Pith. sign in

REVIEW 5 major objections 5 minor 54 references

Holistically Guided Monte Carlo Tree Search for Intricate Information Seeking

T0 review · 5 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read HG-MCTS claims that a checklist-guided Monte Carlo tree search collects more complete evidence and answers multi-hop queries more accurately than existing web-search and RAG baselines.

desk verdict Real idea, confounded evaluation: the adaptive-checklist/MCTS combination is worth a look, but the 'beats all baselines' claim rests on comparing live Google search to DPR/Wikipedia and on a Table 1 that doesn't say what the abstract says. read the letter →

arxiv 2502.04751 v1 pith:P7UFJOUT submitted 2025-02-07 cs.IR cs.CL

Ruiyang Ren , Yuhao Wang , Junyi Li , Jinhao Jiang , Wayne Xin Zhao , Wenjie Wang , Tat-Seng Chua This is my paper · ORCID
classification cs.IRcs.CL
keywords MonteCarlotreesearchintricateinformationseekingmulti-hopquestionansweringadaptivechecklistrewardmodelingretrieval-augmentedgenerationwebprogressfeedback
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

Intricate information seeking tasks—questions with several facets that must be answered by combining facts from different sources—need search processes that are both complete and focused. The paper introduces HG-MCTS, which reformulates the task as progressive information collection into a knowledge memory and runs Monte Carlo tree search under an adaptive checklist of sub-goals. Its claim is that this checklist, together with multi-perspective rewards (exploration reward, retrieval reward, and textual progress feedback), prevents MCTS from getting lost in local branches and from omitting crucial aspects of the query. On five multi-hop and multi-document datasets and across three backbone LLMs, it reports significantly more accurate final answers than closed-book, chain-of-thought, RAG, ReAct, Query2doc, Self-RAG, MindSearch, and Infogent. A sympathetic reader would care because it offers a concrete way to make LLM search assistants comprehensive and explainable rather than answer-focused.

What carries the argument

The engine is the pair formed by the adaptive checklist and the multi-perspective reward model, running inside the standard four MCTS steps. The checklist is an LLM-generated sequence of atomic sub-goals, e.g., for a question about a city's land area: identify the company, its founder, the founder's birth city, the country, then the land area; it narrows the branching of the search tree. Rewards are computed at each expanded node: exploration reward $r_q \in \{0,1\}$ for whether the sub-query targets an unfinished sub-goal, retrieval reward $r_k \in \{0,1,2\}$ for whether the retrieved knowledge snippet satisfies the sub-query, combined as $r = r_q \cdot r_k$ and backpropagated through the UCT visit-count update. The progress feedback $u$ is a text output listing completed and remaining sub-goals, used to update the checklist via $p = \text{Update}(p,u;\mathcal{M}_{\theta})$ and to recognise when all sub-goals are done.

What would settle it

Reproduce the evaluation with the retrieval backend equalized—run HG-MCTS with DPR over the Wikipedia corpus used by Standard RAG, and run the baselines with the same Google top-3 retriever; if the accuracy gap collapses when the sources are the same, the central claim that holistic guidance drives the improvement is falsified.

Watch

Extended reading notes

Core claim

The central discovery is a mechanism for giving Monte Carlo tree search a global sense of the information need: before search, the policy LLM generates a natural-language checklist of atomic sub-goals; during expansion, the model proposes sub-queries conditioned on that checklist, the running progress feedback, and the knowledge memory; and during evaluation, a reward model produces an exploration reward in {0,1}, a retrieval reward in {0,1,2}, and free-text progress feedback that updates the checklist and can terminate the search when every sub-goal is marked complete. The paper argues that this synergy turns MCTS from a locally guided exploration into a goal-directed collector of evidence, and reports that the resulting method outperforms all baselines on HotpotQA, 2WikiMultihopQA, MuSiQue, StrategyQA, and FanOutQA, with each component of the guidance confirmed by ablation.

Load-bearing premise

The load-bearing premise is that the observed gains come from the checklist and reward guidance rather than from HG-MCTS using Google web search (top-3 results) while key baselines retrieve from a Wikipedia corpus with DPR.

Editorial extensions

If this is right

  • Multi-step question answering can be made more complete by making the search process answer an explicit sub-goal checklist rather than freely exploring the web.
  • Progress feedback lets the system adapt its checklist online, so queries whose sub-goals are not determinable in advance, such as 'How many studio albums did this band release?', can still be handled.
  • The scaling experiment shows that answer quality improves with more MCTS simulations up to a saturation point, so test-time compute can be traded against accuracy in deployment.
  • The approach stays effective across different backbone LLMs, including an open-source model, which suggests the guidance mechanism does not depend on a particular proprietary model.
  • Each element of the guidance—exploration reward, retrieval reward, progress feedback, and the checklist itself—contributes to the reported gains, since removing any one degrades accuracy on FanOutQA.

Reading between the lines

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

  • A testable extension is to distill the checklist and reward signals into a single-pass retrieval planner; if the guidance mechanism is the real driver, a distilled version should retain most of the accuracy at a fraction of the inference cost.
  • If the gains survive equalized retrieval backends, the checklist mechanism could be ported to non-MCTS iterative retrievers, implying the value lies in explicit sub-goal tracking rather than tree search per se.
  • The paper's recall analysis suggests the benefit is not merely collecting more documents but collecting the right ones; measuring precision and noise in the knowledge memory across methods would sharpen that claim.
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

5 major / 5 minor

Summary. The paper proposes HG-MCTS, a Monte Carlo tree search framework for multi-hop web information seeking. The method maintains a knowledge memory, constructs an adaptive checklist of sub-goals, and uses an LLM-based reward model that returns exploration and retrieval rewards plus textual progress feedback to guide node expansion. The authors evaluate on five multi-hop QA benchmarks with three backbone LLMs, report EM/CEM/F1 or accuracy/ROUGE metrics, and include ablations on FanoutQA. The central claim is that HG-MCTS acquires more comprehensive information and produces more accurate final answers than existing baselines.

Significance. If the central claim were established, the paper would offer a useful test-time search recipe for multi-step retrieval: the adaptive checklist and progress feedback are plausible mechanisms for reducing redundant expansions, and the internal ablation in Table 3 suggests each component contributes. The paper also spans five datasets and three backbone models, which is a broad comparison. However, the current evaluation does not yet establish the central claim because the main comparison varies the retrieval backend together with the reasoning method, and the numerical results contradict the blanket 'all baselines' statement. With controlled retrieval and statistical rigor, the contribution could be meaningful.

major comments (5)
  1. [Section 4.1.3-4.1.4, Tables 1-2] Section 4.1.4 states that HG-MCTS uses Google Search and selects the top-3 web results, while Section 4.1.3 states that Standard RAG retrieves from the Wikipedia corpus with DPR; the retrieval backends for ReAct, Query2doc, and Self-RAG are not specified, so the comparison in Tables 1 and 2 varies both the reasoning method and the retrieval index. Any observed gain could therefore be due to the web index rather than the checklist and reward guidance. The w/o HG-MCTS ablation in Table 3 is a useful within-backend control, but it does not by itself support the abstract's and Section 4.2's comparisons against existing baselines. Please re-run the baselines on the same Google-search backend, or run HG-MCTS on a fixed corpus such as DPR/Wikipedia, and report both.
  2. [Table 1 vs Section 4.2] Section 4.2 states 'significant improvements over all baseline approaches across four multi-hop QA datasets', but Table 1 shows Query2doc outperforming HG-MCTS on HotpotQA for GPT-4o-mini (EM 44.71 vs 41.76; CEM 57.21 vs 45.88) and for DeepSeek-V3-chat (EM 47.19 vs 45.88). The claim must be corrected to specify the metric and dataset, or the discrepancy resolved.
  3. [Section 4.1.2, Tables 1-3] The '†' in Tables 1 and 2 asserts p<0.05 against all baselines, but the paper gives no test type, no variance estimates, no number of seeds, and only 130 sampled queries per dataset (Section 4.1.2); with a single sample and no error bars the significance claim is unverifiable. Table 3 also reports ablation differences without significance or variance. Please report multiple seeds, standard deviations, and the exact significance test.
  4. [Section 3.3, Equations (3), (5), (10)-(12)] The same policy model M_theta is used to generate the checklist, subqueries, exploration reward, retrieval reward, and progress feedback, with no independent validation of the reward model. This creates a risk that the guidance signals merely reflect the model's prior beliefs, so the contribution of the reward/checklist mechanism cannot be cleanly separated from the base model's search ability. An independent reward model, human-judged reward accuracy, or an oracle-reward ablation would address this concern.
  5. [Sections 4.1.3-4.1.4] Section 4.1.3 says some baselines are not reported on every dataset, and Section 4.1.4 refers to prompts in 'provided anonymous codes' without including them in the manuscript; this omits the exact prompts, retrieval engines for each baseline, and the reward-model instruction set, all of which are needed to reproduce the central results.
minor comments (5)
  1. [Section 4.1.3] The heading 'Vinilla reasoning' should be 'Vanilla reasoning'.
  2. [Section 4.3, bullet (b)] The bullet reads 'demonstrating incorporating exploration rewards enables the model...', but the ablated component is the retrieval reward; this appears to be a typo and should say 'retrieval rewards'.
  3. [Figure 5] The axis label 'Accuary' should be 'Accuracy', and the figure would benefit from error bars or confidence intervals.
  4. [Section 5.1] The word 'interoperability' is used twice in the first paragraph; the intended term is likely 'interpretability', which is also used in the paper's introduction and conclusion.
  5. [Section 4.1.2] For FanOutQA the paper says it uses 'string accuracy' but Table 2 reports 'Acc.'; please clarify whether this is exact string match or token-level accuracy.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: core claims are evaluated against external ground truth and a same-backend ablation; the main weaknesses are experimental confounds and internal self-scoring, not equation-level circularity.

full rationale

The paper is an empirical systems paper; it does not derive a formal prediction from an identity. The central comparison in Tables 1 and 2 is against external ground-truth labels (EM, F1, CEM, ROUGE), so the headline claim cannot reduce to the method's own checklist or reward definitions. The internal guidance loop uses the same LLM for checklist generation, subquery generation, reward scoring, and feedback, which is a potential self-validation or attribution confound, but it is not definitional: the loop's contribution is tested by ablations (Table 3: w/o Checklist, w/o Progress Feedback, w/o HG-MCTS) on an external benchmark, so the claim is falsifiable. The paper contains minor self-citations, e.g., references [9] and [10] by overlapping authors, used for framing and for the 130-query sampling setup, but these are not the load-bearing justification of the central result. The larger validity issue is the retrieval-backend confound: HG-MCTS uses Google Search top-3 web results (Section 4.1.4) while Standard RAG uses DPR over Wikipedia (Section 4.1.3), so part of the gain may be attributable to the backend rather than the method; that is an experimental confound, not evidence of a circular derivation. Table 1 also shows Query2doc beating HG-MCTS on HotpotQA under GPT-4o-mini, contradicting the blanket 'all baselines' wording, but again this is an accuracy/consistency issue, not circularity. No fitted parameter is renamed as a prediction, and no uniqueness theorem or ansatz is smuggled in via self-citation. Overall, no significant circularity.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The method relies on hand-picked hyperparameters (UCT weight, depth, simulation count, subqueries per node), on the assumption that the same LLM can generate reliable checklists and self-scored rewards, and on the coverage of the Google Search top-3 results. No independent validation is provided for the reward signals.

free parameters (6)
  • UCT exploration weight w = 0.2
    Set in Section 4.1.4 without sensitivity analysis.
  • Maximum MCTS simulations = 40
    Set in Section 4.1.4; scaling analysis shows diminishing returns but no optimality argument.
  • Search depth limit = 6
    Set in Section 4.1.4.
  • Sub-queries per expansion m_q = 3
    Set in Section 4.1.4.
  • Number of web results per query = 3 (top-3)
    Set in Section 4.1.4.
  • Decoding temperature and top-p = temperature=0.9, top_p=1.0
    Set in Section 4.1.4 for all generation tasks.
assumptions (5)
  • domain assumption The same LLM can serve as policy, checklist generator, reward model, and final answer generator without reward validation.
    Sections 3.2 and 3.3 use M_theta and R_theta with R_theta consistent with the policy; no evidence that these self-scores correlate with answer quality.
  • domain assumption The LLM-generated adaptive checklist correctly enumerates the sub-goals needed to answer the query.
    Section 3.2.1 generates the checklist from the model's internal knowledge; errors here propagate to all subsequent search guidance.
  • domain assumption Google Search top-3 results contain sufficient information to answer the benchmark questions.
    Section 4.1.4 selects top-3 web results; no analysis of recall or coverage of this retrieval source.
  • domain assumption Ground-truth Wikipedia pages in each benchmark are the correct reference for measuring information collection.
    Section 4.4 computes recall against annotated Wikipedia pages; web pages may contain equivalent information not in the annotation set.
  • standard math UCT with a fixed exploration weight provides a reasonable balance for this search problem.
    Equation 2 uses standard UCT; the choice of w=0.2 is a hyperparameter, not derived.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Holistically Guided Monte Carlo Tree Search for Intricate Information Seeking." pith.science (2026). https://pith.science/paper/P7UFJOUT

@misc{pith2026250204751,
  author       = {Pith},
  title        = {Pith review of: Holistically Guided Monte Carlo Tree Search for Intricate Information Seeking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P7UFJOUT}},
  note         = {Machine review of arXiv:2502.04751}
}
read the original abstract

In the era of vast digital information, the sheer volume and heterogeneity of available information present significant challenges for intricate information seeking. Users frequently face multistep web search tasks that involve navigating vast and varied data sources. This complexity demands every step remains comprehensive, accurate, and relevant. However, traditional search methods often struggle to balance the need for localized precision with the broader context required for holistic understanding, leaving critical facets of intricate queries underexplored. In this paper, we introduce an LLM-based search assistant that adopts a new information seeking paradigm with holistically guided Monte Carlo tree search (HG-MCTS). We reformulate the task as a progressive information collection process with a knowledge memory and unite an adaptive checklist with multi-perspective reward modeling in MCTS. The adaptive checklist provides explicit sub-goals to guide the MCTS process toward comprehensive coverage of complex user queries. Simultaneously, our multi-perspective reward modeling offers both exploration and retrieval rewards, along with progress feedback that tracks completed and remaining sub-goals, refining the checklist as the tree search progresses. By striking a balance between localized tree expansion and global guidance, HG-MCTS reduces redundancy in search paths and ensures that all crucial aspects of an intricate query are properly addressed. Extensive experiments on real-world intricate information seeking tasks demonstrate that HG-MCTS acquires thorough knowledge collections and delivers more accurate final responses compared with existing baselines.

Figures

Figures reproduced from arXiv: 2502.04751 by the authors.

Figure 1
Figure 1. Illustration of the pitfalls in handling intricate [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overall framework of the proposed HG-MCTS method. The left panel outlines the iterative Monte Carlo tree [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. An illustration of the checklist corresponding to an [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Evaluation results of HG-MCTS with various simu [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 22 canonical work pages

  1. [1]

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2024. Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection. In The Twelfth International Conference on Learning Representations

    work page 2024

  2. [2]

    Cameron B Browne, Edward Powley, Daniel Whitehouse, Simon M Lucas, Peter I Cowling, Philipp Rohlfshagen, Stephen Tavener, Diego Perez, Spyridon Samoth- rakis, and Simon Colton. 2012. A survey of monte carlo tree search methods. IEEE Transactions on Computational Intelligence and AI in games 4, 1 (2012), 1–43

    2012

  3. [3]

    Guoxin Chen, Minpeng Liao, Chengxi Li, and Kai Fan. 2024. AlphaMath Almost Zero: process Supervision without process.arXiv preprint arXiv:2405.03553 (2024)

    arXiv 2024

  4. [4]

    Zehui Chen, Kuikun Liu, Qiuchen Wang, Jiangning Liu, Wenwei Zhang, Kai Chen, and Feng Zhao. 2024. Mindsearch: Mimicking human minds elicits deep ai searcher. arXiv preprint arXiv:2407.20183 (2024)

    arXiv 2024

  5. [5]

    Mor Geva, Daniel Khashabi, Elad Segal, Tushar Khot, Dan Roth, and Jonathan Berant. 2021. Did aristotle use a laptop? a question answering benchmark with implicit reasoning strategies. Transactions of the Association for Computational Linguistics 9 (2021), 346–361

    2021

  6. [6]

    Peter E Hart, Nils J Nilsson, and Bertram Raphael. 1968. A formal basis for the heuristic determination of minimum cost paths. IEEE transactions on Systems Science and Cybernetics 4, 2 (1968), 100–107

    1968

  7. [7]

    Xanh Ho, Anh-Khoa Duong Nguyen, Saku Sugawara, and Akiko Aizawa. 2020. Constructing A Multi-hop QA Dataset for Comprehensive Evaluation of Reason- ing Steps. In Proceedings of the 28th International Conference on Computational Linguistics. 6609–6625

    work page 2020

  8. [8]

    Mohanna Hoveyda, Arjen P de Vries, Maarten de Rijke, Harrie Oosterhuis, and Faegheh Hasibi. 2024. AQA: Adaptive question answering in a society of LLMs via contextual multi-armed bandit. arXiv preprint arXiv:2409.13447 (2024)

    work page Pith review arXiv 2024

Show all 54 references
  1. [9]

    Jinhao Jiang, Jiayi Chen, Junyi Li, Ruiyang Ren, Shijie Wang, Wayne Xin Zhao, Yang Song, and Tao Zhang. 2024. RAG-Star: Enhancing Deliberative Reason- ing with Retrieval Augmented Verification and Refinement. arXiv preprint arXiv:2412.12881 (2024)

    2024 arXiv

  2. [10]

    Jinhao Jiang, Zhipeng Chen, Yingqian Min, Jie Chen, Xiaoxue Cheng, Jiapeng Wang, Yiru Tang, Haoxiang Sun, Jia Deng, Wayne Xin Zhao, et al. 2024. Technical Report: Enhancing LLM Reasoning with Reward-guided Tree Search. arXiv preprint arXiv:2411.11694 (2024)

    2024 arXiv

  3. [11]

    Daniel Kahneman. 2011. Thinking, fast and slow. Farrar, Straus and Giroux (2011)

    2011

  4. [12]

    Jikun Kang, Xin Zhe Li, Xi Chen, Amirreza Kazemi, Qianyi Sun, Boxing Chen, Dong Li, Xu He, Quan He, Feng Wen, et al . 2024. Mindstar: Enhancing math reasoning in pre-trained llms at inference time. arXiv preprint arXiv:2405.16265 (2024)

    2024 arXiv

  5. [13]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense Passage Retrieval for Open- Domain Question Answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP...

    2020

  6. [14]

    Tushar Khot, Harsh Trivedi, Matthew Finlayson, Yao Fu, Kyle Richardson, Peter Clark, and Ashish Sabharwal. 2023. Decomposed Prompting: A Modular Ap- proach for Solving Complex Tasks. In The Eleventh International Conference on Learning Representations

    2023

  7. [15]

    Levente Kocsis and Csaba Szepesvári. 2006. Bandit Based Monte-Carlo Planning. In Machine Learning: ECML 2006, 17th European Conference on Machine Learning, Berlin, Germany, September 18-22, 2006, Proceedings (Lecture Notes in Computer Science, Vol. 4212), Johannes Fürnkranz, T...

    2006

  8. [16]

    Seongmin Lee, Jaewook Shin, Youngjin Ahn, Seokin Seo, Ohjoon Kwon, and Kee-Eung Kim. 2024. Zero-Shot Multi-Hop Question Answering via Monte-Carlo Tree Search with Large Language Models. arXiv preprint arXiv:2409.19382 (2024)

    2024 arXiv

  9. [17]

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2024. Let’s Verify Step by Step. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, M...

    2024

  10. [18]

    Chin-Yew Lin. 2004. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out. 74–81

    2004

  11. [19]

    Liangchen Luo, Yinxiao Liu, Rosanne Liu, Samrat Phatale, Harsh Lara, Yunxuan Li, Lei Shu, Yun Zhu, Lei Meng, Jiao Sun, et al . 2024. Improve Mathematical Reasoning in Language Models by Automated Process Supervision.arXiv preprint arXiv:2406.06592 (2024)

    2024 arXiv

  12. [20]

    Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al

  13. [21]

    Pranav Putta, Edmund Mills, Naman Garg, Sumeet Motwani, Chelsea Finn, Di- vyansh Garg, and Rafael Rafailov. 2024. Agent q: Advanced reasoning and learning for autonomous ai agents. arXiv preprint arXiv:2408.07199 (2024)

    2024 arXiv

  14. [22]

    Revanth Gangi Reddy, Sagnik Mukherjee, Jeonghwan Kim, Zhenhailong Wang, Dilek Hakkani-Tur, and Heng Ji. 2024. Infogent: An agent-based framework for web information aggregation. arXiv preprint arXiv:2410.19054 (2024)

    2024 arXiv

  15. [23]

    Amrith Setlur, Chirag Nagpal, Adam Fisch, Xinyang Geng, Jacob Eisenstein, Rishabh Agarwal, Alekh Agarwal, Jonathan Berant, and Aviral Kumar. 2024. Rewarding Progress: Scaling Automated Process Verifiers for LLM Reasoning. arXiv preprint arXiv:2410.08146 (2024)

    2024 arXiv

  16. [24]

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2024. Reflexion: Language agents with verbal reinforcement learning. Advances in Neural Information Processing Systems 36 (2024). Holistically Guided Monte Carlo Tree Search for Intricate Inform...

    2024

  17. [25]

    David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Laurent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershel- vam, Marc Lanctot, et al . 2016. Mastering the game of Go with deep neural networks and tree search. nature 529, 7587 (2...

    2016

  18. [26]

    David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Vedavyas Pan- neershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy P. Lillicra...

    2016

  19. [27]

    David Silver, Thomas Hubert, Julian Schrittwieser, Ioannis Antonoglou, Matthew Lai, Arthur Guez, Marc Lanctot, Laurent Sifre, Dharshan Kumaran, Thore Graepel, et al. 2017. Mastering chess and shogi by self-play with a general reinforcement learning algorithm. arXiv preprint ar...

    2017 arXiv

  20. [28]

    Charlie Snell, Jaehoon Lee, Kelvin Xu, and Aviral Kumar. 2024. Scaling llm test- time compute optimally can be more effective than scaling model parameters. arXiv preprint arXiv:2408.03314 (2024)

    2024 arXiv

  21. [29]

    Trevor Strohman, Howard Turtle, and W Bruce Croft. 2005. Optimization strate- gies for complex queries. In Proceedings of the 28th annual international acm SIGIR conference on research and development in Information Retrieval . 219–225

    2005

  22. [30]

    Richard Sutton. 2019. The bitter lesson. Incomplete Ideas (blog) 13, 1 (2019), 38

    2019

  23. [31]

    Alon Talmor and Jonathan Berant. 2018. The Web as a Knowledge-Base for Answering Complex Questions. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) . 641–651

    2018

  24. [33]

    Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal

  25. [34]

    Chaojie Wang, Yanchen Deng, Zhiyi Lv, Zeng Liang, Jujie He, Shuicheng Yan, and Bo An. 2024. Q*: Improving Multi-step Reasoning for LLMs with Deliberative Planning. CoRR abs/2406.14283 (2024). https://doi.org/10.48550/ARXIV.2406. 14283 arXiv:2406.14283

    2024 doi arXiv

  26. [35]

    Chaojie Wang, Yanchen Deng, Zhiyi Lyu, Liang Zeng, Jujie He, Shuicheng Yan, and Bo An. 2024. Q*: Improving multi-step reasoning for llms with deliberative planning. arXiv preprint arXiv:2406.14283 (2024)

    2024 arXiv

  27. [36]

    Liang Wang, Nan Yang, and Furu Wei. 2023. Query2doc: Query Expansion with Large Language Models. In The 2023 Conference on Empirical Methods in Natural Language Processing

    2023

  28. [37]

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. [n. d.]. Self-Consistency Improves Chain of Thought Reasoning in Language Models. In The Eleventh International Conference on Learning Representations

  29. [38]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems 35 (2022), 24824–24837

    2022

  30. [39]

    Shicheng Xu, Liang Pang, Huawei Shen, Xueqi Cheng, and Tat-Seng Chua

  31. [40]

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language...

    2018

  32. [41]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. ReAct: Synergizing Reasoning and Acting in Language Models. In International Conference on Learning Representations (ICLR)

    2023

  33. [42]

    Weirui Ye, Shaohuai Liu, Thanard Kurutach, Pieter Abbeel, and Yang Gao. 2021. Mastering Atari Games with Limited Data. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, v...

    2021

  34. [43]

    Eric Zelikman, Georges Harik, Yijia Shao, Varuna Jayasiri, Nick Haber, and Noah D. Goodman. 2024. Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking. CoRR abs/2403.09629 (2024). https://doi.org/10.48550/ ARXIV.2403.09629 arXiv:2403.09629

    2024 doi arXiv

  35. [44]

    Di Zhang, Jianbo Wu, Jingdi Lei, Tong Che, Jiatong Li, Tong Xie, Xiaoshui Huang, Shufei Zhang, Marco Pavone, Yuqiang Li, et al. 2024. LLaMA-Berry: Pairwise Optimization for O1-like Olympiad-Level Mathematical Reasoning.arXiv preprint arXiv:2410.02884 (2024)

    2024 arXiv

  36. [45]

    Dan Zhang, Sining Zhoubian, Ziniu Hu, Yisong Yue, Yuxiao Dong, and Jie Tang

  37. [46]

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, et al. 2023. A survey of large language models. arXiv preprint arXiv:2303.18223 (2023)

    2023 arXiv

  38. [47]

    Andy Zhou, Kai Yan, Michal Shlapentokh-Rothman, Haohan Wang, and Yu- Xiong Wang. [n. d.]. Language Agent Tree Search Unifies Reasoning, Acting, and Planning in Language Models. In Forty-first International Conference on Machine Learning

  39. [48]

    Andrew Zhu, Alyssa Hwang, Liam Dugan, and Chris Callison-Burch. 2024. FanOutQA: A multi-hop, multi-document question answering benchmark for large language models. In Proceedings of the 62nd Annual Meeting of the Associa- tion for Computational Linguistics (Volume 2: Short Pap...

    2024

  40. [49]

    Maciej Świechowski, Konrad Godlewski, Bartosz Sawicki, and Jacek Mańdziuk

  41. [50]

    arXiv preprint arXiv:2406.03816 (2024)

    Rest-mcts*: Llm self-training via process reward guided tree search. arXiv preprint arXiv:2406.03816 (2024)

    2024 arXiv

  42. [55]

    Artificial Intelligence Review 56, 3 (July 2022), 2497–2562

    Monte Carlo Tree Search: a review of recent modifications and applications. Artificial Intelligence Review 56, 3 (July 2022), 2497–2562. https://doi.org/10.1007/ s10462-022-10228-y

    2022

  43. [2021]

    Webgpt: Browser-assisted question-answering with human feedback.arXiv preprint arXiv:2112.09332 (2021)

    2021 arXiv

  44. [2022]

    Transactions of the Association for Computational Linguistics 10 (2022), 539–554

    MuSiQue: Multi-hop Questions via Single-hop Question Composition. Transactions of the Association for Computational Linguistics 10 (2022), 539–554

    2022

  45. [2023]

    In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Interleaving Retrieval with Chain-of-Thought Reasoning for Knowledge- Intensive Multi-Step Questions. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) . 10014–10037

  46. [2024]

    In Proceedings of the ACM on Web Conference 2024

    Search-in-the-Chain: Interactively Enhancing Large Language Models with Search for Knowledge-intensive Tasks. In Proceedings of the ACM on Web Conference 2024. 1362–1373

    2024

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.