Pith. sign in

REVIEW 3 major objections 6 minor 26 references

AgentGroupChat-V2: Divide-and-Conquer Is What LLM-Based Multi-Agent System Need

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read AgentGroupChat-V2 claims that dividing queries into parallel task forests with role-specialized agents makes LLM multi-agent systems broadly outperform single-agent and prior multi-agent baselines.

desk verdict A role-specialization study with a useful ablation, wrapped in an overclaimed 'universal superiority' narrative that its own tables contradict. read the letter →

arxiv 2506.15451 v1 pith:MITP7FDE submitted 2025-06-18 cs.CL

Zhouhong Gu , Xiaoxuan Zhu , Yin Cai , Hao Shen , Xingzhou Chen , Qingyi Wang , Jialin Li , Xiaoran Shi
show 7 more authors
Haoran Guo Wenxuan Huang Hongwei Feng Yanghua Xiao Zheyu Ye Yao Hu Shaosheng Cao
This is my paper
classification cs.CL
keywords AgentGroupChat-V2multi-agentLLMsystemsdivide-and-conquertaskdecompositionforestparallelarchitecturemathematicalreasoningcodegeneration
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 show that divide-and-conquer is the organizing principle LLM multi-agent systems need: decompose a user query into a hierarchical task forest, run independent subtasks in parallel with specialized agents, then merge results. On that design, it claims a general-purpose framework that outperforms single-agent methods (naive prompting, chain-of-thought, ReAct) and prior multi-agent frameworks (AutoGen, Multi-Agent Debate) across math, code, structured text, commonsense, and domain knowledge. The headline numbers are 91.50% on GSM8K, 30.4% on AIME, and 79.20% pass@1 on HumanEval, with the largest margins on the hardest MATH problems. A sympathetic reading is that the framework's advantage grows with task complexity because decomposition and role specialization convert a hard monolithic problem into parallel, independently checkable pieces.

What carries the argument

The central object is the hierarchical task forest: a user query is split into a tree of subtasks with explicit parent-child dependencies, where subtasks not connected by dependencies execute in parallel and child results become context for parent tasks. Around it, the paper builds three manager modules (Query, Task, and Group) and a group-chat loop in which agents perceive the environment, decide an action, communicate by broadcast or directed dialogue, and update state; the Group Manager assigns each agent an LLM, a role, a scratch space, and an object. The task forest carries the divide-and-conquer claim by making both decomposition and dependency-aware parallel scheduling structural, rather than prompt-level.

What would settle it

Run AgentGroupChat-V2 on MATH-100 or GSM8K with the Query Manager's task-forest decomposition replaced by a single flat task (or by a fixed, non-adaptive decomposition) while keeping the same agents, roles, dialogue rounds, and underlying LLM; if accuracy does not fall below the reported levels, the divide-and-conquer claim is not supported. A second check is to disable parallel scheduling and execute the same task forest sequentially: if wall-clock cost rises but accuracy stays, parallel execution, not decomposition, may be doing the work.

Watch

Extended reading notes

Core claim

AgentGroupChat-V2 claims that a three-module architecture, consisting of a Query Manager, a Task Manager, and a Group Manager, can serve as a general-purpose substrate for LLM multi-agent collaboration. The Query Manager decomposes each user query into a task forest; the Task Manager tracks parent-child dependencies and schedules independent subtasks in parallel; the Group Manager assembles role-specialized agents to solve each subtask through bounded multi-turn dialogue. The paper's central claim is that this division of labor at both task level and execution level is what multi-agent systems need, and that it delivers accuracy gains that grow with task difficulty, including a 30.4% solve rate on AIME that roughly doubles the best baseline and Level 5 MATH improvements of more than 11 percentage points over the best baselines.

Load-bearing premise

The load-bearing assumption is that the measured gains come from divide-and-conquer task decomposition, but the ablation varies only agent count and dialogue rounds and never removes or alters the Query Manager's task-forest construction, leaving role specialization, parallel scheduling, and prompt design as confounded causes.

Editorial extensions

If this is right

  • If the central claim holds, a single general-purpose multi-agent framework can replace domain-specific systems while improving accuracy on math and code benchmarks.
  • Accuracy gains should concentrate on high-complexity tasks; on easy commonsense questions the framework can lose to a single model, so a practical deployment could route easy queries away from multi-agent collaboration.
  • Scaling the number of specialized agents should keep helping on hard reasoning tasks, whereas scaling homogeneous agents or dialogue rounds beyond a moderate depth should saturate or hurt.
  • The framework's parallel scheduling makes multi-agent collaboration cheaper in wall-clock time than sequential conversation, so its per-LLM-call efficiency is part of the claimed advantage.

Reading between the lines

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

  • The paper never isolates the Query Manager's task-forest decomposition from role specialization, parallel scheduling, and prompt design; an obvious test is to swap in a fixed decomposition or none while holding agents and rounds constant, and see how much of the GSM8K and AIME gains survive.
  • If decomposition is the active ingredient, the framework's math and code gains should transfer to any planner that produces a dependency graph, independent of the specific manager prompts; if the gains vanish without specialized roles, the title claim would need to be narrowed.
  • Because the framework can underperform single models on commonsense tasks, a production system could use a difficulty router that sends only sufficiently hard queries through the multi-agent pipeline, lowering cost while keeping benchmark gains.
  • The reported AIME advantage deserves re-testing on newer contest problems to check whether the architectural gain persists as the underlying LLMs improve.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper introduces AgentGroupChat-V2, a multi-agent framework built on a Query Manager, Task Manager, and Group Manager, with a claimed divide-and-conquer task-forest decomposition, fully parallel group-chat execution, and adaptive collaboration. It evaluates the framework on mathematical reasoning (GSM8K, MATH, AIME), code generation (MBPP, HumanEval), commonsense reasoning (HellaSwag, WinoGrande), structured text understanding (StrucText-Eval), and domain-specific QA (JEC-QA, FinQual, MedmcQA) using Qwen2.5-72B-Instruct and Llama-3.1-70B-Instruct-Turbo, against Naive, Naive-CoT, ReAct, AutoGen, and Multi-Agent Debate. The headline claim is consistent superiority across all benchmark categories, with the title and abstract attributing the gains to divide-and-conquer. The paper also reports ablations on the number of agents, dialogue rounds, and role configuration on a MATH-100 subset, plus two case studies.

Significance. If the central claims held, the framework would be a practically useful general-purpose multi-agent system with publicly released code, and the reported gains on AIME, MATH Level 5, and HumanEval would be notable. The evaluation uses standard external benchmarks with no fitted parameters in the central comparison, which is a strength. However, the significance is substantially reduced because the manuscript's own tables contradict the claimed consistent superiority, and the attribution of the observed advantages to divide-and-conquer is not tested by the reported ablations. As written, the paper does not establish either the universal performance claim or the causal role of its core mechanism.

major comments (3)
  1. [Abstract; Figure 1; Tables I, III, IV, V] The Abstract and Figure 1 caption claim that AgentGroupChat-V2 'consistently outperforms existing multi-agent approaches and baseline methods across both models and all benchmark categories.' The paper's own tables contradict this. In Table III, HellaSwag for Qwen2.5-72B is 70.3 for AgentGroupChat-V2 versus 73.7 for Naive, and WinoGrande for Qwen2.5-72B is 82.7 versus 85.5 for Naive-CoT. In Table IV, at Width=1/Depth=1, Qwen2.5-72B scores 83.5 versus Naive-CoT's 86.7. In Table V, MedmcQA for Qwen2.5-72B is 79.00 versus Naive's 89.30, FinQual for Qwen2.5-72B is 77.11 versus Multi-Agent Debate's 80.20, and JEC-QA for Qwen2.5-72B is 41.20 versus Naive's 42.56. Table I also shows losses on MATH Level 3 and Level 4 for Qwen2.5-72B (84.07 vs 88.05 and 71.79 vs 88.05 in the relevant rows) and on Level 3 and Level 4 for Llama-3.1-70B. The text itself concedes these cases in Finding 3 and Finding 5. Consequently, the universal claim in the Abstract and Figure 1 is not merely overstated; it is falsified by the reported numbers. The authors should either restrict the claim to the settings where the advantage actually holds or substantially revise the framing.
  2. [Section VI-B; title and Abstract] The title and Abstract assert that divide-and-conquer is what LLM-based multi-agent systems need, but the ablation study never isolates the divide-and-conquer task decomposition. Section VI-B varies only the number of agents (2-5), dialogue rounds (2-5), and role configuration (general vs. specialized). It does not ablate the Query Manager's task-forest construction or the Task Manager's dependency management; every run of AgentGroupChat-V2 uses the full decomposition pipeline. The comparison against AutoGen and Multi-Agent Debate is also confounded: the frameworks differ in role specialization, parallel scheduling, prompt design, and orchestration, so any performance difference cannot be attributed specifically to divide-and-conquer. An ablation that removes or perturbs the decomposition module—or applies decomposition to a baseline framework—is needed to support the title-level claim.
  3. [Section V; Tables I-V; Abstract] The experimental comparison reports single accuracy numbers without variance, number of independent runs, or statistical significance tests. This is particularly problematic for small benchmarks such as AIME, which has only 30 problems: the difference between 30.4% and 20.0% for Llama-3.1-70B is about three problems, and the difference between 21.4% and 16.7% for Qwen2.5-72B is about 1.4 problems. The Abstract's claim that AIME results 'nearly double other methods' is not supported by Table I: 30.4 vs. 20.0 is a 1.52x ratio, and 21.4 vs. 16.7 is 1.28x. The authors should report confidence intervals or significance tests and should temper claims that depend on small absolute differences.
minor comments (6)
  1. [Section IV-A] In the task structure description, the text says the parent task of a root task is marked as 'Nan'; this should be 'NaN' or 'none'.
  2. [Section II] The Related Work section has numerous formatting issues, including citations concatenated with the following sentence (e.g., 'behavioral economics principles Bates et al. (1994).Xie et al.') and missing spaces after periods; these should be cleaned up.
  3. [Section I, Abstract] There are grammatical errors in the Abstract and Introduction, such as 'We introduces AgentGroupChat-V2' and the Figure 1 caption containing 'demonstrating the effectiveness, robustness and promising of multi-agent approach'; these need copyediting.
  4. [Table I] Some table cells appear to have merged or missing separators, such as '71.7954.86' and '21.4 71.79 59.10 21.4'; please ensure all numerical entries are cleanly separated and correctly aligned in the final version.
  5. [Section V-B] The baseline descriptions do not specify key hyperparameters such as temperature, max tokens, number of debate rounds for Multi-Agent Debate, or the number of ReAct iterations; these details are needed for reproducibility.
  6. [Section VII] The case-study tables and text sometimes refer to 'AgentGroupChat' instead of 'AgentGroupChat-V2' (e.g., Table VII caption); the naming should be consistent throughout.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the framework's central evaluation is against external standard benchmarks, and the self-citation to AgentGroupChat is background, not load-bearing.

full rationale

The paper's central claims are empirical comparisons on external benchmarks (GSM8K, MATH, AIME, MBPP, HumanEval, HellaSwag, WinoGrande, StrucText-Eval, FinQual, JEC-QA, MedmcQA) against standard baselines (Naive, Naive-CoT, ReAct, AutoGen, Multi-Agent Debate). No parameter is fitted to the evaluation data and then renamed as a prediction; the reported numbers are direct accuracy or pass@k results on held-out test sets. The only self-citation, 'Building upon the AgentGroupChat framework Gu et al. (2024)' (Section III), is a statement of lineage and does not carry the argument: the claimed advantages rest on the tables and heat maps, not on an imported theorem or on the cited prior work. The absence of an ablation isolating the Query Manager's task-forest decomposition is a methodological weakness about attribution, but it is not circularity, because the framework is not defined in terms of its performance scores and the central comparisons do not reduce to fitted inputs. Likewise, the internal contradiction between the Abstract/Figure 1 claim of 'consistently outperforms ... across both models and all benchmark categories' and the paper's own Tables III-V showing losses on HellaSwag, FinQual, MedmcQA, JEC-QA, and low-complexity StrucText-Eval is a correctness and over-claiming issue, not a self-referential derivation. The honest finding is therefore no significant circularity.

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

The empirical claims rest on benchmark evaluations and hand-designed prompts. The main free parameters (agent count, dialogue rounds) are tuned on a 100-problem subset, and the task decomposition module is never isolated, so its contribution is assumed.

free parameters (3)
  • number_of_agents = not reported for main results; ablation uses 2-5
    A core hyperparameter of the group chat; the paper tunes it on MATH-100 (Section VI-B) but does not state the values used in the main experiments.
  • dialogue_rounds = not reported; ablation uses 2-5 with best at 3
    Controls interaction depth; chosen based on MATH-100 ablation, not held-out validation.
  • specialized_role_prompts = hand-designed (Table VI)
    The five math expert role descriptions are manually authored; no evidence they are optimal or transfer to other domains.
assumptions (3)
  • domain assumption The Query Manager's LLM-based task decomposition produces correct and useful task trees that do not lose information.
    The entire framework depends on this (Section III-A); no ablation tests decomposition accuracy.
  • standard math Benchmark accuracy is computed with standard evaluation protocols (exact match for math, unit tests for code, EM for QA).
    Standard practice; the pass@k formula is used (Eq. 1) though without citing its origin.
  • domain assumption The 30 problems in AIME 2024 are a sufficient sample to compare methods.
    The paper reports AIME(2024) with 30.4% vs 20.0%, but with 30 items the difference is ~3 correct answers; no confidence intervals.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AgentGroupChat-V2: Divide-and-Conquer Is What LLM-Based Multi-Agent System Need." pith.science (2026). https://pith.science/paper/MITP7FDE

@misc{pith2026250615451,
  author       = {Pith},
  title        = {Pith review of: AgentGroupChat-V2: Divide-and-Conquer Is What LLM-Based Multi-Agent System Need},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MITP7FDE}},
  note         = {Machine review of arXiv:2506.15451}
}
read the original abstract

Large language model based multi-agent systems have demonstrated significant potential in social simulation and complex task resolution domains. However, current frameworks face critical challenges in system architecture design, cross-domain generalizability, and performance guarantees, particularly as task complexity and number of agents increases. We introduces AgentGroupChat-V2, a novel framework addressing these challenges through three core innovations: (1) a divide-and-conquer fully parallel architecture that decomposes user queries into hierarchical task forest structures enabling dependency management and distributed concurrent processing. (2) an adaptive collaboration engine that dynamically selects heterogeneous LLM combinations and interaction modes based on task characteristics. (3) agent organization optimization strategies combining divide-and-conquer approaches for efficient problem decomposition. Extensive experiments demonstrate AgentGroupChat-V2's superior performance across diverse domains, achieving 91.50% accuracy on GSM8K (exceeding the best baseline by 5.6 percentage points), 30.4% accuracy on competition-level AIME (nearly doubling other methods), and 79.20% pass@1 on HumanEval. Performance advantages become increasingly pronounced with higher task difficulty, particularly on Level 5 MATH problems where improvements exceed 11 percentage points compared to state-of-the-art baselines. These results confirm that AgentGroupChat-V2 provides a comprehensive solution for building efficient, general-purpose LLM multi-agent systems with significant advantages in complex reasoning scenarios. Code is available at https://github.com/MikeGu721/AgentGroupChat-V2.

Figures

Figures reproduced from arXiv: 2506.15451 by the authors.

Figure 1
Figure 1. Upper: The performance of Qwen2.5-72B. Downer: The performance of Llama3.1-70B. Both models are evaluated across five diverse reasoning domains: commonsense reasoning, domain-specific knowledge, structural text understanding, mathematics, and code generation. AGENTGROUPCHAT-V2 consistently outperforms existing multi-agent approaches and baseline methods across both models and all benchmark categories, demonstrating … view at source ↗
Figure 2
Figure 2. Illustration of AgentGroupChat-V2 framework, which composes of three main components: Query Manager, Task Manager and Group Manager. The [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Detailed implementation of Task Manager and Group Manager modules [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: AGENTGROUPCHAT-V2 w/ Specified Role performance on MATH-100 2 3 4 5 2 3 4 5 42 40 39 37 50 44 43 41 45 42 40 38 43 40 38 36 Dialogue Rounds Number of Agents 30 40 50 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: AutoGen performance on MATH-100 2 3 4 5 2 3 4 5 44 49 46 43 42 45 44 41 40 43 42 39 38 41 40 37 Dialogue Rounds Number of Agents 30 40 50 [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 8
Figure 8. Figure 8: A schematic illustration of task decomposition for the “Interactive [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Another illustration of task decomposition for “Writing Blockchain [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

26 extracted references · 3 canonical work pages

  1. [5]

    AgentGroupChat: An Interactive Group Chat Simulacra For Better Eliciting Emergent Behavior

    Zhouhong Gu, Xiaoxuan Zhu, Haoran Guo, Lin Zhang, Yin Cai, Hao Shen, Jiangjie Chen, Zheyu Ye, Yifei Dai, Yan Gao, et al. Agentgroupchat: An interactive group chat simulacra for better eliciting emergent behavior.arXiv preprint arXiv:2403.13433,

    work page Pith review arXiv

  2. [6]

    Large language model based multi-agents: A survey of progress and challenges.arXiv preprint arXiv:2402.01680,

    Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V Chawla, Olaf Wiest, and Xiangliang Zhang. Large language model based multi-agents: A survey of progress and challenges.arXiv preprint arXiv:2402.01680,

    arXiv

  3. [7]

    "Guinea Pig Trials" Utilizing GPT: A Novel Smart Agent-Based Modeling Approach for Studying Firm Competition and Collusion

    Xu Han, Zengqing Wu, and Chuan Xiao. "guinea pig trials" utilizing gpt: A novel smart agent-based modeling approach for studying firm competition and collusion.arXiv preprint arXiv:2308.10974,

    work page Pith review arXiv

  4. [9]

    Abhishek Kumar

    URL https://arxiv.org/abs/2402.02716. Abhishek Kumar. Large language model based multi-agent sys- tem augmented complex event processing pipeline for inter- net of multimedia things.arXiv preprint arXiv:2501.00906,

    arXiv

  5. [11]

    Encouraging divergent thinking in large language models through multi-agent debate.arXiv preprint arXiv:2305.19118,

    Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi, and Zhaopeng Tu. Encouraging divergent thinking in large language models through multi-agent debate.arXiv preprint arXiv:2305.19118,

    arXiv

  6. [12]

    From skepticism to acceptance: Simulating the attitude dynamics toward fake news.arXiv preprint arXiv:2403.09498,

    Yuhan Liu, Xiuying Chen, Xiaoqing Zhang, Xing Gao, Ji Zhang, and Rui Yan. From skepticism to acceptance: Simulating the attitude dynamics toward fake news.arXiv preprint arXiv:2403.09498,

    arXiv

  7. [13]

    Large language model agent: A survey on methodology, applications and challenges.arXiv preprint arXiv:2503.21460,

    Junyu Luo et al. Large language model agent: A survey on methodology, applications and challenges.arXiv preprint arXiv:2503.21460,

    arXiv

  8. [14]

    Yoav Shoham and Kevin Leyton-Brown.Multiagent systems: Algorithmic, game-theoretic, and logical foundations

    URL https://arxiv.org/abs/2307.16789. Yoav Shoham and Kevin Leyton-Brown.Multiagent systems: Algorithmic, game-theoretic, and logical foundations. Cam- bridge University Press,

    arXiv

Show all 26 references
  1. [16]

    V oyager: An open-ended embodied agent with large language models.arXiv preprint arXiv:2305.16291,

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandku- mar. V oyager: An open-ended embodied agent with large language models.arXiv preprint arXiv:2305.16291,

    arXiv

  2. [18]

    The rise and potential of large language model based agents: A survey.arXiv preprint arXiv:2309.07864,

    Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, et al. The rise and potential of large language model based agents: A survey.arXiv preprint arXiv:2309.07864,

    arXiv

  3. [19]

    Can large language model agents simulate human trust behaviors?arXiv preprint arXiv:2402.04559,

    Chengxing Xie, Canyu Chen, Feiran Jia, Ziyu Ye, Kai Shu, Adel Bibi, Ziniu Hu, Philip Torr, Bernard Ghanem, and Guohao Li. Can large language model agents simulate human trust behaviors?arXiv preprint arXiv:2402.04559,

    arXiv

  4. [20]

    Examining inter-consistency of large language models collaboration: An in-depth analysis via debate.arXiv preprint arXiv:2305.11595,

    Kai Xiong, Xiao Ding, Yixin Cao, Ting Liu, and Bing Qin. Examining inter-consistency of large language models collaboration: An in-depth analysis via debate.arXiv preprint arXiv:2305.11595,

    arXiv

  5. [21]

    React: Synergizing reasoning and acting in language models.arXiv preprint arXiv:2210.03629,

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models.arXiv preprint arXiv:2210.03629,

    arXiv

  6. [23]

    Electionsim: Massive population election simulation powered by large language model driven agents

    Xinnong Zhang, Jiayu Lin, Libo Sun, Weihong Qi, Yihang Yang, Yue Chen, Hanjia Lyu, Xinyi Mou, Siming Chen, Jiebo Luo, et al. Electionsim: Massive population election simulation powered by large language model driven agents. arXiv preprint arXiv:2410.20746,

    arXiv

  7. [24]

    Multi-agent reinforcement learning: A comprehensive survey

    Yifan Zhang, Jingqin Wang, Jianye Yu, and Joey Tianyi Wen. Multi-agent reinforcement learning: A comprehensive survey. arXiv preprint arXiv:2312.10256,

    arXiv

  8. [25]

    Sotopia: Interactive evaluation for social intelligence in language agents.arXiv preprint arXiv:2310.11667,

    Xuhui Zhou, Hao Zhu, Leena Mathur, Ruohong Zhang, Haofei Yu, Zhengyang Qi, Louis-Philippe Morency, Yonatan Bisk, Daniel Fried, Graham Neubig, et al. Sotopia: Interactive evaluation for social intelligence in language agents.arXiv preprint arXiv:2310.11667,

    arXiv

  9. [26]

    Multi-agent reinforcement learning in wireless distributed networks for 6g.arXiv preprint arXiv:2502.05812, 2025

    Yiyang Zhu. Multi-agent reinforcement learning in wireless distributed networks for 6g.arXiv preprint arXiv:2502.05812, 2025

    2025 arXiv

  10. [1994]

    Why do multi-agent llm systems fail?arXiv preprint arXiv:2503.13657,

    Mert Cemri, Melissa Z Pan, Shuyi Yang, Lakshya A Agrawal, Bhavya Chopra, Rishabh Tiwari, Kurt Keutzer, Aditya Parameswaran, Dan Klein, Kannan Ramchandran, et al. Why do multi-agent llm systems fail?arXiv preprint arXiv:2503.13657,

    arXiv

  11. [1999]

    S3: Social-network simulation system with large language model- empowered agents.arXiv preprint arXiv:2307.14984,

    Chen Gao, Xiaochong Lan, Zhihong Lu, Jinzhu Mao, Jinghua Piao, Huandong Wang, Depeng Jin, and Yong Li. S3: Social-network simulation system with large language model- empowered agents.arXiv preprint arXiv:2307.14984,

    arXiv

  12. [2008]

    Lawluo: A chinese law firm co-run by llm agents.arXiv preprint arXiv:2407.16252,

    Jingyun Sun, Chengxiao Dai, Zhongze Luo, Yangbo Chang, and Yang Li. Lawluo: A chinese law firm co-run by llm agents.arXiv preprint arXiv:2407.16252,

    arXiv

  13. [2009]

    Autogen: Enabling next-gen llm applications via multi-agent conversation framework.arXiv preprint arXiv:2308.08155,

    Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Shaokun Zhang, Erkang Zhu, Beibin Li, Li Jiang, Xiaoyun Zhang, and Chi Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework.arXiv preprint arXiv:2308.08155,

    arXiv

  14. [2021]

    Improving factuality and reasoning in language models through multiagent debate.arXiv preprint arXiv:2305.14325,

    Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. Improving factuality and reasoning in language models through multiagent debate.arXiv preprint arXiv:2305.14325,

    arXiv

  15. [2022]

    From mooc to maic: Reshaping online teaching and learning through llm-driven agents.arXiv preprint arXiv:2409.03512,

    Jifan Yu, Zheyuan Zhang, Daniel Zhang-li, Shangqing Tu, Zhanxin Hao, Rui Miao Li, Haoxuan Li, Yuanchun Wang, Hanming Li, Linlu Gong, et al. From mooc to maic: Reshaping online teaching and learning through llm-driven agents.arXiv preprint arXiv:2409.03512,

    arXiv

  16. [2023]

    Metagpt: Meta programming for multi-agent collaborative framework.arXiv preprint arXiv:2308.00352,

    Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, et al. Metagpt: Meta programming for multi-agent collaborative framework.arXiv preprint arXiv:2308.00352,

    arXiv

  17. [2024]

    Allan Dafoe, Yoram Bachrach, Gillian Hadfield, Eric Horvitz, Kate Larson, and Thore Graepel

    URL https://arxiv.org/abs/2409.08069. Allan Dafoe, Yoram Bachrach, Gillian Hadfield, Eric Horvitz, Kate Larson, and Thore Graepel. Cooperative ai: machines must learn to find common ground,

    arXiv

  18. [2025]

    Can large language models cap- ture public opinion about global warming? an empirical assessment of algorithmic fidelity and bias.arXiv preprint arXiv:2311.00217,

    Sanguk Lee, Tai-Quan Peng, Matthew H Goldberg, Seth A Rosenthal, John E Kotcher, Edward W Maibach, and Anthony Leiserowitz. Can large language models cap- ture public opinion about global warming? an empirical assessment of algorithmic fidelity and bias.arXiv preprint arXiv:23...

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.