Pith. sign in

REVIEW 4 major objections 7 minor 3 cited by

Understanding the Information Propagation Effects of Communication Topologies in LLM-based Multi-Agent Systems

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read By measuring how single-agent errors and insights flip a multi-agent system's final answer, this paper argues that optimal communication topologies sit at moderate sparsity and presents a learner, EIB-Learner, that builds such topologies.

desk verdict A plausible design principle for LLM-MAS topologies, backed by a useful new metric and a reasonable learner, but the headline numbers rest on single unseeded runs. read the letter →

arxiv 2505.23352 v1 pith:WGI4M3KY submitted 2025-05-29 cs.MA cs.AI

Xu Shen , Yixin Liu , Yiwei Dai , Yili Wang , Rui Miao , Yue Tan , Shirui Pan , Xin Wang This is my paper
classification cs.MAcs.AI
keywords LLMmulti-agentsystemscommunicationtopologygraphneuralnetworkscounterfactualanalysiserrorpropagationinsightoptimizationdebate
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 asks why sparse communication graphs help or hurt LLM multi-agent teams. Using counterfactual experiments, it claims that every topology is caught between two effects: dense graphs spread useful insights but also amplify an individual agent's mistakes, while sparse graphs block both. The balance point, it concludes, is moderate sparsity, where a topology suppresses errors yet still lets correct information through, and accuracy peaks there. The paper then introduces EIB-Learner, a graph-neural-network method that learns a query-specific topology by fusing a sparse-view and a dense-view estimate of connectivity. Across six benchmarks in reasoning, math, and code, it reports 91.38% average accuracy, ahead of earlier automated topology designers.

What carries the argument

The load-bearing machinery is a pair of counterfactual metrics plus a dual-view graph-neural-network learner. CAPE (Counterfactual Agent Propagation Effect) measures how often changing one agent's output to an injected correct or incorrect answer flips the final system answer; TCTE (Total Counterfactual Topology Effect) averages CAPE across agents with a $1/\sqrt{d_i}$ weight that downweights high-degree agents. On the learning side, one GNN runs message passing over the sparsest chain graph to model error suppression, another over the fully connected graph to model insight diffusion, and an inner-product decoder converts each view's node embeddings into edge-coefficient matrices. A query-aware softmax gate fuses the two matrices into a single topology, trained by policy gradient against final-answer accuracy.

What would settle it

Run the flip-rate experiment with many phrasings of the injected wrong (or right) output per agent; if the flip rates vary so widely that the monotone trends in Figure 2 do not reproduce, the balancing principle lacks support. An equally decisive test is matching EIB-Learner against random graphs with the same edge count on all six benchmarks: if random pruning performs as well, the learned fusion contributes nothing beyond moderate sparsity.

Watch

Extended reading notes

Core claim

The central claim is that the value of a communication topology for an LLM-based multi-agent system consists of two opposed causal effects—error propagation and insight propagation—and that task accuracy is maximized when the two are balanced, which in practice occurs at intermediate sparsity. To measure these effects, the paper defines CAPE, which records whether injecting a deliberately wrong or deliberately correct output into one agent flips the system's final answer, and TCTE, which averages CAPE over agents with degree-based normalization. On MMLU, dense graphs show the highest error-flip rates, sparse chains the lowest insight-flip rates, and accuracy is highest between the extremes. The paper's method, EIB-Learner, operationalizes this balance: one GNN simulates communication on the sparsest chain graph, another on the fully connected graph, and a gated fusion combines the two views into the final topology. EIB-Learner is reported to outperform both fixed and learned topologies on all six benchmarks.

Load-bearing premise

The analysis assumes that injecting one hand-written correct or incorrect output into a single agent and recording whether the final answer flips faithfully captures how real errors and insights spread; because LLM outputs are stochastic, a single forced output may not represent typical agent behavior.

Editorial extensions

If this is right

  • A topology's accuracy can be predicted from its position on the error-flip versus insight-flip curve, so future designers should report both metrics instead of edge count alone.
  • Sparsity-based optimization that only removes edges will stop short of the best topologies; the target is the error-insight balance, which may keep some dense clusters around reliable agents.
  • EIB-Learner achieves its accuracy gains without extra communication, with token consumption comparable to a sparsity-optimized baseline on both MMLU and GSM8K.
  • Because the fusion weight depends on the query, the same system can route a high-confidence query through denser connections and an uncertain query through sparser ones, giving an adaptive reliability mechanism within one model.

Reading between the lines

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

  • An unstated corollary is that per-agent reliability, if it can be estimated from past answers, could replace the query-only gate: the topology would rebalance in real time as agents prove accurate or error-prone.
  • A testable extension the paper does not run is matching EIB-Learner against random graphs with the same edge count on all six benchmarks; the paper's own analysis suggests moderate-sparsity random graphs are strong, so this control would separate the learned fusion from mere sparsity.
  • The CAPE/TCTE recipe is a generic auditing tool: the same flip-rate experiment could be applied to human team structures or other distributed decision systems to find which connections are dangerous.
  • The authors note that evaluation is limited to reasoning, math, and code with fixed roles; open-domain dialogue and real-world decision-making would test whether the error-insight balance generalizes outside well-scoped tasks.
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 / 7 minor

Summary. The paper presents a causal framework, CAPE and TCTE, to analyze how agent-level correct or erroneous outputs propagate through LLM-based multi-agent communication topologies of varying sparsity. Empirical analysis on MMLU with GPT-3.5 leads to the claim that moderately sparse topologies best balance error suppression and insight propagation. Based on this insight, the authors propose EIB-Learner, a dual-view GNN-based topology learner that fuses sparse and dense graph representations, and report state-of-the-art average accuracy of 91.38% across six benchmarks versus 90.04% for the best baseline. The paper also reports token-efficiency and robustness advantages. The code is publicly released.

Significance. If the empirical claims are ultimately supported, the paper would provide a useful design principle for LLM-based multi-agent communication and a practical, query-adaptive topology learner. The release of code, the use of publicly available benchmarks, and the attempt to define quantitative propagation metrics are notable strengths. However, the paper's central evidence currently consists of single-run point estimates without uncertainty quantification, and the causal analysis is conducted with a different model than the evaluation; these issues materially weaken the significance of the reported findings until addressed.

major comments (4)
  1. [Sec. 5.2, Table 1; Appendix D, Table 4] The state-of-the-art claim rests on single unseeded point estimates without significance tests. For MMLU the evaluation set is only 153 questions (Appendix D, Table 4), so the reported 1.98-point advantage over G-designer (88.90 vs. 86.92) is smaller than the approximate 4-point standard error of the difference between two proportions; several other per-dataset gaps, such as the 0.33-point gap on MultiArith, are also within sampling noise. Because EIB-Learner is trained with a policy gradient on a non-differentiable reward using only 40–60 queries (Sec. 4.2, Eq. 11), seed-to-seed variation is a first-order concern. The authors should report multiple random seeds and temperatures with confidence intervals or significance tests, at least for the headline average and the main comparisons.
  2. [Sec. 3.2, Fig. 2] The central 'moderate sparsity' finding is presented without uncertainty quantification. Each TCTE value in Fig. 2 is computed with a single forced-injection sample per agent per question (Definition 3.1, Eq. 5), and each sparsity level corresponds to one random edge-deletion or edge-addition draw; no temperature, seed, or repeated-sampling protocol is reported. Because LLM outputs are stochastic, a single hand-written injected output may not faithfully represent typical error or insight propagation, and the monotonic trends in TCTE and accuracy could be sampling artifacts. The authors should add error bars over repeated interventions and repeated graph draws, and ideally a statistical test of the moderation claim.
  3. [Sec. 3.2 vs. Sec. 5.1] The causal analysis is performed with GPT-3.5 on a 1000-question MMLU sample, while EIB-Learner is evaluated with GPT-4o on six different task types. The paper assumes, but does not demonstrate, that the error/insight propagation trade-off and the moderate-sparsity optimum transfer across models and domains. At minimum, the authors should re-run the CAPE/TCTE analysis on GPT-4o for MMLU, or provide evidence that the findings are stable across models and task families.
  4. [Sec. 3.1, Eq. (6)] The 1/sqrt(d_i) degree weighting in TCTE is introduced without derivation or sensitivity analysis. Because degree distributions differ substantially across the compared topologies, such as Chain versus Full, this arbitrary normalization can systematically change the TCTE comparison and consequently the support for Findings 1 and 2. The authors should justify this weighting theoretically or report unweighted TCTE as a robustness check.
minor comments (7)
  1. [Abstract] The abstract contains the typo 'EIB-leanrner' instead of 'EIB-Learner'.
  2. [Sec. 2] The edge-direction convention is inconsistent: the text says e_{ij}=(v_i,v_j) means agent v_i receives from v_j, but Eq. (2) defines N(v_i) using (v_j,v_i). Please align the definitions of e_{ij}, N(v_i), and the adjacency matrix A.
  3. [Sec. 3.2.4] The sentence 'indicating that topology optimization methods that explicitly balance error and insight propagation.' is grammatically incomplete; it appears to be missing a predicate.
  4. [Fig. 2] The x-axis labels 'Full 0.2 0.4 0.6 0.8Chain' and 'Chain 0.2 0.4 0.6 0.8Full' lack separators and should clearly distinguish sparsity from density axes.
  5. [Sec. 3.2 and Appendix D, Table 4] Section 3.2 reports two 500-question sets sampled from MMLU, while Appendix D, Table 4 lists MMLU as having 153 test questions; please clarify the sampling procedure and how these sets relate to the evaluation set used in Table 1.
  6. [Appendix E, Fig. 5] The text says the HumanEval case is case A in Figure 5, but the figure labels indicate case A is GSM8K and case B is HumanEval; please correct the cross-reference.
  7. [References] Several references contain the placeholder 'and 1 others' instead of complete author lists; please fix the bibliography.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the causal metrics and EIB-Learner's evaluation are independent by construction.

full rationale

The paper's causal analysis defines CAPE (Eq. 5) and TCTE (Eq. 6) directly from LLM-based MAS runs and final-answer flips, with no dependence on EIB-Learner's parameters. The empirical findings (Finding 1, Finding 2, and the moderate-sparsity Insight) are observations from Figure 2, not outputs of the proposed method. EIB-Learner is motivated by those findings, but its dual-view GNN design (Eqs. 7-10) and policy-gradient objective (Eq. 11) are separate, and its claimed success is measured on held-out benchmarks in Table 1 against external baselines. No fitted quantity is relabeled as a prediction, and no load-bearing argument reduces to a self-citation or an imported uniqueness claim. The lack of repeated seeds and error bars is a legitimate robustness concern, but it does not make the derivation circular.

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

The central empirical insight depends on hand-chosen TCTE weighting and sparsity schedules; the method contributes standard learned parameters rather than a parameter-free law. The main axioms are the DAG communication model, binary utility, deterministic counterfactual response, and generalization from GPT-3.5/MMLU to GPT-4o/other benchmarks.

free parameters (5)
  • TCTE degree weighting exponent = 1/sqrt(d_i)
    Hand-chosen normalizer in Definition 3.2 downweights high-degree nodes; no derivation or sensitivity analysis is provided, and it directly affects the reported TCTE curves.
  • Sparsity levels for edge deletion = 0, 0.2, 0.4, 0.6, 0.8, chain
    The empirical curves in Figure 2 depend on how edges are randomly removed from the full graph; the exact deletion schedule and guarantee of reaching a chain are not specified.
  • GNN, decoder, and gating parameters = learned via policy gradient on B in {40,60} queries
    Standard learned parameters of EIB-Learner (Eqs. 8-11); they are fitted to maximize accuracy reward, not derived from first principles.
  • Number of layers and communication rounds K = 3
    Set equal to communication rounds; no sensitivity analysis is reported.
  • MMLU evaluation subset = 153 questions (Table 4)
    The full MMLU test set has 14,042 questions; the paper does not explain the subset selection, and the analysis in Section 3 uses 500 questions, so the final accuracy may not be directly comparable across setups.
assumptions (6)
  • domain assumption Agents are functions of their system prompt and the outputs of their in-neighbors only (Eq. 2).
    The formal model ignores memory, side channels, and bidirectional deliberation beyond the specified DAG.
  • domain assumption The communication graph is a DAG and execution follows a topological order.
    Section 2; this excludes cyclic interaction patterns that occur in some multi-agent debate systems.
  • domain assumption Binary correctness is an adequate utility function for comparing topologies.
    phi(G(Q)) in Eq. 4 treats only whether the final answer is correct; partial credit and answer quality are ignored.
  • ad hoc to paper Counterfactual intervention do(O_i := O_i_hat) is well-defined and the LLM system response is effectively deterministic for measuring flips.
    Definition 3.1 defines CAPE with a single run; no repeated sampling is used despite LLM stochasticity.
  • domain assumption The empirical findings from MMLU with GPT-3.5 generalize to other benchmarks and to GPT-4o.
    The analysis in Section 3 uses GPT-3.5 and MMLU only, while the method is evaluated with GPT-4o on six benchmarks.
  • domain assumption A graph sampled from M_final is a valid DAG.
    Algorithm 1 samples a binary graph from M_final and then performs topological sorting; no cycle prevention or post-processing is described.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Understanding the Information Propagation Effects of Communication Topologies in LLM-based Multi-Agent Systems." pith.science (2026). https://pith.science/paper/WGI4M3KY

@misc{pith2026250523352,
  author       = {Pith},
  title        = {Pith review of: Understanding the Information Propagation Effects of Communication Topologies in LLM-based Multi-Agent Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WGI4M3KY}},
  note         = {Machine review of arXiv:2505.23352}
}
read the original abstract

The communication topology in large language model-based multi-agent systems fundamentally governs inter-agent collaboration patterns, critically shaping both the efficiency and effectiveness of collective decision-making. While recent studies for communication topology automated design tend to construct sparse structures for efficiency, they often overlook why and when sparse and dense topologies help or hinder collaboration. In this paper, we present a causal framework to analyze how agent outputs, whether correct or erroneous, propagate under topologies with varying sparsity. Our empirical studies reveal that moderately sparse topologies, which effectively suppress error propagation while preserving beneficial information diffusion, typically achieve optimal task performance. Guided by this insight, we propose a novel topology design approach, EIB-leanrner, that balances error suppression and beneficial information propagation by fusing connectivity patterns from both dense and sparse graphs. Extensive experiments show the superior effectiveness, communication cost, and robustness of EIB-leanrner.

Figures

Figures reproduced from arXiv: 2505.23352 by the authors.

Figure 1
Figure 1. Illustration of (a) insight suppression caused [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Empirical results of error propagation effect [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The overall framework of EIB-LEARNER. EIB-LEARNER simulates MAS communication via GNNs, generating two connectivity coefficient matrices to suppress error spreading (sparse view) and enhance insight propagation (dense view), which are then combined by a query-aware fusion module into an optimal topology. MAS as an Attributed Graph While the inter￾agent communication can be represented as a graph structure, we furthe… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Experiments on token consumption and ro [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Case study of the communication topologies designed by [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

  1. Attacking and Defending Multi-Agent Collaborative Filtering Systems Through Connectivity

    cs.IR 2026-08 conditional novelty 6.0 of 10

    In agent-based collaborative filtering, attack spread and privacy leakage grow with interaction connectivity, but the effect is asymmetric between user and item agents and differs between early and steady-state phases.

  2. The Social Cost of Intelligence: Emergence, Propagation, and Amplification of Stereotypical Bias in Multi-Agent Systems

    cs.MA 2025-10 conditional novelty 6.0 of 10

    Multi-agent LLM systems are less robust to stereotyping than single agents, and simple bias-injection attacks succeed on most systems.

  3. Secure Multi-LLM Agentic AI and Agentification for Edge General Intelligence by Zero-Trust: A Survey

    cs.NI 2025-08 conditional novelty 4.0 of 10

    A survey proposing zero-trust architecture for multi-LLM systems in edge computing, with a taxonomy of model- and system-level defenses and a conceptual framework.

Reference graph

Works this paper leans on

37 extracted references · 16 canonical work pages · cited by 3 Pith papers

  1. [1]

    Xiaohe Bo, Zeyu Zhang, Quanyu Dai, Xueyang Feng, Lei Wang, Rui Li, Xu Chen, and Ji-Rong Wen. 2024. Reflective multi-agent collaboration based on large language models. Advances in Neural Information Processing Systems, 37:138595--138631

    2024

  2. [2]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde De Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, and 1 others. 2021. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374

    arXiv 2021

  3. [3]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, and 1 others. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168

    arXiv 2021

  4. [4]

    Yilun Du, Shuang Li, Antonio Torralba, Joshua B Tenenbaum, and Igor Mordatch. 2023. Improving factuality and reasoning in language models through multiagent debate. In Forty-first International Conference on Machine Learning

    2023

  5. [5]

    Taicheng Guo, Xiuying Chen, Yaqi Wang, Ruidi Chang, Shichao Pei, Nitesh V Chawla, Olaf Wiest, and Xiangliang Zhang. 2024 a . Large language model based multi-agents: a survey of progress and challenges. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 8048--8057

    work page 2024

  6. [6]

    Xudong Guo, Kaixuan Huang, Jiale Liu, Wenhui Fan, Natalia V \'e lez, Qingyun Wu, Huazheng Wang, Thomas L Griffiths, and Mengdi Wang. 2024 b . Embodied llm agents learn to cooperate in organized teams. In Language Gamification-NeurIPS 2024 Workshop

    work page 2024

  7. [7]

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. 2021. Measuring massive multitask language understanding. In International Conference on Learning Representations

    2021

  8. [8]

    Mengkang Hu, Pu Zhao, Can Xu, Qingfeng Sun, Jianguang Lou, Qingwei Lin, Ping Luo, and Saravan Rajmohan. 2024 a . Agentgen: Enhancing planning abilities for large language model based agent via environment and task generation. arXiv preprint arXiv:2408.00764

    arXiv 2024

Show all 37 references
  1. [9]

    Shengchao Hu, Li Shen, Ya Zhang, and Dacheng Tao. 2024 b . Learning multi-agent communication from graph modeling perspective. In International Conference on Learning Representations

    2024

  2. [10]

    Yoichi Ishibashi and Yoshimasa Nishimura. 2024. Self-organized agents: A llm multi-agent framework toward ultra large-scale code generation and optimization. arXiv preprint arXiv:2404.02183

    2024 arXiv

  3. [11]

    Guohao Li, Hasan Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. 2023. Camel: Communicative agents for" mind" exploration of large language model society. Advances in Neural Information Processing Systems, 36:51991--52008

    2023

  4. [12]

    Yunxuan Li, Yibing Du, Jiageng Zhang, Le Hou, Peter Grabowski, Yeqing Li, and Eugene Ie. 2024. Improving multi-agent debate with sparse communication topology. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 7281--7294

    2024

  5. [13]

    Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi, and Zhaopeng Tu. 2024. Encouraging divergent thinking in large language models through multi-agent debate. In Proceedings of the 2024 Conference on Empirical Methods in Natural Langua...

    2024

  6. [14]

    Wang Ling, Dani Yogatama, Chris Dyer, and Phil Blunsom. 2017. Program induction by rationale generation: Learning to solve and explain algebraic word problems. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), p...

    2017

  7. [15]

    Bang Liu, Xinfeng Li, Jiayi Zhang, Jinlin Wang, Tanjin He, Sirui Hong, Hongzhang Liu, Shaokun Zhang, Kaitao Song, Kunlun Zhu, and 1 others. 2025. Advances and challenges in foundation agents: From brain-inspired intelligence to evolutionary, collaborative, and safe systems. ar...

    2025 arXiv

  8. [16]

    Arkil Patel, Satwik Bhattamishra, and Navin Goyal. 2021. Are NLP models really able to solve simple math word problems? In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2080--2094

    2021

  9. [17]

    Chen Qian, Zihao Xie, Yifei Wang, Wei Liu, Yufan Dang, Zhuoyun Du, Weize Chen, Cheng Yang, Zhiyuan Liu, and Maosong Sun. 2025. Scaling large-language-model-based multi-agent collaboration. In International Conference on Learning Representations

    2025

  10. [18]

    Subhro Roy and Dan Roth. 2015. Solving general arithmetic word problems. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1743--1752

    2015

  11. [19]

    Yashar Talebirad and Amirhossein Nadiri. 2023. Multi-agent collaboration: Harnessing the power of intelligent llm agents. arXiv preprint arXiv:2306.03314

    2023 arXiv

  12. [20]

    Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, Zhiyuan Chen, Jiakai Tang, Xu Chen, Yankai Lin, and 1 others. 2024 a . A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6):186345

    2024

  13. [21]

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2023. Self-consistency improves chain of thought reasoning in language models. In International Conference on Learning Representations

    2023

  14. [22]

    Zhenhailong Wang, Shaoguang Mao, Wenshan Wu, Tao Ge, Furu Wei, and Heng Ji. 2024 b . Unleashing the emergent cognitive synergy in large language models: A task-solving agent through multi-persona self-collaboration. In Proceedings of the 2024 Conference of the North American C...

    2024

  15. [23]

    Zhexuan Wang, Yutong Wang, Xuebo Liu, Liang Ding, Miao Zhang, Jie Liu, and Min Zhang. 2025. Agentdropout: Dynamic agent elimination for token-efficient and high-performance llm-based multi-agent collaboration. arXiv preprint arXiv:2503.18891

    2025 arXiv

  16. [24]

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

    2022

  17. [25]

    Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, Ming Zhang, Junzhe Wang, Senjie Jin, Enyu Zhou, and 1 others. 2025. The rise and potential of large language model based agents: A survey. Science China Information Sciences, 68(2):121101

    2025

  18. [26]

    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

    2023

  19. [27]

    Zihao Yi, Jiarui Ouyang, Yuwen Liu, Tianhao Liao, Zhe Xu, and Ying Shen. 2024. A survey on recent advances in llm-based multi-turn dialogue systems. arXiv preprint arXiv:2402.18013

    2024 arXiv

  20. [28]

    Guibin Zhang, Yanwei Yue, Zhixun Li, Sukwon Yun, Guancheng Wan, Kun Wang, Dawei Cheng, Jeffrey Xu Yu, and Tianlong Chen. 2025 a . Cut the crap: An economical communication pipeline for llm-based multi-agent systems. In International Conference on Learning Representations

    2025

  21. [29]

    Guibin Zhang, Yanwei Yue, Xiangguo Sun, Guancheng Wan, Miao Yu, Junfeng Fang, Kun Wang, Tianlong Chen, and Dawei Cheng. 2025 b . G-designer: Architecting multi-agent communication topologies via graph neural networks. In Forty-second International Conference on Machine Learning

    2025

  22. [30]

    Jintian Zhang, Xin Xu, Ningyu Zhang, Ruibo Liu, Bryan Hooi, and Shumin Deng. 2024 a . Exploring collaboration mechanisms for llm agents: A social psychology view. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)...

    2024

  23. [31]

    Kechi Zhang, Jia Li, Ge Li, Xianjie Shi, and Zhi Jin. 2024 b . CodeAgent : Enhancing code generation with tool-integrated agent systems for real-world repo-level coding challenges. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volu...

    2024

  24. [32]

    Yusen Zhang, Ruoxi Sun, Yanfei Chen, Tomas Pfister, Rui Zhang, and Sercan Arik. 2024 c . Chain of agents: Large language models collaborating on long-context tasks. Advances in Neural Information Processing Systems, 37:132208--132237

    2024

  25. [33]

    Andy Zhou, Kai Yan, Michal Shlapentokh-Rothman, Haohan Wang, and Yu-Xiong Wang. 2024. Language agent tree search unifies reasoning, acting, and planning in language models. In Proceedings of the 41st International Conference on Machine Learning, pages 62138--62160

    2024

  26. [34]

    Han Zhou, Xingchen Wan, Ruoxi Sun, Hamid Palangi, Shariq Iqbal, Ivan Vuli \'c , Anna Korhonen, and Sercan \"O Ar k. 2025. Multi-agent design: Optimizing agents with better prompts and topologies. arXiv preprint arXiv:2502.02533

    2025

  27. [35]

    Mingchen Zhuge, Wenyi Wang, Louis Kirsch, Francesco Faccio, Dmitrii Khizbullin, and J \"u rgen Schmidhuber. 2024. Gptswarm: Language agents as optimizable graphs. In Forty-first International Conference on Machine Learning

    2024

  28. [36]

    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...

  29. [37]

    write newline

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

Pith tools

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