \'a rton Kardos, Ashwin Mathur, David Stap, Jay Gala, Wissam Siblini, Dominik Krzemi \'n ski, Genta Indra Winata, Saba Sturua, Saiteja Utpala, Mathieu Ciancone, Marion Schaeffer, Diganta Misra, Shreeya Dhakal, Jonathan Rystr m,...

    2025

  4. [12]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  5. [13]

    Xinyu Guan, Li Lyna Zhang, Yifei Liu, Ning Shang, Youran Sun, Yi Zhu, Fan Yang, and Mao Yang. 2025. https://arxiv.org/abs/2501.04519 rstar-math: Small llms can master math reasoning with self-evolved deep thinking . Preprint, arXiv:2501.04519

    2025 arXiv

  6. [14]

    Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and Zhiting Hu. 2023. https://openreview.net/forum?id=VTWWvYtF1R Reasoning with language model is planning with world model . In The 2023 Conference on Empirical Methods in Natural Language Processing

    2023

  7. [15]

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. https://arxiv.org/abs/2103.03874 Measuring mathematical problem solving with the math dataset . Preprint, arXiv:2103.03874

    2021 arXiv

  8. [16]

    Minda Hu, Licheng Zong, Hongru Wang, Jingyan Zhou, Jingjing Li, Yichen Gao, Kam-Fai Wong, Yu Li, and Irwin King. 2024. https://arxiv.org/abs/2406.11258 Serts: Self-rewarding tree search for biomedical retrieval-augmented generation . Preprint, arXiv:2406.11258

    2024 arXiv

  9. [17]

    Yunhai Hu, Yilun Zhao, Chen Zhao, and Arman Cohan. 2025. https://arxiv.org/abs/2503.20757 Mcts-rag: Enhancing retrieval-augmented generation with monte carlo tree search . Preprint, arXiv:2503.20757

    2025

  10. [18]

    Edward Beeching Jia LI. 2024. Numinamath. [https://github.com/project-numina/aimo-progress-prize](https://github.com/project-numina/aimo-progress-prize/blob/main/report/numina_dataset.pdf)

    2024

  11. [19]

    Bowen Jin, Hansi Zeng, Zhenrui Yue, Jinsung Yoon, Sercan Arik, Dong Wang, Hamed Zamani, and Jiawei Han. 2025. https://arxiv.org/abs/2503.09516 Search-r1: Training llms to reason and leverage search engines with reinforcement learning . Preprint, arXiv:2503.09516

    2025 arXiv

  12. [20]

    Levente Kocsis and Csaba Szepesv \'a ri. 2006. Bandit based monte-carlo planning. In Machine Learning: ECML 2006, pages 282--293, Berlin, Heidelberg. Springer Berlin Heidelberg

    2006

  13. [21]

    Jinhyuk Lee, Zhuyun Dai, Xiaoqi Ren, Blair Chen, Daniel Cer, Jeremy R. Cole, Kai Hui, Michael Boratko, Rajvi Kapadia, Wen Ding, Yi Luan, Sai Meher Karthik Duddu, Gustavo Hernandez Abrego, Weiqiang Shi, Nithi Gupta, Aditya Kusupati, Prateek Jain, Siddhartha Reddy Jonnalagadda, ...

    2024 arXiv

  14. [22]

    Yibin Lei, Liang Ding, Yu Cao, Changtong Zan, Andrew Yates, and Dacheng Tao. 2023. https://doi.org/10.18653/v1/2023.findings-acl.695 Unsupervised dense retrieval with relevance-aware contrastive pre-training . In Findings of the Association for Computational Linguistics: ACL 2...

    2023 doi

  15. [24]

    Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie, and Meishan Zhang. 2023 b . Towards general text embeddings with multi-stage contrastive learning. arXiv preprint arXiv:2308.03281

    2023 arXiv

  16. [25]

    Xueguang Ma, Luyu Gao, Shengyao Zhuang, Jiaqi Samantha Zhan, Jamie Callan, and Jimmy Lin. 2025. Tevatron 2.0: Unified document retrieval toolkit across scale, language, and modality. arXiv preprint arXiv:2505.02466

    2025 arXiv

  17. [26]

    Xueguang Ma, Liang Wang, Nan Yang, Furu Wei, and Jimmy Lin. 2023. https://arxiv.org/abs/2310.08319 Fine-tuning llama for multi-stage text retrieval . Preprint, arXiv:2310.08319

    2023 arXiv

  18. [27]

    Oard, and Richard Zanibbi

    Behrooz Mansouri, V \'i t Novotn \'y , Anurag Agarwal, Douglas W. Oard, and Richard Zanibbi. 2022. https://api.semanticscholar.org/CorpusID:251471943 Overview of arqmath-3 (2022): Third clef lab on answer retrieval for questions on math (working notes version) . In Conference ...

    2022

  19. [28]

    Niklas Muennighoff, Hongjin Su, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Amanpreet Singh, and Douwe Kiela. 2024. https://arxiv.org/abs/2402.09906 Generative representational instruction tuning . Preprint, arXiv:2402.09906

    2024 arXiv

  20. [29]

    Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, Xu Jiang, Karl Cobbe, Tyna Eloundou, Gretchen Krueger, Kevin Button, Matthew Knight, Benjamin Chess, and John Schulman. 2022...

    2022 arXiv

  21. [30]

    Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2017. https://openreview.net/forum?id=Hk1iOLcle MS MARCO : A human-generated MA chine reading CO mprehension dataset

    2017

  22. [31]

    Rodrigo Nogueira, Zhiying Jiang, Ronak Pradeep, and Jimmy Lin. 2020. https://doi.org/10.18653/v1/2020.findings-emnlp.63 Document ranking with a pretrained sequence-to-sequence model . In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 708--718, Onl...

    2020 doi

  23. [32]

    Zhenting Qi, Mingyuan MA, Jiahang Xu, Li Lyna Zhang, Fan Yang, and Mao Yang. 2025. https://openreview.net/forum?id=6aHUmotXaw Mutual reasoning makes smaller LLM s stronger problem-solver . In The Thirteenth International Conference on Learning Representations

    2025

  24. [33]

    Walker, S

    Stephen Robertson, S. Walker, S. Jones, M. M. Hancock-Beaulieu, and M. Gatford. 1995. https://www.microsoft.com/en-us/research/publication/okapi-at-trec-3/ Okapi at trec-3 . In Overview of the Third Text REtrieval Conference (TREC-3), pages 109--126. Gaithersburg, MD: NIST

    1995

  25. [34]

    Ohad Rubin, Jonathan Herzig, and Jonathan Berant. 2022. https://doi.org/10.18653/v1/2022.naacl-main.191 Learning to retrieve prompts for in-context learning . In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: ...

    2022 doi

  26. [35]

    Chris Samarinas and Hamed Zamani. 2025. https://arxiv.org/abs/2504.03947 Distillation and refinement of reasoning in small language models for document re-ranking . Preprint, arXiv:2504.03947

    2025 arXiv

  27. [36]

    Timo Schick, Jane Dwivedi-Yu, Roberto Dessi, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. https://openreview.net/forum?id=Yacmpz84TH Toolformer: Language models can teach themselves to use tools . In Thirty-seventh C...

    2023

  28. [37]

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

    2025 arXiv

  29. [38]

    Zhihong Shao, Yeyun Gong, Yelong Shen, Minlie Huang, Nan Duan, and Weizhu Chen. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.620 Enhancing retrieval-augmented large language models with iterative retrieval-generation synergy . In Findings of the Association for Comput...

    2023 doi

  30. [39]

    Huatong Song, Jinhao Jiang, Yingqian Min, Jie Chen, Zhipeng Chen, Wayne Xin Zhao, Lei Fang, and Ji-Rong Wen. 2025. https://arxiv.org/abs/2503.05592 R1-searcher: Incentivizing the search capability in llms via reinforcement learning . Preprint, arXiv:2503.05592

    2025 arXiv

  31. [40]

    Smith, Luke Zettlemoyer, and Tao Yu

    Hongjin Su, Weijia Shi, Jungo Kasai, Yizhong Wang, Yushi Hu, Mari Ostendorf, Wen-tau Yih, Noah A. Smith, Luke Zettlemoyer, and Tao Yu. 2023. https://doi.org/10.18653/v1/2023.findings-acl.71 One embedder, any task: Instruction-finetuned text embeddings . In Findings of the Asso...

    2023 doi

  32. [41]

    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. 2024. https://arxiv.org/abs/2407.12883 Bright: A realistic and challenging ...

    2024 arXiv

  33. [42]

    Hieu Tran, Zonghai Yao, Junda Wang, Yifan Zhang, Zhichao Yang, and Hong Yu. 2024. https://arxiv.org/abs/2412.02830 Rare: Retrieval-augmented reasoning enhancement for large language models . Preprint, arXiv:2412.02830

    2024 arXiv

  34. [43]

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. 2023. https://arxiv.org/abs/2212.10509 Interleaving retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions . Preprint, arXiv:2212.10509

    2023 arXiv

  35. [44]

    Voyage. 2024. Voyage embedding models. https://docs.voyageai.com/docs/embeddings

    2024

  36. [45]

    Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, Daxin Jiang, Rangan Majumder, and Furu Wei. 2023 a . https://arxiv.org/abs/2207.02578 Simlm: Pre-training with representation bottleneck for dense passage retrieval . Preprint, arXiv:2207.02578

    2023 arXiv

  37. [46]

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2024. https://doi.org/10.18653/v1/2024.acl-long.642 Improving text embeddings with large language models . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguisti...

    2024 doi

  38. [47]

    Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H. Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023 b . https://openreview.net/forum?id=1PL1NIMMrw Self-consistency improves chain of thought reasoning in language models . In The Eleventh International Confer...

    2023

  39. [48]

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

    2025 arXiv

  40. [49]

    Yuan Xia, Jingbo Zhou, Zhenhui Shi, Jun Chen, and Haifeng Huang. 2025. https://doi.org/10.1609/aaai.v39i24.34743 Improving retrieval augmented language model with self-reasoning . Proceedings of the AAAI Conference on Artificial Intelligence, 39(24):25534--25542

    2025 doi

  41. [50]

    Chenghao Xiao, G Thomas Hudson, and Noura Al Moubayed. 2024. https://arxiv.org/abs/2404.06347 Rar-b: Reasoning as retrieval benchmark . Preprint, arXiv:2404.06347

    2024 arXiv

  42. [51]

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

    2024 arXiv

  43. [52]

    Griffiths, Yuan Cao, and Karthik R Narasimhan

    Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Thomas L. Griffiths, Yuan Cao, and Karthik R Narasimhan. 2023 a . https://openreview.net/forum?id=5Xc1ecxO1h Tree of thoughts: Deliberate problem solving with large language models . In Thirty-seventh Conference on Neural Infor...

    2023

  44. [53]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023 b . ReAct : Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR)

    2023

  45. [54]

    Wei Zhong, Jheng-Hong Yang, Yuqing Xie, and Jimmy Lin. 2022. https://doi.org/10.18653/v1/2022.findings-emnlp.78 Evaluating token-level and passage-level dense retrieval models for math information retrieval . In Findings of the Association for Computational Linguistics: EMNLP ...

    2022 doi

  46. [55]

    Wei Zhong, Xinyu Zhang, Ji Xin, Richard Zanibbi, and Jimmy Lin. 2021. Approach zero and anserini at the clef-2021 arqmath track: Applying substructure search and bm25 on operator tree path tokens. Proc. CLEF 2021 (CEUR Working Notes)

    2021

Pith tools

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