Pith. sign in

REVIEW 5 major objections 4 minor 3 cited by

Adaptive Graph Pruning for Multi-Agent Communication

T0 review · 5 major / 4 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 Adaptive Graph Pruning lets an LLM agent team choose its own size and communication pattern per task, reaching 91.04% average accuracy while cutting prompt tokens by up to 90%.

desk verdict AGP is a genuinely new dual-pruning method for adaptive multi-agent communication, but the central task-adaptivity claim is undercut by an unverified Stage-I/evaluation split and the abstract overstates token savings. read the letter →

arxiv 2506.02951 v3 pith:IDIBSDMA submitted 2025-06-03 cs.CL cs.MA

Boyi Li , Zhonghan Zhao , Der-Horng Lee , Gaoang Wang This is my paper
classification cs.CLcs.MA
keywords multi-agentLLMsystemsadaptivecommunicationtopologygraphpruninghardsofttokenefficiencytaskadaptivityneuralnetwork
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

This paper argues that a multi-agent LLM system should not be stuck with a fixed roster of agents and a fixed communication layout. It proposes Adaptive Graph Pruning (AGP), which decides for each incoming query both which agents should participate (hard pruning) and how strongly they should talk to each other (soft pruning). Over six benchmarks spanning general knowledge, math reasoning, and code generation, AGP reports the highest average accuracy among twelve baselines, including topologies that are optimized once, and claims token savings of up to 90% compared with other multi-agent setups. The authors take this as evidence that task-adaptive, size-variable topologies are both feasible and worth pursuing: the best team for an easy arithmetic problem can be a two-agent chain, while harder or subtler tasks may retain an apparently irrelevant expert that nonetheless helps.

What carries the argument

The central object is the maximum complete graph $K_{N_{\max}}$ over a fixed pool of heterogeneous agents, together with two learned structures: a directed edge-weight matrix $W \in [0,1]^{N_{\max} \times N_{\max}}$ (soft-pruning) and a binary node mask $m \in \{0,1\}^{N_{\max}}$ (hard-pruning). A two-layer GCN encodes agent profiles and the task into a shared latent space; a bilinear head produces edge weights and an MLP head produces node masks, trained jointly with edge loss and node loss. The Gumbel-Sigmoid trick lets gradients flow through the discrete mask decisions. The claim is that these two levers, operating together, are what let the topology shrink and rewire per query rather than staying fixed.

What would settle it

Take one of the six benchmarks, split its queries so that the 100 to 200 optimization queries used in Stage I are disjoint from the evaluation set (or remove them entirely), and re-run the AGP pipeline. If the accuracy gain over fixed-topology baselines mostly disappears, the central claim of task-adaptive generalization is not supported; if the gain survives on truly held-out queries, the claim is strengthened.

Watch

Extended reading notes

Core claim

AGP jointly optimizes agent quantity (hard-pruning) and communication topology (soft-pruning) within a maximum complete graph of fifteen heterogeneous LLM agents. Stage I mines near-optimal subgraphs by sampling complete subgraphs, fine-tuning each on a task, and keeping the top two performers as ground-truth edge-weight matrices and node masks; Stage II trains a dual-branch GNN that shares a latent representation and outputs both a directed weighted adjacency and a binary node mask, using Gumbel-Sigmoid to bridge discrete choices. The result is a single forward pass that produces a task-specific, variable-size topology. The paper reports 91.04% average accuracy, ahead of the best static-graph competitor by +1.48 and the strongest single-agent baseline by +5.62, with up to 90% fewer prompt tokens and baseline-beating performance after about ten training steps.

Load-bearing premise

Stage I assumes that the top two graphs found by scoring a few hundred sampled subgraphs on 100 to 200 queries per benchmark are the right teaching signal for the whole benchmark and for unseen tasks, and no held-out split is reported; if those queries overlap the evaluation set, the reported gains could be benchmark fitting rather than genuine task adaptivity.

Editorial extensions

If this is right

  • A single trained AGP model produces a different team size and communication graph for each query, removing the need for per-task manual topology design at inference.
  • Prompt-token budgets drop sharply because pruned graphs keep only useful agents and attenuate noisy edges; on MMLU the paper reports about 90% fewer tokens than the GPTSwarm baseline with higher accuracy.
  • Ablations show that neither edge-weight learning alone nor node selection alone is sufficient; removing either branch lowers accuracy by roughly 2 to 5 points on MMLU, GSM8K, and HumanEval.
  • AGP surpasses the baselines after about ten training steps, indicating that the dual-pruning objective learns faster than fixed-graph or edge-only optimization.
  • Learned topologies can be counter-intuitive, sometimes retaining an apparently irrelevant agent that improves accuracy, meaning human intuition is not always the best guide.

Reading between the lines

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

  • Stage I's mining cost, which involves fine-tuning every sampled subgraph on 100 to 200 queries per benchmark, is not counted in the reported training efficiency; a fair comparison with search-based baselines would need to include it.
  • The learned adaptivity is bounded by the 460 supervision pairs and the fifteen hand-written roles; a broader role pool or a different task mix could change which topologies are discovered.
  • The 'counter-intuitive' agents may be functioning as general critics or as sources of alternative reasoning rather than as topic experts; testing which agents contribute and why would clarify the mechanism.
  • A direct generalization test would be to evaluate AGP on task families absent from the Stage I corpus, or to verify that the 100 to 200 optimization queries are disjoint from the evaluation sets.
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 / 4 minor

Summary. AGP proposes a two-stage framework that learns task-adaptive communication topologies for multi-agent LLM systems by jointly pruning nodes (hard pruning) and edges (soft pruning). In Stage I, the method samples complete subgraphs from a heterogeneous agent pool, scores each on optimization queries, and keeps the top-performing graphs per task as supervision; in Stage II, it trains a GNN with a shared latent space to output a weighted adjacency and a node mask for any new query. The paper reports state-of-the-art average accuracy of 91.04% across six benchmarks (MMLU, GSM8K, MultiArith, SVAMP, AQuA, HumanEval), claims token reductions up to 90%, and demonstrates fast convergence of the Stage II network in about ten training steps.

Significance. The conceptual contribution is timely and interesting: existing methods either fix the number of agents or optimize edge weights within a fixed pool, whereas AGP is the first in this line to treat node count and communication graph as jointly learnable per task. The two-stage supervision idea is pragmatic, and the case study of counter-intuitive agent combinations is compelling. The paper also provides a clean ablation separating soft- and hard-pruning, and the authors state that code and demos are publicly available, which is a positive reproducibility step. However, the empirical support for the headline claims is weakened by unresolved questions about whether the Stage I supervision and the evaluation sets overlap, by the absence of statistical reliability measures, and by an overstatement of the token savings. The underlying approach remains promising, but the evidence in its current form does not yet establish that the reported gains reflect task adaptivity rather than benchmark fitting.

major comments (5)
  1. [Section 3.2 / Section 4.1 / Table A3] The paper never states whether the Q∈{100,200} optimization queries used to mine Stage I supervision are disjoint from the evaluation instances in Table A4 (e.g., MMLU #Test 153, GSM8K 1,319, HumanEval 164), and Table A3 confusingly labels the training-set column "#Test". If the optimization queries overlap the test sets, the reported gains (MMLU +9.84, GSM8K +7.56) would reflect fitting to the evaluation distribution rather than task adaptivity. The authors must specify the exact origin and split of these queries, and should validate on a held-out benchmark family that was not used for supervision.
  2. [Section 3.2 / Table A3] The description "Per task, we keep the top-2 performers" appears inconsistent with the corpus statistics of 460 supervision graphs for 200+100+160 tasks. If top-2 per task were kept, the corpus should contain roughly twice as many graphs. The paper should clarify the exact counting (number of queries, number of graphs per query) and make Table A3's column headers consistent, since this directly affects the interpretation of the supervision pool.
  3. [Abstract / Section 4.2 / Figure 5] The abstract's claim of a "90%+" token decrease is contradicted by Figure 5 and the text, which report reductions of about 65% on GSM8K, 67% on SVAMP, and 22% on HumanEval relative to GPTSwarm, with 90% only on MMLU. The abstract should be revised to "up to 90%" and should report the per-benchmark reductions or clearly state the basis of the aggregate claim.
  4. [Section 4.1] The paper reports only the average of two evaluation runs and provides no standard deviations or significance tests, yet the headline claims (e.g., +1.48 over G-Designer, +5.62 over SC) are of the same magnitude as typical run-to-run variability of LLM sampling at temperature 1. The authors should provide multiple independent runs with variance and, if possible, a paired significance test to support the state-of-the-art claim.
  5. [Section 3.2 / Section 4.2] The efficiency claims ("training-efficient", "few training steps") refer only to Stage II GNN training, while Stage I samples B=2,000 graphs and scores each sampled graph by fine-tuning or running it on the task, a potentially enormous LLM-call budget that is never reported. The paper must quantify the Stage I computational and token cost (and compare it with baseline training or search costs) before claiming token economy and training efficiency.
minor comments (4)
  1. [Appendix A.4] The Gaussian fit reports "σ = −0.607"; a standard deviation cannot be negative, so this is likely a typo that should be corrected.
  2. [Section 4.2 / Abstract] There are typos such as "Similar gains aear on GSM8K" and "a increase"; the abstract and main text should be copyedited.
  3. [Section 3.3.2, Eq. (5)] The notation "λ_s⟨ˆy⟩" is used before being clearly defined; please define the operator ⟨·⟩ explicitly when it is introduced.
  4. [Table 1] Some entries are missing (e.g., MetaGPT has only one reported score), which makes the average comparison across methods unclear; please either fill in the missing values or explain why they are omitted.

Circularity Check

0 steps flagged · score 0.0 of 10

No constructional circularity; the benchmark-overlap concern is a data-leakage risk, not a derivation that reduces to its inputs.

full rationale

Stage I (Section 3.2) mines supervision graphs by scoring sampled subgraphs on Q in {100, 200} optimization queries per benchmark and keeps the top-2 performers as labels, and Stage II is trained on those labels and then evaluated on the same six benchmarks (Table A4). This would be circular only if the optimization queries coincided with the evaluation queries, making the final "prediction" a retrieval of labels computed from the same instances. The paper never states that the sets are disjoint, but it also never states that they overlap; for GSM8K (Q <= 200 vs. #Test 1,319), MultiArith (#Test 600), SVAMP (#Test 1,000), and HumanEval (#Test 164), the reported test sizes are large enough that a disjoint evaluation is feasible, and the wording "queries for optimization" versus "evaluation" suggests separate query sets. The absence of an explicit held-out split is a reproducibility and correctness concern, not a reduction by construction. No equation in the paper identifies the final reported accuracy with the Stage-I scoring function ut(G) = Acc(G; t); the learned GNN must generalize to new queries if the splits are disjoint. The only self-citation, [22] in Section 3.1, points to the paper's own appendix for notation and is not load-bearing. The definition of "fully task-adaptive" in Section 4.2 is descriptive categorization, not a derivation of the reported gains. Therefore, no circular step meets the evidentiary standard required by the analysis. The strongest attack on the paper is a potential data-hygiene problem, but that is distinct from circularity by construction.

Assumptions & free parameters 11 free parameters · 4 assumptions · 0 invented entities

The method depends on hand-set hyperparameters, a fixed 15-agent pool, and the assumption that Stage I graph mining on benchmark queries provides valid supervision. The most consequential assumption, that no distribution shift or data overlap exists between Stage I optimization queries and the final evaluation, is not verified. No new physical entities are introduced.

free parameters (11)
  • lambda_off (Eq. 4) = 0.5
    Off-diagonal edge penalty weight; chosen by hand, no sensitivity analysis.
  • lambda_s (Eq. 5) = 0.1
    Node-mask sparsity penalty; chosen by hand.
  • lambda_c (Eq. 5) = 0.05
    Coherence penalty for edges from pruned nodes; chosen by hand.
  • beta (total objective) = 1.0
    Node/edge loss balance; appendix tests 0.75, 1, 1.333 and reports instability at 1.333.
  • Sampling budget B = 2000
    Stage I size of the sampled graph pool; chosen by hand.
  • Sampling Gaussian sigma = 2
    Width of truncated Gaussian over graph orders; chosen by hand.
  • Top-k supervision graphs per task = 2
    Stage I keeps the two best graphs per task; no sensitivity analysis.
  • Optimization queries Q per benchmark = 100 or 200
    Number of queries used to score candidate graphs; the paper does not report whether these are disjoint from the final evaluation set.
  • Node mask threshold = 0.5
    Sigmoid threshold for hard pruning at inference; chosen by hand.
  • Communication rounds K = 3
    Fixed number of dialogue rounds plus a decision agent in all experiments.
  • Agent pool size Nmax = 15
    Fixed roster of hand-authored roles; the paper does not test sensitivity to pool size.
assumptions (4)
  • domain assumption The hand-authored 15-agent pool suffices for all six benchmarks.
    Table A5 lists roles; performance depends on this pool, with no pool-ablation study.
  • domain assumption Three communication rounds with a decision agent produce reliable utility estimates U(A|Q).
    Candidate graphs are scored under this protocol; no study of K.
  • ad hoc to paper Graphs mined from Q=100/200 queries per benchmark are valid supervision for the full benchmark and for unseen tasks.
    No held-out split or cross-task evaluation is reported; this is the paper's weakest premise.
  • ad hoc to paper Complete subgraphs (cliques) of the pool suffice to represent optimal communication topologies.
    Section 3.2 samples only complete subgraphs, so edge-weight supervision cannot encode arbitrary sparse graphs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Adaptive Graph Pruning for Multi-Agent Communication." pith.science (2026). https://pith.science/paper/IDIBSDMA

@misc{pith2026250602951,
  author       = {Pith},
  title        = {Pith review of: Adaptive Graph Pruning for Multi-Agent Communication},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IDIBSDMA}},
  note         = {Machine review of arXiv:2506.02951}
}
abstract

Large Language Model (LLM) based multi-agent systems have shown remarkable performance in various tasks, especially when enhanced through collaborative communication. However, current methods often rely on a fixed number of agents and static communication structures, limiting their ability to adapt to varying task complexities. In this paper, we propose Adaptive Graph Pruning (AGP), a novel task-adaptive multi-agent collaboration framework that jointly optimizes agent quantity (hard-pruning) and communication topology (soft-pruning). Specifically, our method employs a two-stage training strategy: firstly, independently training soft-pruning networks for different agent quantities to determine optimal agent-quantity-specific complete graphs and positional masks across specific tasks; and then jointly optimizing hard-pruning and soft-pruning within a maximum complete graph to dynamically configure the number of agents and their communication topologies per task. Extensive experiments demonstrate that our approach is: (1) High-performing, achieving state-of-the-art results across six benchmarks and consistently generalizes across multiple mainstream LLM architectures, with a increase in performance of $2.58\%\sim 9.84\%$; (2) Task-adaptive, dynamically constructing optimized communication topologies tailored to specific tasks, with an extremely high performance in all three task categories (general reasoning, mathematical reasoning, and code generation); (3) Token-economical, having fewer training steps and token consumption at the same time, with a decrease in token consumption of $90\%+$; and (4) Training-efficient, achieving high performance with very few training steps compared with other methods. The performance will surpass the existing baselines after about ten steps of training under six benchmarks.

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. MasFACT: Continual Multi-Agent Topology Learning via Geometry-Aware Posterior Transfer

    cs.LG 2026-05 unverdicted novelty 7.0 of 10

    MasFACT transfers historical topology priors across tasks via Fused Gromov-Wasserstein optimal transport and PAC-Bayes conservative adaptation to reduce topology forgetting in continual multi-agent settings.

  2. From Cognitive Architectures to Language Agents: A Mechanism-Level Review of Lineage, Convergence, and Migration Gaps

    cs.AI 2026-07 conditional novelty 6.0 of 10

    Coding each mechanism for evidence of lineage and implementation depth, the review closes one candidate gap (GraSP) and isolates five residual control bundles for language agents.

  3. Graphs Meet AI Agents: Taxonomy, Progress, and Future Opportunities

    cs.AI 2025-06 conditional novelty 5.0 of 10

    A survey that groups graph-empowered AI agent research into planning, execution, memory, and multi-agent coordination, plus agents-for-graphs and applications.

Reference graph

Works this paper leans on

48 extracted references · 18 canonical work pages · cited by 3 Pith papers

  1. [1]

    Besta, N

    M. Besta, N. Blach, A. Kubicek, R. Gerstenberger, L. Gianinazzi, J. Gajda, T. Lehmann, M. Podstawski, H. Niewiadomski, P. Nyczyk, and T. Hoefler. Graph of thoughts: Solving elaborate problems with large language models, August 01, 2023 2023

    work page 2023

  2. [2]

    C.-M. Chan, W. Chen, Y . Su, J. Yu, W. Xue, S. Zhang, J. Fu, and Z. Liu. ChatEval: Towards Better LLM-based Evaluators through Multi-Agent Debate. arXiv e-prints, Aug. 2023

    work page 2023

  3. [3]

    G. Chen, S. Dong, Y . Shu, G. Zhang, J. Sesay, B. F. Karlsson, J. Fu, and Y . Shi. Autoagents: A framework for automatic agent generation.arXiv preprint arXiv:2309.17288, 2023

    arXiv 2023

  4. [4]

    L. Chen, J. Q. Davis, B. Hanin, P. Bailis, I. Stoica, M. Zaharia, and J. Zou. Are more llm calls all you need? towards scaling laws of com- pound inference systems. arXiv preprint arXiv:2403.02419, 2024

    arXiv 2024

  5. [5]

    M. Chen, J. Tworek, H. Jun, and etc. Evaluating large language models trained on code, July 01, 2021 2021

    work page 2021

  6. [6]

    W. Chen, Y . Su, J. Zuo, C. Yang, C. Yuan, C. Qian, C.-M. Chan, Y . Qin, Y . Lu, R. Xie, Z. Liu, M. Sun, and J. Zhou. Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors in agents, 2023

    work page 2023

  7. [7]

    Cobbe, V

    K. Cobbe, V . Kosaraju, M. Bavarian, M. Chen, H. Jun, L. Kaiser, M. Plappert, J. Tworek, J. Hilton, R. Nakano, C. Hesse, and J. Schul- man. Training verifiers to solve math word problems. arXiv prepring, abs/2110.14168, 2021

    arXiv 2021

  8. [8]

    Y . Du, S. Li, A. Torralba, J. B. Tenenbaum, and I. Mordatch. Improving factuality and reasoning in language models through multiagent debate. CoRR, abs/2305.14325, 2023

    arXiv 2023

Show all 48 references
  1. [9]

    Fernando, D

    C. Fernando, D. Banarse, H. Michalewski, S. Osindero, and T. Rock- täschel. Promptbreeder: Self-referential self-improvement via prompt evolution. arXiv preprint arXiv:2309.16797, 2023

    2023 arXiv

  2. [10]

    Y . Fu, H. Peng, A. Sabharwal, P. Clark, and T. Khot. Complexity-based prompting for multi-step reasoning. In The Eleventh International Con- ference on Learning Representations, 2022

    2022

  3. [11]

    Q. Guo, R. Wang, J. Guo, B. Li, K. Song, X. Tan, G. Liu, J. Bian, and Y . Yang. Connecting large language models with evolution- ary algorithms yields powerful prompt optimizers. arXiv preprint arXiv:2309.08532, 2023

    2023 arXiv

  4. [12]

    R. Hao, L. Hu, W. Qi, Q. Wu, Y . Zhang, and L. Nie. Chatllm network: More brains, more intelligence, April 01, 2023 2023

    2023

  5. [13]

    Hendrycks, C

    D. Hendrycks, C. Burns, S. Basart, A. Zou, M. Mazeika, D. Song, and J. Steinhardt. Measuring massive multitask language understanding. Proceedings of the International Conference on Learning Representa- tions (ICLR), 2021

    2021

  6. [14]

    S. Holt, M. R. Luyten, and M. van der Schaar. L2mac: Large lan- guage model automatic computer for extensive code generation. In The Twelfth International Conference on Learning Representations, 2024

    2024

  7. [15]

    S. Hong, X. Zheng, J. Chen, Y . Cheng, J. Wang, C. Zhang, Z. Wang, S. K. S. Yau, Z. Lin, L. Zhou, C. Ran, L. Xiao, and C. Wu. Metagpt: Meta programming for multi-agent collaborative framework, August 01, 2023 2023

    2023

  8. [16]

    S. Hu, C. Lu, and J. Clune. Automated design of agentic systems. arXiv preprint arXiv:2408.08435, 2024

    2024 arXiv

  9. [17]

    S. Hu, L. Shen, Y . Zhang, and D. Tao. Learning multi-agent communication from graph modeling perspective. arXiv preprint arXiv:2405.08550, 2024

    2024 arXiv

  10. [18]

    Y . Hu, Y . Cai, Y . Du, X. Zhu, X. Liu, Z. Yu, Y . Hou, S. Tang, and S. Chen. Self-evolving multi-agent collaboration networks for software development. arXiv preprint arXiv:2410.16946, 2024

    2024 arXiv

  11. [19]

    Ishibashi and Y

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

    2024 arXiv

  12. [20]

    Jiang, X

    D. Jiang, X. Ren, and B. Y . Lin. LLM-blender: Ensembling large lan- guage models with pairwise ranking and generative fusion. In Proceed- ings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages 14165–14178, Toronto, Can...

    2023

  13. [21]

    Khattab, A

    O. Khattab, A. Singhvi, P. Maheshwari, Z. Zhang, K. Santhanam, S. Vardhamanan, S. Haq, A. Sharma, T. T. Joshi, H. Moazam, et al. Dspy: Compiling declarative language model calls into self-improving pipelines. arXiv preprint arXiv:2310.03714, 2023

    arXiv

  14. [22]

    B. Li, Z. Zhao, D.-H. Lee, and G. Wang. Adaptive graph pruning for multi-agent communication. arXiv preprint arXiv:2506.02951, 2025

    2025 arXiv

  15. [23]

    W. Ling, D. Yogatama, C. Dyer, and P. Blunsom. Program induction by rationale generation: Learning to solve and explain algebraic word problems. arXiv preprint arXiv:1705.04146, 2017

    2017 arXiv

  16. [24]

    Z. Liu, Y . Zhang, P. Li, Y . Liu, and D. Yang. Dynamic llm-agent net- work: An llm-agent collaboration framework with agent team optimiza- tion. CoRR, abs/2310.02170, 2023

    2023 arXiv

  17. [25]

    Patel, S

    A. Patel, S. Bhattamishra, and N. Goyal. Are nlp models really able to solve simple math word problems? arXiv preprint arXiv:2103.07191, 2021

    2021 arXiv

  18. [26]

    Pesce and G

    E. Pesce and G. Montana. Learning multi-agent coordination through connectivity-driven communication. Machine Learning, 112(2):483– 514, 2023

    2023

  19. [27]

    C. Qian, X. Cong, C. Yang, W. Chen, Y . Su, J. Xu, Z. Liu, and M. Sun. Communicative agents for software development, July 01, 2023 2023. 25 pages, 9 figures, 2 tables

    2023

  20. [28]

    C. Qian, Z. Xie, Y . Wang, W. Liu, Y . Dang, Z. Du, W. Chen, C. Yang, Z. Liu, and M. Sun. Scaling large-language-model-based multi-agent collaboration. arXiv preprint arXiv:2406.07155, 2024

    2024 arXiv

  21. [29]

    Roy and D

    S. Roy and D. Roth. Solving general arithmetic word problems. arXiv preprint arXiv:1608.01413, 2016

    2016 arXiv

  22. [30]

    Shang, Y

    Y . Shang, Y . Li, K. Zhao, L. Ma, J. Liu, F. Xu, and Y . Li. Agentsquare: Automatic llm agent search in modular design space. arXiv preprint arXiv:2410.06153, 2024

    2024 arXiv

  23. [31]

    Shinn, B

    N. Shinn, B. Labash, and A. Gopinath. Reflexion: an autonomous agent with dynamic memory and self-reflection. arXiv preprint , abs/2303.11366, 2023. doi: 10.48550/arXiv.2303.11366. URL https: //doi.org/10.48550/arXiv.2303.11366

    2023 doi arXiv

  24. [32]

    G. Wang, Y . Xie, Y . Jiang, A. Mandlekar, C. Xiao, Y . Zhu, L. Fan, and A. Anandkumar. V oyager: An Open-Ended Embodied Agent with Large Language Models. arXiv e-prints, art. arXiv:2305.16291, May 2023

    2023 arXiv

  25. [33]

    X. Wang, J. Wei, D. Schuurmans, Q. V . Le, E. H. Chi, S. Narang, A. Chowdhery, and D. Zhou. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations, 2023

    2023

  26. [34]

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. Chi, Q. Le, and D. Zhou. Chain-of-thought prompting elicits reasoning in large language models, January 01, 2022 2022

    2022

  27. [35]

    Q. Wu, G. Bansal, J. Zhang, Y . Wu, S. Zhang, E. Zhu, B. Li, L. Jiang, X. Zhang, and C. Wang. Autogen: Enabling next-gen llm applications via multi-agent conversation framework, August 01, 2023 2023

    2023

  28. [36]

    S. Yao, D. Yu, J. Zhao, I. Shafran, T. L. Griffiths, Y . Cao, and K. Narasimhan. Tree of thoughts: Deliberate problem solving with large language models, May 01, 2023 2023

    2023

  29. [37]

    S. Yuan, K. Song, J. Chen, X. Tan, D. Li, and D. Yang. Evoagent: Towards automatic multi-agent generation via evolutionary algorithms. arXiv preprint arXiv:2406.14228, 2024

    2024 arXiv

  30. [38]

    Zhang, Y

    G. Zhang, Y . Yue, Z. Li, S. Yun, G. Wan, K. Wang, D. Cheng, J. X. Yu, and T. Chen. Cut the crap: An economical communication pipeline for llm-based multi-agent systems. arXiv preprint arXiv:2410.02506 , 2024

    2024 arXiv

  31. [39]

    Zhang, Y

    G. Zhang, Y . Yue, X. Sun, G. Wan, M. Yu, J. Fang, K. Wang, T. Chen, and D. Cheng. G-designer: Architecting multi-agent communication topologies via graph neural networks.arXiv preprint arXiv:2410.11782, 2024

    2024 arXiv

  32. [40]

    Zhang, L

    G. Zhang, L. Niu, J. Fang, K. Wang, L. Bai, and X. Wang. Multi- agent architecture search via agentic supernet. arXiv preprint arXiv:2502.04180, 2025

    2025 arXiv

  33. [41]

    Zhang, X

    J. Zhang, X. Xu, and S. Deng. Exploring collaboration mecha- nisms for llm agents: A social psychology view. arXiv preprint arXiv:2310.02124, 2023

    2023 arXiv

  34. [42]

    Zhang, J

    J. Zhang, J. Xiang, Z. Yu, F. Teng, X. Chen, J. Chen, M. Zhuge, X. Cheng, S. Hong, J. Wang, et al. Aflow: Automating agentic workflow generation. arXiv preprint arXiv:2410.10762, 2024

    2024 arXiv

  35. [43]

    Z. Zhao, W. Chai, X. Wang, L. Boyi, S. Hao, S. Cao, T. Ye, J.-N. Hwang, and G. Wang. See and think: Embodied agent in virtual environment. arXiv preprint arXiv:2311.15209, 2023

    2023 arXiv

  36. [44]

    Z. Zhao, K. Chen, D. Guo, W. Chai, T. Ye, Y . Zhang, and G. Wang. Hierarchical auto-organizing system for open-ended multi-agent navi- gation. arXiv preprint arXiv:2403.08282, 2024

    2024 arXiv

  37. [45]

    Z. Zhao, K. Ma, W. Chai, X. Wang, K. Chen, D. Guo, Y . Zhang, H. Wang, and G. Wang. Do we really need a complex agent sys- tem? distill embodied agent into a single model. arXiv preprint arXiv:2404.04619, 2024

    2024 arXiv

  38. [46]

    Z. Zhao, W. Zhang, H. Huang, K. Liu, J. Gao, G. Wang, and K. Chen. Rig: Synergizing reasoning and imagination in end-to-end generalist policy. arXiv preprint arXiv:2503.24388, 2025

    2025 arXiv

  39. [47]

    Z. Zhou, B. Hu, C. Zhao, P. Zhang, and B. Liu. Large language model as a policy teacher for training reinforcement learning agents. arXiv preprint arXiv:2311.13373, 2023

    2023 arXiv

  40. [48]

    Dialogue History

    M. Zhuge, W. Wang, L. Kirsch, F. Faccio, D. Khizbullin, and J. Schmid- huber. Gptswarm: Language agents as optimizable graphs. InForty-first International Conference on Machine Learning, 2024. Prompt Format for a Single Agent of AGP in mathematical Reasoning Input Format: • Ta...

    2024

Pith tools

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