Pith. sign in

REVIEW 4 major objections 5 minor 41 references

MASTER: A Multi-Agent System with LLM Specialized MCTS

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A multi-agent system replaces MCTS simulations with LLM self-evaluation, reporting 76.0% exact match on HotpotQA and 80.0% on WebShop, new state-of-the-art results that use about 6% of the tokens of a prior tree-search agent.

desk verdict The mechanism is more interesting than the numbers, and the numbers as presented do not yet support the SOTA claims. read the letter →

arxiv 2501.14304 v2 pith:EVQGUZIE submitted 2025-01-24 cs.AI

Bingzheng Gan , Yufan Zhao , Tianyi Zhang , Jing Huang , Yusu Li , Shu Xian Teo , Changwang Zhang , Wei Shi This is my paper · ORCID
classification cs.AI
keywords largelanguagemodelsMonteCarlotreesearchmulti-agentsystemsself-evaluationconfidenceweightingquestionansweringdecisionmakingcodegeneration
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

The paper sets out to remove the main blocker to Monte Carlo tree search for language models: the need for many simulations and for objective rewards. MASTER replaces the simulation step with a single LLM-driven pass in which an agent produces a solution, validates its own facts, and scores its progress together with a confidence value. That self-assigned reward feeds a modified UCT formula and backpropagation, and the tree doubles as an agent-recruitment and communication protocol: child agents continue their parent's context, and a terminal agent that passes an LLM evaluation ends the task. The authors report 76.0% exact match on HotpotQA and 80.0% accuracy on WebShop, which they call new state-of-the-art, and they measure about 10,937 tokens per HotpotQA question versus 185,392 for LATS. If these claims hold, tree-structured LLM agents can plan without ground-truth rewards or expensive rollouts, and task complexity automatically decides how many agents get recruited.

What carries the argument

The load-bearing object is a reasoning tree where each node is an agent carrying Solution, Validation, and Assessment, and where selection is governed by a modified UCT formula. With $n_i$ backpropagations to agent $i$, initial reward $r_0$, confidence $c_0$, and parent visit count $N_i$, the value used for selection is $UCT = r_0$ when $n_i=0$, and otherwise $$UCT = c_0 r_0 + (1-c_0)\frac{1}{n_i}\sum_{n=1}^{n_i} r_n + \frac{1}{10\sqrt{2c_0}} \sqrt{\frac{\ln N_i}{n_i}}.$$ The confidence $c_0$ appears twice: it reweights the initial reward against backpropagated updates, and it inversely scales the exploration bonus through the factor $1/(10\sqrt{2c_0})$, so low-confidence branches are explored more. A terminal agent that fails evaluation triggers backpropagation that updates every ancestor's reward and visit count, and one that passes evaluation terminates the task, which is the mechanism that makes token cost adapt to task difficulty.

What would settle it

A decisive test is to take the same 100-question HotpotQA and WebShop samples, corrupt the LLM's self-assigned scores and confidence values with increasing random noise while freezing every other component, and measure how accuracy moves: if the reported 76.0% and 80.0% results barely change under corruption, the tree structure alone is doing the work; if they degrade sharply, the claim that self-evaluation can replace simulation depends on the reliability of the self-scores.

Watch

Extended reading notes

Core claim

MASTER's central claim is that the simulation phase of MCTS can be dropped entirely in LLM settings and replaced by structured self-evaluation. Each node in the tree is an agent that writes a thought, takes an action through a tool, receives an observation, validates the facts in its partial solution, and then assigns itself a score and a confidence in that score; the score acts as the initial reward and the confidence controls both how much the score is trusted against backpropagated updates and how strongly the exploration term pulls selection toward uncertain branches. When a terminal agent states a final answer, the LLM evaluates it: acceptance ends the task, and a failed evaluation backpropagates the terminal reward up the path, lowering the UCT values of the agents that led to it. The paper argues that this design keeps the selection and exploration structure of MCTS while making rewards available without ground truth and making token cost scale with task difficulty, and it reports the framework's measured results of 76.0% exact match on HotpotQA and 80.0% accuracy on WebShop as new state-of-the-art, with roughly 6% of LATS's token consumption.

Load-bearing premise

The load-bearing premise is that the LLM's self-assigned correctness scores, confidence values, and terminal pass/fail judgments are accurate enough to steer both the tree selection and the early-stop decision; if these are biased or noisy, the UCT selection and early termination go wrong no matter how carefully the tree is built.

Editorial extensions

If this is right

  • If MASTER's numbers reproduce, HotpotQA's best reported exact match would move from 73.3% (Beam Retrieval) to 76.0%.
  • WebShop's best reported accuracy would move from 70.2% (AgentKit) to 80.0%, a much larger jump than HotpotQA's.
  • At 10,937 tokens per HotpotQA question against LATS's 185,392, a tree-based LLM planner would become roughly 17 times cheaper at higher measured accuracy, which changes the cost calculus for deploying search-based agents.
  • Because agents are task-agnostic and only take different actions, the same MASTER code transfers across question answering, web shopping, and coding without reconfiguring agent roles.
  • Ablations show the reward-assignment chain is responsible for most of the gain: removing validation drops HotpotQA from 76.0% to 62.3% and WebShop from 80.0% to 56.3%, while replacing assessment with random rewards drops HotpotQA to 23.3% and WebShop to 15.7%.

Reading between the lines

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

  • Editorial inference: the same confidence-weighted UCT update could be applied to other LLM search settings, such as code repair or theorem proving, where a single deterministic evaluator replaces rollouts and the confidence term reports a calibrated uncertainty.
  • Editorial inference: if the self-evaluation scores are well calibrated, the acceptance threshold at the terminal node could be tuned continuously instead of using a binary pass/fail prompt, which might recover additional accuracy or cut tokens further on easy questions.
  • Editorial inference: because the paper evaluates on 100-question subsamples, the full-test-set version of the WebShop claim is the most direct way to test whether the 80.0% accuracy is stable; the web-shopping margin over the prior 70.2% is particularly worth checking on the whole benchmark.
  • Editorial inference: the ablation pattern suggests the validation step is what stabilizes reward assignment, so a useful extension is to measure how performance degrades when the same validation module is replaced by a weaker or faster model, which would show how much of the result depends on GPT-4's evaluative strength.
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

4 major / 5 minor

Summary. The paper presents MASTER, a multi-agent framework that replaces the simulation step of MCTS with LLM self-evaluation, validation, and confidence-weighted rewards. Agents are organized in a reasoning tree, selected by a modified UCT formula, and backpropagated when terminal answers fail an LLM-based evaluation. The authors report 76.0% exact match on HotpotQA, 80.0% accuracy on WebShop, and 91.0% pass@1 on MBPP, claiming new state-of-the-art results on HotpotQA and WebShop while using about 6% of LATS's token budget. The paper includes ablation studies, parameter studies, pseudo-code, and detailed case studies.

Significance. If the empirical claims are substantiated, the core idea is significant: it shows that MCTS-style search can be driven by LLM self-evaluation rather than by ground-truth rewards, and that dynamic agent recruitment guided by such a search can be both effective and token-efficient. The efficiency comparison against LATS is a useful and concrete contribution. The paper is also clearly structured, with pseudo-code and case-study traces that make the mechanism understandable. However, the headline SOTA claims rest on comparisons with external, non-reproduced baseline numbers on different evaluation protocols, and all reported results lack error bars or significance tests. The central mechanism also depends on an unmeasured self-verification reliability. These issues currently weaken the evidence for the paper's strongest claims, though they do not refute the underlying approach.

major comments (4)
  1. [§4.2, Table 1] The state-of-the-art claims for HotpotQA and WebShop are not supported by comparable evaluations. Beam Retrieval (0.733) and AgentKit (0.702) are taken from their original papers, while MASTER is evaluated on a randomly selected 100-question sample. The paper explicitly states for AgentKit that "we rely on the original performance claims from their paper." Since the reported MASTER numbers are means over 100 questions, the 2.7-point HotpotQA margin is within sampling noise (standard error of a 0.76 proportion on n=100 is about 0.043), and the WebShop margin cannot be attributed to the algorithm unless AgentKit is run under the identical protocol. Please either reproduce these baselines on the same 100-question sample, or present the results as indicative rather than as new SOTA.
  2. [§4.3, §5, Tables 1–5] All reported accuracies are means of three runs on 100 questions, with no variance, confidence intervals, or significance tests. Many of the differences that drive the ablative conclusions, such as 0.760 vs. 0.737 in Table 2 or 0.760 vs. 0.770 in Table 5, are smaller than the sampling error and are not interpretable as evidence for the contribution of a UCT component. Please report per-run or per-question breakdowns, confidence intervals, and paired statistical tests where relevant, or explicitly temper the claims that depend on these differences.
  3. [§5.4, Table 5] The claimed ability to "autonomously adjust the number of agents based on task complexity" is strongly qualified by the sensitivity to Maximum of Expansion. WebShop accuracy collapses from 0.800 to 0.013 when the limit is set to 3 instead of 8, and the paper adopts task-specific values (3 for HotpotQA and MBPP, 8 for WebShop). This is a manual hyperparameter, not an autonomous mechanism. The contribution statement should be revised to reflect that the expansion depth is task-specific, and ideally the paper should provide guidance or an automatic criterion for setting this parameter.
  4. [§3.2, §7] The framework's terminal acceptance and backpropagation both rely entirely on the LLM's own Evaluation of correctness, and Section 7 acknowledges the reliance on accurate self-assessed scores and confidence. The paper does not measure the reliability of this self-verification—e.g., the false-accept rate, the agreement between the LLM's pass/fail judgment and ground truth, or the impact of replacing the self-evaluation with an oracle. Since the mechanism's interpretation depends on this signal, an analysis of its reliability is needed to support the claim that the tree-search coordination, rather than the self-verifier, drives the gains.
minor comments (5)
  1. [§3.3, Eqns. (1), (4), (5)] The notation for the exploration term is inconsistent: Eqn. (1) has sqrt(ln(N_i)/(2 n_i)), while Eqns. (4) and (5) use sqrt(ln(N_i)/n_i) with a separate multiplier. Please clarify the relationship between the Hoeffding-derived constant and the exploration weight, and add parentheses to resolve expressions like "1/ 10 √ 2c0".
  2. [§4.3] The paper says the same random seed is used across datasets to select 100 questions, but it does not specify the total pool size for WebShop or MBPP or provide the resulting question IDs. Making the sample list available would improve reproducibility.
  3. [Table 2] The row label "Fixed Exploration Weight0.700" is missing a space and a separating symbol; it should read "Fixed Exploration Weight" with the value in the cell. Also consider explaining in the text why the fixed-weight variant underperforms even the no-exploration variant.
  4. [§5.2] The statement that LATS's lower reported token cost is "because their tests were conducted on correctly answered questions" is a claim about the LATS evaluation protocol; it needs a citation or evidence from the LATS paper.
  5. [Appendix C] Several typos appear in the case studies (e.g., "Qustion", "assessement", "informaion", "assessement"). A careful proofread is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central mechanism is an empirically tested design and the SOTA claim rests on external benchmarks, not on self-referential derivation.

full rationale

The core proposal is to replace MCTS simulation rewards with LLM self-evaluation, confidence weighting, validation, and backpropagation. This is an algorithmic design whose effectiveness is measured against external benchmark accuracy, not a quantity defined in terms of those self-evaluations. The modified UCT formula (Eqn. 5) is explicitly introduced as a design choice and is tested by ablations, not derived as a forced consequence of any fitted parameter. The only derivation in the paper, Appendix B, is the standard Hoeffding-based derivation of UCT from prior bandit literature, and it does not depend on the paper's own claims. There are no load-bearing self-citations: the cited baselines and methods are all external works. The reliance on the LLM's self-evaluation for scores and confidence is disclosed in Section 7 as a limitation and is an architectural assumption, not a circular reduction. The use of reported numbers for AgentKit and other baselines is an experimental verification concern about comparable evaluation, not a definitional equivalence between the paper's inputs and its claimed outputs. Therefore no step in the paper's derivation chain reduces to its own inputs by construction.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The framework does not introduce new physical entities. The load-bearing assumptions are that LLM self-evaluation is a reliable reward signal and that the hand-set exploration and depth parameters are appropriate. The exploration constant 10, the confidence clamp 0.1, and the per-dataset maximum expansion depth are free, fitted or hand-picked settings that the reported results depend on.

free parameters (5)
  • exploration_constant_10 = 10
    Appears in Eqn 5 as 1/(10*sqrt(2*c0)). No derivation or sensitivity study is provided; the number is hand-selected to balance exploration in LLM search trees.
  • number_of_branches = 2
    Set to 2 for all datasets after the parameter study in Table 4; chosen as a trade-off between performance and token cost.
  • maximum_of_expansion = 3 for HotpotQA and MBPP, 8 for WebShop
    Chosen per dataset based on the expected number of steps needed to solve the problem (Section 5.4). Table 5 shows that setting it too low collapses performance, indicating the value is effectively tuned to the benchmark.
  • minimum_confidence_c0 = 0.1
    The text (Section 3.3) notes 'when the minimum confidence of 0.1 is used, the exploration weight equals 1/sqrt(2)', implying confidence is clamped to [0.1, 1]; this clamp is not justified.
  • temperature = 0.6 for Thought/Action, 0.0 for Validation/Assessment
    Chosen to encourage diverse reasoning paths while keeping scoring stable; no sensitivity analysis is given.
assumptions (4)
  • standard math Hoeffding's Inequality holds for the LLM-generated reward distribution and justifies the UCT bound.
    Used in Appendix B to derive the original UCT formula; the paper assumes the bound transfers to its modified reward setting.
  • domain assumption LLM self-evaluation scores and confidence values are correlated with true solution quality.
    The whole reward mechanism in Section 3.4 depends on assessment and validation being meaningful. The paper explicitly flags this reliance in Section 7.
  • domain assumption The LLM's terminal evaluation (pass/fail) is reliable enough that accepting a 'pass' early termination does not hurt accuracy.
    In Section 3.2, a terminal agent that passes Evaluation is submitted as the final answer; this assumes the LLM's verdict is usually correct.
  • ad hoc to paper Per-dataset Maximum of Expansion values are known or knowable a priori.
    The paper selects 8 for WebShop and 3 for HotpotQA/MBPP based on the expected number of steps; this is task-specific knowledge fitted to the benchmarks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MASTER: A Multi-Agent System with LLM Specialized MCTS." pith.science (2026). https://pith.science/paper/EVQGUZIE

@misc{pith2026250114304,
  author       = {Pith},
  title        = {Pith review of: MASTER: A Multi-Agent System with LLM Specialized MCTS},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EVQGUZIE}},
  note         = {Machine review of arXiv:2501.14304}
}
read the original abstract

Large Language Models (LLM) are increasingly being explored for problem-solving tasks. However, their strategic planning capability is often viewed with skepticism. Recent studies have incorporated the Monte Carlo Tree Search (MCTS) algorithm to augment the planning capacity of LLM. Despite its potential, MCTS relies on extensive sampling simulations to approximate the true reward distribution, which leads to two primary issues. Firstly, MCTS is effective for tasks like the Game of Go, where simulation results can yield objective rewards (e.g., 1 for a win and 0 for a loss). However, for tasks such as question answering, the result of a simulation is the answer to the question, which cannot yield an objective reward without the ground truth. Secondly, obtaining statistically significant reward estimations typically requires a sample size exceeding 30 simulations, resulting in excessive token usage and time consumption. To address these challenges, we present the Multi-Agent System with Tactical Execution and Reasoning using LLM Specialized MCTS (MASTER), a novel framework that coordinates agent recruitment and communication through LLM specialized MCTS. This system autonomously adjusts the number of agents based on task complexity and ensures focused communication among them. Comprehensive experiments across various tasks demonstrate the effectiveness of our proposed framework. It achieves 76% accuracy on HotpotQA and 80% on WebShop, setting new state-of-the-art performance on these datasets.

Figures

Figures reproduced from arXiv: 2501.14304 by the authors.

Figure 1
Figure 1. Reasoning Tree of MASTER. Starting from Agent0 , Agent1 and Agent2 are created in the first expansion. Then the system first selects Agent1 for expansion due to its higher UCT. Its child agent Agent3 is a terminal agent that failed evaluation which triggers a backpropagation and lowers the UCT of Agent1 . Now Agent2 has the highest UCT and is selected for next expansion. Its child agent, Agent6 is a terminal agent a… view at source ↗
Figure 2
Figure 2. Reasoning Tree of case studies on three datasets. [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 10 canonical work pages

  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]

    Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, and Charles Sutton. 2021. https://arxiv.org/abs/2108.07732 Program synthesis with large language models . Preprint, arXiv:2108.07732

    arXiv 2021

  4. [4]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin...

    arXiv 2020

  5. [5]

    Karlsson, Jie Fu, and Yemin Shi

    Guangyao Chen, Siwei Dong, Yu Shu, Ge Zhang, Jaward Sesay, Börje F. Karlsson, Jie Fu, and Yemin Shi. 2024 a . https://arxiv.org/abs/2309.17288 Autoagents: A framework for automatic agent generation . Preprint, arXiv:2309.17288

    arXiv 2024

  6. [6]

    Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chi-Min Chan, Heyang Yu, Yaxi Lu, Yi-Hsin Hung, Chen Qian, Yujia Qin, Xin Cong, Ruobing Xie, Zhiyuan Liu, Maosong Sun, and Jie Zhou. 2024 b . https://openreview.net/forum?id=EHg5GDnyq1 Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors . In The Twelfth Internat...

    2024

  7. [7]

    Xinyun Chen, Maxwell Lin, Nathanael Sch \"a rli, and Denny Zhou. 2024 c . https://openreview.net/forum?id=KuPixIqPiq Teaching large language models to self-debug . In The Twelfth International Conference on Learning Representations

    work page 2024

  8. [8]

    Zheng Chu, Jingchang Chen, Qianglong Chen, Weijiang Yu, Tao He, Haotian Wang, Weihua Peng, Ming Liu, Bing Qin, and Ting Liu. 2023. https://arxiv.org/abs/2309.15402 A survey of chain of thought reasoning: Advances, frontiers and future . Preprint, arXiv:2309.15402

    arXiv 2023

Show all 41 references
  1. [9]

    Rémi Coulom. 2006. https://inria.hal.science/inria-00116992/document Efficient selectivity and backup operators in monte-carlo tree search . In 5th International Conference on Computer and Games

    2006

  2. [10]

    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

  3. [11]

    Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Ceyao Zhang, Jinlin Wang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, Lingfeng Xiao, Chenglin Wu, and Jürgen Schmidhuber. 2023. https://arxiv.org/abs/2308.00352 Metagpt: Meta program...

    2023 arXiv

  4. [12]

    Zhang, Michael Luck, Qingwen Bu, Yuhao Qing, and Heming Cui

    Dong Huang, Jie M. Zhang, Michael Luck, Qingwen Bu, Yuhao Qing, and Heming Cui. 2024. https://arxiv.org/abs/2312.13010 Agentcoder: Multi-agent-based code generation with iterative testing and optimisation . Preprint, arXiv:2312.13010

    2024 arXiv

  5. [13]

    Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong, Zhangyin Feng, Haotian Wang, Qianglong Chen, Weihua Peng, Xiaocheng Feng, Bing Qin, and Ting Liu. 2023. https://arxiv.org/abs/2311.05232 A survey on hallucination in large language models: Principles, taxonomy, challenges, and ...

    2023 arXiv

  6. [14]

    Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. 2023. https://arxiv.org/abs/2205.11916 Large language models are zero-shot reasoners . Preprint, arXiv:2205.11916

    2023 arXiv

  7. [15]

    Szepesvári C Lattimore T. 2020. Bandit Algorithms, page 80. Cambridge University Press

    2020

  8. [16]

    Guohao Li, Hasan Abed Al Kader Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. 2023. https://openreview.net/forum?id=3IyL2XWDkG CAMEL : Communicative agents for ''mind'' exploration of large language model society . In Thirty-seventh Conference on Neural Informati...

    2023

  9. [17]

    Jianghao Lin, Rong Shan, Chenxu Zhu, Kounianhua Du, Bo Chen, Shigang Quan, Ruiming Tang, Yong Yu, and Weinan Zhang. 2024. https://openreview.net/forum?id=Vc39mtFSUW Re LL a: Retrieval-enhanced large language models for lifelong sequential behavior comprehension in recommendati...

    2024

  10. [18]

    Xuefei Ning, Zinan Lin, Zixuan Zhou, Zifu Wang, Huazhong Yang, and Yu Wang. 2024. https://arxiv.org/abs/2307.15337 Skeleton-of-thought: Prompting llms for efficient parallel generation . Preprint, arXiv:2307.15337

    2024 arXiv

  11. [19]

    Shravan Pargaonkar. 2023. https://doi.org/10.29322/IJSRP.13.08.2023.p14015 A comprehensive research analysis of software development life cycle (sdlc) agile & waterfall model advantages, disadvantages, and application suitability in software quality engineering . International...

    2023 doi

  12. [20]

    Chen Qian, Xin Cong, Wei Liu, Cheng Yang, Weize Chen, Yusheng Su, Yufan Dang, Jiahao Li, Juyuan Xu, Dahai Li, Zhiyuan Liu, and Maosong Sun. 2023. https://arxiv.org/abs/2307.07924 Communicative agents for software development . Preprint, arXiv:2307.07924

    2023 arXiv

  13. [21]

    Linlu Qiu, Liwei Jiang, Ximing Lu, Melanie Sclar, Valentina Pyatkin, Chandra Bhagavatula, Bailin Wang, Yoon Kim, Yejin Choi, Nouha Dziri, and Xiang Ren. 2024. https://openreview.net/forum?id=bNt7oajl2a Phenomenal yet puzzling: Testing inductive reasoning capabilities of langua...

    2024

  14. [22]

    Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik R Narasimhan, and Shunyu Yao. 2023. https://openreview.net/forum?id=vAElhFcKW6 Reflexion: language agents with verbal reinforcement learning . In Thirty-seventh Conference on Neural Information Processing Systems

    2023

  15. [23]

    David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap, Madele...

    2016 doi

  16. [24]

    Mirac Suzgun and Adam Tauman Kalai. 2024. https://arxiv.org/abs/2401.12954 Meta-prompting: Enhancing language models with task-agnostic scaffolding . Preprint, arXiv:2401.12954

    2024 arXiv

  17. [25]

    Karthik Valmeekam, Matthew Marquez, Alberto Olmo, Sarath Sreedharan, and Subbarao Kambhampati. 2023. https://openreview.net/forum?id=YXogl4uQUO Planbench: An extensible benchmark for evaluating large language models on planning and reasoning about change . In Thirty-seventh Co...

    2023

  18. [26]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf Attention is all you need . In Advances in Ne...

    2017

  19. [27]

    Arun Verma and Manjesh Kumar Hanawal. 2021. https://openreview.net/forum?id=jHAAHg8T7Nx Stochastic multi-armed bandits with control variates . In Advances in Neural Information Processing Systems

    2021

  20. [28]

    Xinyuan Wang, Chenxi Li, Zhen Wang, Fan Bai, Haotian Luo, Jiayou Zhang, Nebojsa Jojic, Eric Xing, and Zhiting Hu. 2024. https://openreview.net/forum?id=22pyNMuIoa Promptagent: Strategic planning with language models enables expert-level prompt optimization . In The Twelfth Int...

    2024

  21. [29]

    Chi, Quoc V Le, and Denny Zhou

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, brian ichter, Fei Xia, Ed H. Chi, Quoc V Le, and Denny Zhou. 2022. https://openreview.net/forum?id=_VjQlMeSB_J Chain of thought prompting elicits reasoning in large language models . In Advances in Neural Information Proc...

    2022

  22. [30]

    Jimmy Wei, Kurt Shuster, Arthur Szlam, Jason Weston, Jack Urbanek, and Mojtaba Komeili. 2023. https://arxiv.org/abs/2304.13835 Multi-party chat: Conversational agents in group settings with humans and models . Preprint, arXiv:2304.13835

    2023 arXiv

  23. [31]

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W White, Doug Burger, and Chi Wang. 2023. https://arxiv.org/abs/2308.08155 Autogen: Enabling next-gen llm applications via mu...

    2023 arXiv

  24. [32]

    Yue Wu, Yewen Fan, So Yeon Min, Shrimai Prabhumoye, Stephen McAleer, Yonatan Bisk, Ruslan Salakhutdinov, Yuanzhi Li, and Tom Mitchell. 2024. https://arxiv.org/abs/2404.11483 Agentkit: Flow engineering with graphs, not coding . Preprint, arXiv:2404.11483

    2024 arXiv

  25. [33]

    Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, Rui Zheng, Xiaoran Fan, Xiao Wang, Limao Xiong, Yuhao Zhou, Weiran Wang, Changhao Jiang, Yicheng Zou, Xiangyang Liu, Zhangyue Yin, Shihan Dou, Rongxiang Weng, W...

    2023 arXiv

  26. [34]

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. 2018. https://doi.org/10.18653/v1/D18-1259 H otpot QA : A dataset for diverse, explainable multi-hop question answering . In Proceedings of the 2018 Conference...

    2018 doi

  27. [35]

    Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. 2022. https://proceedings.neurips.cc/paper_files/paper/2022/file/82ad13ec01f9fe44c01cb91814fd7b8c-Paper-Conference.pdf Webshop: Towards scalable real-world web interaction with grounded language agents . In Advances i...

    2022

  28. [36]

    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

  29. [37]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2023 b . https://openreview.net/forum?id=WE_vluYUL-X React: Synergizing reasoning and acting in language models . In The Eleventh International Conference on Learning Representations

    2023

  30. [38]

    Bin Zhang, Hangyu Mao, Jingqing Ruan, Ying Wen, Yang Li, Shao Zhang, Zhiwei Xu, Dapeng Li, Ziyue Li, Rui Zhao, Lijuan Li, and Guoliang Fan. 2024 a . https://arxiv.org/abs/2311.13884 Controlling large language model-based agents for large-scale decision-making: An actor-critic ...

    2024 arXiv

  31. [39]

    Jiahao Zhang, Haiyang Zhang, Dongmei Zhang, Yong Liu, and Shen Huang. 2024 b . https://arxiv.org/abs/2308.08973 End-to-end beam retrieval for multi-hop question answering . Preprint, arXiv:2308.08973

    2024 arXiv

  32. [40]

    Yifan Zhang, Yang Yuan, and Andrew Chi-Chih Yao. 2024 c . https://arxiv.org/abs/2311.11482 Meta prompting for ai systems . Preprint, arXiv:2311.11482

    2024 arXiv

  33. [41]

    Andy Zhou, Kai Yan, Michal Shlapentokh-Rothman, Haohan Wang, and Yu-Xiong Wang. 2024. https://openreview.net/forum?id=6LNTSrJjBe Language agent tree search unifies reasoning acting and planning in language models

    2024

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.