Pith. sign in

REVIEW 4 major objections 4 minor 14 references

This paper claims that collaboration failures in unstructured multi-agent LLM systems can be detected and corrected in real time from the causal structure of agent interactions alone, without access to agent internals or task semantics.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

A dynamic interaction graph can expose and partially heal structural failures in emergent multi-agent LLM collaboration, with mixed empirical support.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection A genuinely useful graph-based monitoring framework for emergent multi-agent LLM collaboration, but the 'protocol-agnostic' claim is untested because the evaluation forces agents to emit the exact labels DIG is built from. the 4 major comments →

arxiv 2603.00309 v2 pith:4RVNVVGG submitted 2026-02-27 cs.AI cs.MA

DIG to Heal: Scaling General-purpose Agent Collaboration via Explainable Dynamic Decision Paths

Hanqing Yang , Hyungwoo Lee , Yuhang Yao , Zhiwei Liu , Kay Liu , Jingdi Chen , Carlee Joe-Wong This is my paper
classification cs.AI cs.MA
keywords Dynamic Interaction Graphmulti-agent LLM systemsemergent collaborationfailure detectiongraph rewritingexplainabilitysystem healingagent interaction traces
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 argues that failures in unstructured multi-LLM agent systems can be understood and repaired from interaction structure alone, without reading agent internals or task semantics. It introduces the Dynamic Interaction Graph (DIG), a time-evolving bipartite causal graph of agent activations and events, and defines local graph rewrites—consume, wait, reroute, discard, and submit—that convert raw execution traces into labeled causal paths. On this structure it defines a failure taxonomy (early or missing termination, orphaned events, deadlock, excessive rerouting, cross-lineage aggregation, repeated subproblem solving) and a healing mechanism that injects or reroutes information in real time. In verifiable two-task evaluations with homogeneous and heterogeneous agent teams, DIG-based monitoring cuts final task error in most settings, including an 83% reduction on one parallelizable task and 44% and 18% reductions on the two dependency-constrained settings, at the cost of detecting more failures and sometimes running longer.

Core claim

DIG makes emergent collaboration observable by turning every agent activation and message delivery into nodes and edges of a labeled bipartite directed acyclic graph, with each edge labeled by how the recipient handled it. The paper's claim is that cooperation patterns, and the failures they produce, live in this graph: a topological characterization maps specific graph shapes—unreachable work, empty recipient sets, reroute loops, cross-lineage merges, overlapping consumption—to distinct error types, and any observable system-level diagnosis function reduces to inference over the DIG (Theorem 1). Therefore monitoring and healing need not look inside agents or understand the task; they operat

What carries the argument

The Dynamic Interaction Graph and its canonical edge rewrite operators form the central mechanism. Agent activation nodes and event nodes make up a bipartite DAG; each incoming edge receives one of four fates—CONSUME, DELAY (wait), REROUTE, or DISCARD—and these fates act as local graph rewrites that update the DIG. The failure taxonomy and healing logic read these labels and the reachability structure: for example, an event whose recipient set is empty becomes an orphaned event, repeated REROUTEs become excessive rerouting, and a SUBMIT issued while reachable work remains is early termination. The graph carries the argument because all diagnosis is a function of its topology and edge labels.

Load-bearing premise

The construction assumes that every agent interaction can be reliably labeled with an action primitive (CONSUME, WAIT, REROUTE, DISCARD, SUBMIT) and that event payloads and recipient sets are observable; in the experiments this is guaranteed by forcing agents to output structured input_actions, so without that guarantee the DIG cannot be built from truly unconstrained natural-language conversations.

What would settle it

Run the same two tasks with agents whose messages are free-form natural language and no structured output protocol, then attempt to reconstruct DIG from the messages alone; if the edge-label mapping is unrecoverable or ambiguous, the protocol-agnostic claim is falsified. Alternatively, a task where unseen structural errors produce DIG topologies identical to successful ones would falsify the topological characterization.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the graph-level characterization holds, real-time failure detection in multi-agent LLM systems can run without semantic judgment or access to agent reasoning, detecting deadlocks, premature submission, orphaned work, and redundant effort purely from trace structure.
  • Healing can be automated at the interaction level: injecting guidance, rerouting events, or creating system events when a pattern is detected, without retraining or modifying agents.
  • The approach scales to dense collaborations: in the paper's 20-agent case study, the DIG-enabled system produced a valid answer within 70 seconds where the unmonitored system failed to finish in 120 seconds, with 16 structural failures detected and corrected.
  • The framework is intended as a compatibility layer: topology-level signals can feed into semantic or behavioral error detectors rather than replacing them, so existing failure-attribution methods can be augmented.
  • Because DIG treats interaction structure as the object of analysis, it offers a protocol-agnostic ground for comparing different agent designs and for studying how cooperation emerges.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: If DIG is applied to agents that communicate in unrestricted natural language, the edge-labeling assumption becomes the bottleneck; the paper's evaluation forces agents to emit structured input_actions, so extending to free-text messages requires an additional parser whose reliability is not established. The absence of such a parser is the main openness gap.
  • Editorial inference: The graph-rewrite view suggests a testable prediction: error frequency should be predictable from local topology statistics—reroute counts, orphaned-event counts, lineage-merger counts—before task completion, enabling classifiers trained on DIG features rather than agent outputs.
  • Editorial inference: The failure taxonomy may generalize beyond LLM agents to any asynchronous message-passing system with comparable primitives, such as human crowdsourced workflows; if the same topology-to-failure mapping appears there, DIG would support a general theory of collaboration failure rather than one specific to LLM agents.
  • Editorial inference: The trade-off observed—more detected failures and longer runtime with healing—suggests that naive structure-driven intervention can overcorrect; a cost-sensitive policy that learns which patterns warrant intervention versus mere logging would be a natural extension the paper does not develop.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper introduces the Dynamic Interaction Graph (DIG), a time-evolving bipartite causal graph meant to represent asynchronous interactions among general-purpose LLM agents that collaborate without predefined roles or control flow. DIG labels each interaction edge with one of four primitives (CONSUME/WAIT/REROUTE/DISCARD) and defines local graph rewrite operators. The authors claim a topological characterization of collaboration-induced failures, a theorem that all observable system-level reasoning reduces to inference over DIG, and a structure-driven healing mechanism. They evaluate on two synthetic tasks (Count Frequency and Research Job) with homogeneous and heterogeneous agent teams, reporting error reductions, detection counts, and runtime, plus a 20-agent case study.

Significance. If the central claims held, DIG would offer a general, protocol-agnostic layer for monitoring and correcting emergent multi-agent LLM collaboration without access to agent internals or task semantics. The paper's conceptual framework is original and the evaluation setting is verifiable; the inclusion of per-run statistics in Appendix E and the 20-agent case study are strengths. However, the load-bearing assumption that edge labels are observable from natural-language interactions is not tested, and the empirical support for the headline error-reduction claim is statistically weak in most settings. The theoretical result is a tautology. The paper is therefore a promising but not yet established contribution.

major comments (4)
  1. [§4.2, Appendix B] The entire DIG construction, rewrite dynamics, and failure taxonomy depend on the per-edge action labels φ_v ∈ {CONSUME, DELAY, REROUTE, DISCARD}. In the evaluation, these labels are not inferred from natural language: Appendix B forces every agent to emit structured `input_actions` with an action and `reroute_to` recipients. This is a communication protocol, not a protocol-agnostic trace. The paper claims to be 'protocol-agnostic' (Fig. 1, §4), but no evidence is provided that arbitrary natural-language messages can be mapped reliably to these four primitives. A single message can simultaneously answer, forward, and request, so the labeling step is underdetermined. This is a load-bearing gap: without an automatic labeling mechanism, DIG cannot be built in the claimed general setting. Please either specify and evaluate such a mechanism, or substantially soften the protocol-agnostic claim
  2. [§4.3, Appendix A] Theorem 1 is effectively a restatement of its own assumption. The proof defines I(Ψ(T)) = Φ(T) by choosing a representative trace, which is well-defined only because Eq. (10) already assumes Φ is invariant over traces with isomorphic DIGs. The theorem therefore does not establish any substantive reduction; it merely rephrases the invariance condition. Since the paper presents this as a formal foundation for 'topological inference reduction,' it should be reframed as a definitional observation or removed. It does not, by itself, justify that failures detected from DIG structure correspond to actual task errors; final task error is measured externally via Eq. (9).
  3. [§7, Table 3] The text states that 'DIG consistently reduces final error in all settings,' but the significance tests in Table 3 show p<0.05 for task error only in three CF settings (4-agent homo 8000, 4-agent hetero 2222, 6-agent homo 12000). In the remaining five settings—including both Research Job rows—the task-error p-values are 0.061, 0.096, 0.124, 0.271, and 0.867, none significant at conventional levels. The claim of consistent improvement is therefore not statistically supported. Additionally, the MAS+LLM Judge baseline is only shown qualitatively in Fig. 5; no quantitative comparison of error, detection, or runtime is reported, so the statement that DIG 'avoids expensive semantic judging and provides more targeted intervention' is unsupported.
  4. [§5, §7] Detection and healing depend on several unspecified thresholds: 'reasonable time window' (MC, OE, DL) and 'reasonable number of times' (ER). No values are given in the main text or appendices, and no sensitivity analysis is reported. These free parameters can substantially affect how many failures are detected and healed, and therefore the reported error reductions are not fully reproducible. Please specify the thresholds and, ideally, show that the main results are robust to their variation.
minor comments (4)
  1. [§4.2] Terminology is inconsistent: the mapping φ_v is defined as {CONSUME, DELAY, REROUTE, DISCARD}, but the operators are called Wait (W) and the taxonomy uses 'WAIT'. Please unify the term (e.g., use 'WAIT' throughout).
  2. [Abstract and §8] The phrase 'for the first time' is used twice and is difficult to verify. I recommend softening to 'a framework that enables' or similar, unless a systematic comparison with all prior interaction-modeling approaches is provided.
  3. [§6] The paper claims to introduce 'the first dataset of execution traces' but no data repository or artifact URL is provided (only a project webpage). If the traces are part of the contribution, please include a link or appendix with the raw data.
  4. [Appendix G] The trace figures (Figs. 7–13) are very small and difficult to read in the PDF. Higher-resolution versions or vector graphics would help the reader verify the qualitative claims about interaction patterns.

Circularity Check

1 steps flagged

Theorem 1 is a definitional tautology (I is defined as Φ of a representative trace), so the formal 'reduction' restates its own invariance premise; the empirical evaluations provide independent grounding, and no load-bearing self-citation chain is present.

specific steps
  1. self definitional [Sec. 4.3, Theorem 1; Appendix A.1, Eqs. (10)-(12)]
    "For any system-level functional Φ : T→ F that depends only on observable interactions, equivalently, is invariant over traces with isomorphic DIGs, there exists a graph functional I:{G(t)} t∈Z → F such that, for every trace T ∈T, Φ(T) = I(Ψ(T)). ... Define I on the range of Ψ by selecting any trace T ∈T such that Ψ(T) ={G(t)} t∈Z and setting I({G(t)} t∈Z) ≜ Φ(T)."

    The theorem's conclusion is contained in its assumption. Invariance under DIG isomorphism is exactly the statement that Φ factors through Ψ; the proof then defines I to be Φ evaluated on a representative trace. No use is made of DIG topology, edge labels, causality, or any derived property. Eq. (12) is true by the definition in Eq. (11), so the claimed 'topological inference reduction' is a vacuous reformulation rather than an derived result.

full rationale

The only clear circularity is Theorem 1, which is a tautology: the proof constructs I by picking a representative trace and setting I(Ψ(T)) = Φ(T), so under the stated invariance assumption the result is immediate by definition. This does not independently establish that DIG topology enables failure inference; it merely renames the invariance assumption. The rest of the paper has substantial independent content: final task error is measured externally via coverage/over in Eq. (9), and the MAS+DIG vs MAS-Only comparisons are empirical, not forced by the construction. The largest validity gap, not a circular derivation, is that the edge labels CONSUME/WAIT/REROUTE/DISCARD are supplied by the agents themselves under the structured protocol in Appendix B, so the protocol-agnostic claim is untested for natural-language interaction. The paper's self-citations are not load-bearing. Overall, the formal centerpiece is definitional, but the empirical failure-detection and healing results are externally grounded, warranting a moderate circularity score rather than a high one.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The framework's load-bearing assumptions are observational: reliable action labels and well-formed events. The formal theorem is a tautology. No new physical or conceptual entities beyond the DIG representation are introduced.

free parameters (2)
  • detection time-window threshold
    Detection rules for missing termination, orphaned events, and deadlock use 'reasonable time window'; the value is never reported or analyzed.
  • excessive-rerouting count threshold
    Excessive rerouting detection depends on 'a reasonable number of times', which is undefined and could be tuned per setting.
axioms (4)
  • domain assumption Agents' observable outputs include reliable structured action labels (input_actions) mapping each event to CONSUME/WAIT/REROUTE/DISCARD.
    Sec. 4.2 defines the edge-label function from these labels, and Appendix B forces the format in experiments. Raw LLM outputs do not naturally provide them.
  • domain assumption The system-level functional Φ depends only on observable interactions and is invariant under DIG isomorphism.
    Theorem 1 assumes this rather than proving it, making the theorem a restatement of the assumption.
  • domain assumption Event payloads and delivery policies are fully observable and well-formed.
    Sec. 3 models events as (payload, policy) pairs; real LLM agent messages may not have explicit recipient policies.
  • domain assumption LLM agents follow the prompt instructions and output machine-readable actions accurately.
    The experiments depend on agents staying in the required output format; deviations would break DIG construction.

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of DIG to Heal: Scaling General-purpose Agent Collaboration via Explainable Dynamic Decision Paths." pith.science (2026). https://pith.science/paper/4RVNVVGG

@misc{pith2026260300309,
  author       = {Pith},
  title        = {Pith review of: DIG to Heal: Scaling General-purpose Agent Collaboration via Explainable Dynamic Decision Paths},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4RVNVVGG}},
  note         = {Machine review of arXiv:2603.00309}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The increasingly popular agentic AI paradigm promises to harness the power of multiple, general-purpose large language model (LLM) agents to collaboratively complete complex tasks. While many agentic AI systems reduce complexity through predefined workflows or fixed agent roles, the ideal is to support truly autonomous agents capable of emergent collaboration across many interacting agents. Yet in practice, such unstructured interactions often lead to redundant work and cascading failures that are difficult to interpret or correct. In this work, we study multi-agent systems composed of general-purpose LLM agents that solve problems through emergent collaboration, without relying on predefined roles, control flows, or communication constraints. We introduce the Dynamic Interaction Graph (DIG), which captures emergent collaboration as a time-evolving causal network of agent activations and interactions. DIG makes emergent collaboration observable and explainable for the first time, enabling real-time identification, explanation, and correction of collaboration-induced error patterns directly from agents' collaboration paths. Thus, DIG fills a critical gap in understanding how general LLM agents solve problems together in truly agentic multi-agent systems. The project webpage can be found at: https://happyeureka.github.io/dig.

Figures

Figures reproduced from arXiv: 2603.00309 by Carlee Joe-Wong, Hanqing Yang, Hyungwoo Lee, Jingdi Chen, Kay Liu, Yuhang Yao, Zhiwei Liu.

Figure 1
Figure 1. Figure 1: Cooperative problem solving with general-purpose agents. A pool of autonomous agents works on a shared task without prede￾fined roles, control flow, or communication con￾straints, each operating independently and non￾deterministically and interacting through emer￾gent cooperative strategies. We make emergent cooperation analyzable by modeling agent inter￾action in a protocol-agnostic manner [PITH_FULL_IMA… view at source ↗
Figure 3
Figure 3. Figure 3: Structural failure patterns in DIG, showing collaboration-invariant vi￾olations as interaction structure. The system defines errors not by agents’ Internal Rea￾soning, but by observable Interaction Primitives (actions taken by agents) and the corresponding DIG structures. Failure Taxonomy. Empirically, cooperative general agent systems exhibit failures that arise not from iso￾lated reasoning mistakes withi… view at source ↗
Figure 4
Figure 4. Figure 4: Task structures. Left: Count Frequency (fully parallelizable), where tasks are decomposed into independent subtasks that can be processed and merged once completed. Right: Research Job (dependency￾constrained), where tasks are organized into stages with inter-stage dependencies, and violations cause process￾ing or merging failures. Open-ended, verifiable task-solving. We introduce a task-solving setting to… view at source ↗
Figure 5
Figure 5. Figure 5: With DIG (a), the timeline of agent activations, event propagation, and edge-level rewrites [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: MAS + DIG (20 agents): Large￾scale DIG showing structured task prop￾agation, controlled rerouting, and stable convergence despite dense interactions and high concurrency. These gains come with a trade-off: DIG keeps the sys￾tem active to repair failures, which may expose addi￾tional failures and increase runtime. MAS appears to have fewer detected errors and shorter runtime simply because it leaves failure… view at source ↗
Figure 7
Figure 7. Figure 7: Count Frequency example (Homogeneous_4000, 4 agents). The problem is within the group’s capability. Agent 1 decomposes the task for parallel execution, but due to asynchrony and partial observability, agents may ignore subtasks, reroute unnecessarily, or duplicate work. These stochastic behaviors make cooperation brittle, which DIG captures and helps diagnose [PITH_FULL_IMAGE:figures/full_fig_p019_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Count Frequency example (Homoge￾neous_8000, 4 agents, w/o DIG). The problem exceeds the group’s capability. Agents continue to decompose tasks, resulting in a more complex structure of problem-solving and interaction [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Count Frequency example (Hetero￾geneous_2222, 4 agents, w/o DIG). The prob￾lem exceeds group capability. Agents decompose tasks based on downstream capacity, leading to uneven interaction patterns [PITH_FULL_IMAGE:figures/full_fig_p020_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: Research domain (Heteroge￾neous_1111, 4 agents, w/o DIG). Agents ini￾tially operate on the same stage and progressively move forward. However, due to the more complex task structure, errors are more likely: agents may merge incompatible results, violate dependencies, or terminate early under partial observability [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

14 extracted references · 11 linked inside Pith

  1. [1]

    wait” instead. If you don’t want it, use “discard

    = Φ(T2).(10) DefineIon the range ofΨby selecting any traceT ∈Tsuch thatΨ(T) ={G(t)} t∈Z and setting I {G(t)}t∈Z ≜Φ(T).(11) This definition is well-defined: if another trace T ′ ∈T satisfies Ψ(T ′) ={G(t)} t∈Z, then Ψ(T ′) ∼= Ψ(T) , so Eq.(10) gives Φ(T ′) = Φ(T) . Hence Eq. (11) does not depend on the chosen representative trace. Finally, for any traceT ∈...

    2008

  2. [9]

    Monitoring llm-based multi-agent systems against corruptions via node evaluation.arXiv preprint arXiv:2510.19420,

    Chengcan Wu, Zhixin Zhang, Mingqian Xu, Zeming Wei, and Meng Sun. Monitoring llm-based multi-agent systems against corruptions via node evaluation.arXiv preprint arXiv:2510.19420,

    Pith/arXiv arXiv

  3. [10]

    Autogen: Enabling next-gen llm applications via multi-agent conversation framework

    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,

    Pith/arXiv arXiv

  4. [12]

    URL https: //arxiv.org/abs/2502.05453. Chengyue Yu, Siyuan Lu, Chenyi Zhuang, Dong Wang, Qintong Wu, Zongyue Li, Runsheng Gan, Chunfeng Wang, Siqi Hou, Gaochi Huang, Wenlong Yan, Lifeng Hong, Aohui Xue, Yanfeng Wang, Jinjie Gu, David Tsai, and Tao Lin. Aworld: Orchestrating the training recipe for agentic ai,

    Pith/arXiv arXiv

  5. [13]

    org/abs/2508.20404

    URL https://arxiv. org/abs/2508.20404. Guibin Zhang, Junhao Wang, Junjie Chen, Wangchunshu Zhou, Kun Wang, and Shuicheng Yan. Agentracer: Who is inducing failure in the llm agentic systems?arXiv preprint arXiv:2509.03312, 2025a. Shaokun Zhang, Ming Yin, Jieyu Zhang, Jiale Liu, Zhiguang Han, Jingyang Zhang, Beibin Li, Chi Wang, Huazheng Wang, Yiran Chen, e...

    Pith/arXiv arXiv

  6. [2002]

    Towards a science of scaling agent systems.arXiv preprint arXiv:2512.08296,

    Yubin Kim, Ken Gu, Chanwoo Park, Chunjong Park, Samuel Schmidgall, A Ali Heydari, Yao Yan, Zhihan Zhang, Yuchen Zhuang, Yun Liu, et al. Towards a science of scaling agent systems.arXiv preprint arXiv:2512.08296,

    Pith/arXiv arXiv

  7. [2008]

    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,

    Pith/arXiv arXiv

  8. [2013]

    When agents go astray: Course- correcting swe agents with prms.arXiv preprint arXiv:2509.02360,

    Shubham Gandhi, Jason Tsay, Jatin Ganhotra, Kiran Kate, and Yara Rizk. When agents go astray: Course- correcting swe agents with prms.arXiv preprint arXiv:2509.02360,

    arXiv

  9. [2014]

    Agent ai: Surveying the horizons of multimodal interaction

    Zane Durante, Qiuyuan Huang, Naoki Wake, Ran Gong, Jae Sung Park, Bidipta Sarkar, Rohan Taori, Yusuke Noda, Demetri Terzopoulos, Yejin Choi, et al. Agent ai: Surveying the horizons of multimodal interaction. arXiv preprint arXiv:2401.03568,

    Pith/arXiv arXiv

  10. [2016]

    Qian Wang, Tianyu Wang, Zhenheng Tang, Qinbin Li, Nuo Chen, Jingsheng Liang, and Bingsheng He

    URLhttps://arxiv.org/abs/1506.05254. Qian Wang, Tianyu Wang, Zhenheng Tang, Qinbin Li, Nuo Chen, Jingsheng Liang, and Bingsheng He. Megaagent: A large-scale autonomous llm-based multi-agent system without predefined sops. InFindings of the Association for Computational Linguistics: ACL 2025, pages 4998–5036,

    Pith/arXiv arXiv 2025

  11. [2023]

    Beyond self-talk: A communication-centric survey of llm-based multi-agent systems

    Bingyu Yan, Zhibo Zhou, Litian Zhang, Lian Zhang, Ziyi Zhou, Dezhuang Miao, Zhoujun Li, Chaozhuo Li, and Xiaoming Zhang. Beyond self-talk: A communication-centric survey of llm-based multi-agent systems. arXiv preprint arXiv:2502.14321,

    Pith/arXiv arXiv

  12. [2024]

    Stop wasting your tokens: Towards efficient runtime multi-agent systems.arXiv preprint arXiv:2510.26585,

    Fulin Lin, Shaowen Chen, Ruishan Fang, Hongwei Wang, and Tao Lin. Stop wasting your tokens: Towards efficient runtime multi-agent systems.arXiv preprint arXiv:2510.26585,

    arXiv

  13. [2025]

    Aegis: Automated error generation and attribution for multi-agent systems.arXiv preprint arXiv:2509.14295,

    Fanqi Kong, Ruijie Zhang, Huaxiao Yin, Guibin Zhang, Xiaofei Zhang, Ziang Chen, Zhaowei Zhang, Xiaoyuan Zhang, Song-Chun Zhu, and Xue Feng. Aegis: Automated error generation and attribution for multi-agent systems.arXiv preprint arXiv:2509.14295,

    Pith/arXiv arXiv

  14. [2026]

    URL https://openreview.net/forum?id= LGsed0QQVq

    ISSN 2835-8856. URL https://openreview.net/forum?id= LGsed0QQVq. Survey Certification. Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chen Qian, Chi-Min Chan, Yujia Qin, Yaxi Lu, Ruobing Xie, et al. Agentverse: Facilitating multi-agent collaboration and exploring emergent behaviors in agents.arXiv preprint arXiv:2308.10848, 2(4):5,

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.