pith. sign in

arxiv: 2604.07767 · v1 · submitted 2026-04-09 · 💻 cs.DC

Administrative Decentralization in Edge-Cloud Multi-Agent for Mobile Automation

Senyao Li , Zhigang Zuo , Haozhao Wang , Junyu Chen , Zhanbo Jin , Ruixuan Li This is my paper

Pith reviewed 2026-05-10 17:54 UTC · model grok-4.3

classification 💻 cs.DC
keywords administrative decentralizationedge-cloud multi-agentmobile automationbimodal edge teamtactical planningtask success ratelatency reduction
0
0 comments X

The pith

AdecPilot decentralizes administration in edge-cloud multi-agent systems so edge agents can handle tactical planning and self-correction independently for mobile automation tasks.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper addresses cognitive lag in existing edge-cloud frameworks where the cloud centrally directs all actions while devices only execute passively. It proposes applying administrative decentralization to separate high-level strategic design from on-device tactical execution. A UI-agnostic cloud component produces abstract milestones while a bimodal edge team manages autonomous planning and correction without further cloud input. A Hierarchical Implicit Termination protocol is added to enforce clean stops and avoid hallucinations after task completion. Experiments show these changes produce measurable gains in success rate, token use, and latency on mobile automation benchmarks.

Core claim

By redefining edge agency through administrative decentralization, AdecPilot pairs a UI-agnostic cloud designer that emits abstract milestones with a bimodal edge team that performs autonomous tactical planning and self-correction. The Hierarchical Implicit Termination protocol supplies deterministic stopping conditions. This structure yields a 21.7 percent higher task success rate and 37.5 percent lower cloud token consumption than EcoAgent together with an 88.9 percent latency reduction versus CORE.

What carries the argument

Administrative decentralization that decouples cloud-generated abstract milestones from an autonomous bimodal edge team plus the Hierarchical Implicit Termination protocol that enforces deterministic completion.

If this is right

  • Task success rate rises 21.7 percent over EcoAgent.
  • Cloud token consumption falls 37.5 percent.
  • End-to-end latency drops 88.9 percent compared with CORE.
  • Edge agents operate without constant cloud oversight for real-time UI changes.
  • Post-completion hallucinations are prevented by deterministic termination.

Where Pith is reading between the lines

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

  • The same separation of strategic milestones from tactical execution could be tested in non-UI domains such as robotic navigation or sensor-driven control loops.
  • Greater edge autonomy might further lower privacy exposure by reducing the volume of UI state sent to the cloud.
  • The termination protocol could be adapted to other multi-agent systems to curb runaway agent behavior after goal achievement.

Load-bearing premise

The bimodal edge team can reliably perform autonomous tactical planning and self-correction without cloud intervention across varied real-world mobile UI dynamics and task types.

What would settle it

Real-world trials in which the edge team repeatedly fails to self-correct on dynamic UI changes and must query the cloud, eliminating the reported latency and token reductions while dropping success rate below baseline.

Figures

Figures reproduced from arXiv: 2604.07767 by Haozhao Wang, Junyu Chen, Ruixuan Li, Senyao Li, Zhanbo Jin, Zhigang Zuo.

Figure 1
Figure 1. Figure 1: Evolution of Mobile Agent Paradigms. Left & Mid￾dle: Conventional methods struggle with high latency (1) or brittle static plans failing under perturbations (2). Right: AdecPilot (3) decouples Strategic Milestones from Tactical Grounding. Autonomous edge Planning ensures robust local resolution, minimal privacy exposure, and minimal cloud consumption. scale with model parameter counts, favoring large cloud… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of AdecPilot. The UI-Agnostic Cloud Designer generates abstract milestones, while the Bimodal Edge Team autonomously executes them via local planning. This hierarchical loop enables real-time self-correction, en￾suring robust and privacy-preserving automation. Task Instruction 𝐿𝑐𝑚𝑑 ∈ L: Natural language goal provided by user. Application Metadata 𝐶𝑚: Static invariant functional schema resolving ap… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the AdecPilot workflow. The Cloud Designer orchestrates Strategic Projection, while the device-based Orchestrator and Executor collaborate to enable autonomous Self-Correction via local planning. Crucially, this closed loop is safeguarded by the HIT protocol, which enforces deterministic termination to prevent post-completion hallucinations. empirical evidence reveals reasoning gap within l… view at source ↗
Figure 4
Figure 4. Figure 4: Performance analysis. (a) Comprehensive comparison across task success rate and token cost. (b) Completion rate [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Holistic System Analysis (a) Average cloud calls (MC) and cloud token usage (MT) comparison within AndroidWorld. (b) [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
read the original abstract

Collaborative edge-cloud frameworks have emerged as the main- stream paradigm for mobile automation, mitigating the latency and privacy risks inherent to monolithic cloud agents. However, existing approaches centralize administration in the cloud while relegating the device to passive execution, inducing a cognitive lag regard- ing real-time UI dynamics. To tackle this, we introduce AdecPilot by applying the principle of administrative decentralization to the edge-cloud multi-agent framework, which redefines edge agency by decoupling high-level strategic designing from tactical grounding. AdecPilot integrates a UI-agnostic cloud designer generating ab- stract milestones with a bimodal edge team capable of autonomous tactical planning and self-correction without cloud intervention. Furthermore, AdecPilot employs a Hierarchical Implicit Termi- nation protocol to enforce deterministic stops and prevent post- completion hallucinations. Extensive experiments demonstrate pro- posed approach improves task success rate by 21.7% while reducing cloud token consumption by 37.5% against EcoAgent and decreas- ing end to end latency by 88.9% against CORE. The source code is available at https://anonymous.4open.science/r/Anonymous_code- B8AB.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes AdecPilot, an edge-cloud multi-agent framework for mobile automation that decentralizes administration by assigning high-level strategic milestone design to a cloud component and autonomous tactical planning plus self-correction to a bimodal edge team. A Hierarchical Implicit Termination protocol is introduced to enforce deterministic task stops. Experiments are claimed to demonstrate a 21.7% higher task success rate and 37.5% lower cloud token consumption versus EcoAgent, together with an 88.9% end-to-end latency reduction versus CORE.

Significance. If the reported gains are reproducible, the work would meaningfully advance practical mobile automation by lowering cloud dependency, token costs, and latency while preserving privacy. The explicit separation of strategy from tactics and the open-source code release are concrete strengths that could influence subsequent multi-agent designs for dynamic UI environments.

major comments (3)
  1. [Abstract and Experiments section] Abstract and Experiments section: the headline metrics (21.7% success-rate gain, 37.5% token reduction vs EcoAgent, 88.9% latency reduction vs CORE) are stated without reporting task-suite composition, number of trials, statistical significance tests, exact baseline re-implementations, or controls for UI variability. These omissions prevent independent verification of the central empirical claims.
  2. [Section 3.2 (Bimodal Edge Team)] Section 3.2 (Bimodal Edge Team): the claim that the edge team performs reliable autonomous tactical planning and self-correction without cloud intervention across varied real-world UI dynamics is load-bearing for all three performance deltas, yet no intervention-frequency statistics, failure-mode analysis, or ablation on ambiguous or rapidly changing UI states are supplied. If fallback frequency is high, both latency and token savings collapse.
  3. [Section 3.3 (Hierarchical Implicit Termination)] Section 3.3 (Hierarchical Implicit Termination): the protocol is described only at a conceptual level; no pseudocode, formal termination condition, or quantitative evaluation of hallucination prevention is given, leaving the deterministic-stop guarantee unverified.
minor comments (2)
  1. [Abstract] The abstract introduces 'bimodal edge team' and 'UI-agnostic cloud designer' without a one-sentence definition or pointer to the corresponding figure; a brief gloss would improve first-read clarity.
  2. [References] Ensure EcoAgent and CORE are cited with full bibliographic details (venue, year) so readers can locate the exact prior implementations used for comparison.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We have addressed each major comment point by point below. Revisions have been made to improve the reproducibility and rigor of the empirical sections without altering the core claims or methodology.

read point-by-point responses

  1. Referee: [Abstract and Experiments section] Abstract and Experiments section: the headline metrics (21.7% success-rate gain, 37.5% token reduction vs EcoAgent, 88.9% latency reduction vs CORE) are stated without reporting task-suite composition, number of trials, statistical significance tests, exact baseline re-implementations, or controls for UI variability. These omissions prevent independent verification of the central empirical claims.

    Authors: We agree that the original presentation omitted key experimental details needed for verification. In the revised manuscript, the Experiments section has been expanded to explicitly report the task-suite composition (including the specific mobile automation scenarios and UI environments used), the total number of trials per condition, statistical significance tests (including p-values and confidence intervals), precise descriptions of how each baseline (EcoAgent and CORE) was re-implemented, and the controls applied for UI variability (such as randomization of UI states and device configurations). These additions enable independent reproduction of the reported gains. revision: yes

  2. Referee: [Section 3.2 (Bimodal Edge Team)] Section 3.2 (Bimodal Edge Team): the claim that the edge team performs reliable autonomous tactical planning and self-correction without cloud intervention across varied real-world UI dynamics is load-bearing for all three performance deltas, yet no intervention-frequency statistics, failure-mode analysis, or ablation on ambiguous or rapidly changing UI states are supplied. If fallback frequency is high, both latency and token savings collapse.

    Authors: The referee correctly identifies that autonomy metrics are essential to substantiate the performance deltas. We have revised Section 3.2 to include intervention-frequency statistics measured across all trials, a comprehensive failure-mode analysis categorizing cases where cloud fallback occurred, and a dedicated ablation study isolating performance under ambiguous and rapidly changing UI states. The added results show low fallback rates, confirming that the edge team's autonomous tactical planning and self-correction hold under the tested conditions and that the latency and token savings are not artifacts of frequent cloud intervention. revision: yes

  3. Referee: [Section 3.3 (Hierarchical Implicit Termination)] Section 3.3 (Hierarchical Implicit Termination): the protocol is described only at a conceptual level; no pseudocode, formal termination condition, or quantitative evaluation of hallucination prevention is given, leaving the deterministic-stop guarantee unverified.

    Authors: We acknowledge that the original description of the Hierarchical Implicit Termination protocol remained at a high level. The revised Section 3.3 now provides pseudocode for the full protocol, a formal mathematical definition of the termination condition (including the hierarchical triggering rules), and quantitative evaluation results from controlled experiments measuring hallucination rates with and without the protocol. These additions verify the deterministic-stop guarantee and its contribution to preventing post-completion hallucinations. revision: yes

Circularity Check

0 steps flagged

No circularity; purely empirical claims with no derivation chain

full rationale

The paper describes an architectural approach (AdecPilot) and reports empirical gains from experiments against baselines (EcoAgent, CORE). No equations, fitted parameters, predictions derived from inputs, or self-citation chains appear in the provided text. The central claims reduce to measured task success, token use, and latency deltas rather than any self-referential construction. The bimodal edge-team autonomy is presented as a design premise whose reliability is asserted via experiment, not derived by definition or prior self-work.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

No free parameters, axioms, or invented entities are explicitly introduced beyond standard concepts in multi-agent and edge-cloud computing; the contribution is the integrated system design.

pith-pipeline@v0.9.0 · 5519 in / 987 out tokens · 36282 ms · 2026-05-10T17:54:14.363332+00:00 · methodology

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

34 extracted references · 34 canonical work pages · 1 internal anchor

  1. [1]

    Theodoros Aslanidis, Sokol Kosta, Spyros Lalis, and Dimitris Chatzopoulos. 2025. Cross-Domain DRL Agents for Efficient Job Placement in the Cloud-Edge Con- tinuum. InProceedings of the 5th Workshop on Machine Learning and Systems, EuroMLSys 2025, World Trade Center, Rotterdam, The Netherlands, 30 March 2025- 3 April 2025. 276–285

    work page 2025

  2. [2]

    Hao Bai, Yifei Zhou, Jiayi Pan, Mert Cemri, Alane Suhr, Sergey Levine, and Aviral Kumar. 2024. DigiRL: Training In-The-Wild Device-Control Agents with Au- tonomous Reinforcement Learning. InAdvances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10...

    work page 2024

  3. [3]

    Ronshee Chawla, Daniel Vial, Sanjay Shakkottai, and R. Srikant. 2023. Col- laborative Multi-Agent Heterogeneous Multi-Armed Bandits. InInternational Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA (Proceedings of Machine Learning Research, Vol. 202). 4189–4217

    work page 2023

  4. [4]

    Xinran Chen, Yuchen Li, Hengyi Cai, Zhuoran Ma, Xuanang Chen, Haoyi Xiong, Shuaiqiang Wang, Ben He, Le Sun, and Dawei Yin. 2025. Multi-agent proactive information seeking with adaptive llm orchestration for non-factoid question answering. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 4341–4352

    work page 2025

  5. [5]

    Tong Cheng, Hang Dong, Lu Wang, Bo Qiao, Si Qin, Qingwei Lin, Dongmei Zhang, Saravan Rajmohan, and Thomas Moscibroda. 2023. Multi-Agent Rein- forcement Learning with Shared Policy for Cloud Quota Management Problem. InCompanion Proceedings of the ACM Web Conference 2023, WWW 2023, Austin, TX, USA, 30 April 2023 - 4 May 2023. 391–395

    work page 2023

  6. [6]

    Alex Clinton, Yiding Chen, Jerry Zhu, and Kirthevasan Kandasamy. 2025. Col- laborative Mean Estimation Among Heterogeneous Strategic Agents: Individual Rationality, Fairness, and Truthful Contribution. InForty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025

    work page 2025

  7. [7]

    Gucongcong Fan, Chaoyue Niu, chengfei lv, Fan Wu, and Guihai Chen. 2025. CORE: Reducing UI Exposure in Mobile Agents via Collaboration Between Cloud and Local LLMs. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems

    work page 2025

  8. [8]

    Boyu Gou, Ruohan Wang, Boyuan Zheng, Yanan Xie, Cheng Chang, Yiheng Shu, Huan Sun, and Yu Su. 2025. Navigating the Digital World as Humans Do: Universal Visual Grounding for GUI Agents. InThe Thirteenth International Conference on Learning Representations

    work page 2025

  9. [9]

    Xuehang Guo, Xingyao Wang, Yangyi Chen, Sha Li, Chi Han, Manling Li, and Heng Ji. 2025. SyncMind: Measuring Agent Out-of-Sync Recovery in Collabora- tive Software Engineering. InForty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025

    work page 2025

  10. [10]

    Xiao Han, Chen Zhu, Hengshu Zhu, and Xiangyu Zhao. 2025. Swarm intelligence in geo-localization: A multi-agent large vision-language model collaborative framework. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 814–825

    work page 2025

  11. [11]

    Senkang Hu, Yihang Tao, Guowen Xu, Yiqin Deng, Xianhao Chen, Yuguang Fang, and Sam Kwong. 2025. CP-Guard: Malicious Agent Detection and Defense in Collaborative Bird’s Eye View Perception. InAAAI-25, Sponsored by the Associ- ation for the Advancement of Artificial Intelligence, February 25 - March 4, 2025, Philadelphia, PA, USA. 23203–23211

    work page 2025

  12. [12]

    Guanjie Huang, Danny HK Tsang, Shan Yang, Guangzhi Lei, and Li Liu. 2025. Cued-Agent: A Collaborative Multi-Agent System for Automatic Cued Speech Recognition. InProceedings of the 33rd ACM International Conference on Multime- dia. 8313–8321

    work page 2025

  13. [13]

    Zaid Khan, Elias Stengel-Eskin, Jaemin Cho, and Mohit Bansal. 2025. DataEn- vGym: Data Generation Agents in Teacher Environments with Student Feedback. InThe Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025

    work page 2025

  14. [14]

    Senyao Li, Haozhao Wang, Wenchao Xu, Rui Zhang, Song Guo, Jingling Yuan, Xian Zhong, Tianwei Zhang, and Ruixuan Li. 2025. Collaborative inference and learning between edge slms and cloud LLMs: A survey of algorithms, execution, and open challenges.arXiv preprint arXiv:2507.16731(2025)

    work page arXiv 2025

  15. [15]

    Kevin Qinghong Lin, Linjie Li, Difei Gao, Zhengyuan Yang, Shiwei Wu, Zechen Bai, Stan Weixian Lei, Lijuan Wang, and Mike Zheng Shou. 2025. Showui: One vision-language-action model for gui visual agent. InProceedings of the Computer Vision and Pattern Recognition Conference. 19498–19508

    work page 2025

  16. [16]

    Wenjin Liu, Haoran Luo, Xueyuan Lin, Haoming Liu, Tiesunlong Shen, Jiapu Wang, Rui Mao, and Erik Cambria. 2025. Prompt-R1: Collaborative Automatic Prompting Framework via End-to-end Reinforcement Learning.arXiv preprint arXiv:2511.01016(2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    Yuhang Liu, Pengxiang Li, Zishu Wei, Congkai Xie, Xueyu Hu, Xinchen Xu, Shengyu Zhang, Xiaotian Han, Hongxia Yang, and Fei Wu. 2026. Infiguiagent: A multimodal generalist gui agent with native reasoning and reflection. InPro- ceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). 1035–1051

    work page 2026

  18. [18]

    Lin Long, Yichen He, Wentao Ye, Yiyuan Pan, Yuan Lin, Hang Li, Junbo Zhao, and Wei Li. 2025. Seeing, listening, remembering, and reasoning: A multimodal agent with long-term memory.arXiv preprint arXiv:2508.09736(2025)

    work page arXiv 2025

  19. [19]

    Jiayuan Rao, Zifeng Li, Haoning Wu, Ya Zhang, Yanfeng Wang, and Weidi Xie

    work page

  20. [20]

    InProceedings of the 33rd ACM International Conference on Multimedia

    Multi-agent system for comprehensive soccer understanding. InProceedings of the 33rd ACM International Conference on Multimedia. 3654–3663

    work page

  21. [21]

    Chris Rawles, Sarah Clinckemaillie, Yifan Chang, Jonathan Waltz, Gabrielle Lau, Marybeth Fair, Alice Li, William Bishop, Wei Li, Folawiyo Campbell-Ajala, Daniel Toyama, Robert Berry, Divya Tyamagundlu, Timothy Lillicrap, and Oriana Riva

    work page

  22. [22]

    InInternational Conference on Representation Learning

    AndroidWorld: A Dynamic Benchmarking Environment for Autonomous Agents. InInternational Conference on Representation Learning. 406–441

    work page

  23. [23]

    Chenyang Shao, Xinyuan Hu, Yutang Lin, and Fengli Xu. 2025. Division-of- Thoughts: Harnessing Hybrid Language Model Synergy for Efficient On-Device Agents. InProceedings of the ACM on Web Conference 2025, WWW 2025, Sydney, NSW, Australia, 28 April 2025- 2 May 2025. 1822–1833

    work page 2025

  24. [24]

    Hongda Sun, Hongzhan Lin, Haiyu Yan, Yang Song, Xin Gao, and Rui Yan. 2025. MockLLM: A Multi-Agent Behavior Collaboration Framework for Online Job Seeking and Recruiting. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining, V.2, KDD 2025, Toronto ON, Canada, August 3-7, 2025. 2714–2724

    work page 2025

  25. [25]

    Hao Sun, Jiayi Wu, Hengyi Cai, Xiaochi Wei, Yue Feng, Bo Wang, Shuaiqiang Wang, Yan Zhang, and Dawei Yin. 2024. AdaSwitch: Adaptive Switching between Small and Large Agents for Effective Cloud-Local Collaborative Learning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, EMNLP 2024, Miami, FL, USA, November 12-16,...

    work page 2024

  26. [26]

    Luyuan Wang, Yongyu Deng, Yiwei Zha, Guodong Mao, Qinmin Wang, Tianchen Min, Wei Chen, and Shoufa Chen. 2024. Mobileagentbench: An efficient and user-friendly benchmark for mobile llm agents.arXiv preprint arXiv:2406.08184 (2024)

    work page arXiv 2024

  27. [27]

    Quanmin Wei, Penglin Dai, Wei Li, Bingyi Liu, and Xiao Wu. 2025. CoPEFT: Fast Adaptation Framework for Multi-Agent Collaborative Perception with Parameter- Efficient Fine-Tuning. InAAAI-25, Sponsored by the Association for the Advance- ment of Artificial Intelligence, February 25 - March 4, 2025, Philadelphia, PA, USA. 23351–23359

    work page 2025

  28. [28]

    Yifan Xu, Xiao Liu, Xueqiao Sun, Siyi Cheng, Hao Yu, Hanyu Lai, Shudan Zhang, Dan Zhang, Jie Tang, and Yuxiao Dong. 2025. Androidlab: Training and systematic benchmarking of android autonomous agents. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2144–2166

    work page 2025

  29. [29]

    Biao Yi, Xavier Hu, Yurun Chen, Shengyu Zhang, Hongxia Yang, Fan Wu, and Fei Wu. 2025. EcoAgent: An Efficient Edge-Cloud Collaborative Multi-Agent Framework for Mobile Automation.arXiv preprint arXiv:2505.05440(2025)

    work page arXiv 2025

  30. [30]

    Junfei Zhan, Haoxun Shen, Zheng Lin, and Tengjiao He. 2025. PRISM: Privacy- Aware Routing for Adaptive Cloud-Edge LLM Inference via Semantic Sketch Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Trovato et al. Collaboration.arXiv preprint arXiv:2511.22788(2025)

    work page arXiv 2025

  31. [31]

    Chi Zhang, Zhao Yang, Jiaxuan Liu, Yanda Li, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. 2025. Appagent: Multimodal agents as smartphone users. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems. 1–20

    work page 2025

  32. [32]

    Guibin Zhang, Luyang Niu, Junfeng Fang, Kun Wang, Lei Bai, and Xiang Wang

    work page

  33. [33]

    InForty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025

    Multi-agent Architecture Search via Agentic Supernet. InForty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-19, 2025

    work page 2025

  34. [34]

    click the blue button

    Shaokun Zhang, Ming Yin, Jieyu Zhang, Jiale Liu, Zhiguang Han, Jingyang Zhang, Beibin Li, Chi Wang, Huazheng Wang, Yiran Chen, and Qingyun Wu. 2025. Which Agent Causes Task Failures and When? On Automated Failure Attribution of LLM Multi-Agent Systems. InForty-second International Conference on Machine Learning, ICML 2025, Vancouver, BC, Canada, July 13-1...

    work page 2025