pith. sign in

arxiv: 2606.31966 · v1 · pith:FYV3ICB5new · submitted 2026-06-30 · 💻 cs.MA · cs.AI· cs.CL· cs.CV

MECoBench: A Systematic Study of Multimodal Agent Collaboration in Embodied Environments

Qingyun Liu , Jiwen Zhang , Jingyi Hu , Siyuan Wang , Zhongyu Wei This is my paper

Pith reviewed 2026-07-01 02:14 UTC · model grok-4.3

classification 💻 cs.MA cs.AIcs.CLcs.CV
keywords multimodal agentsembodied environmentsagent collaborationcooperation benchmarkMLLMstask completioncommunication modesrobustness to noise
0
0 comments X

The pith

Multimodal agents complete embodied tasks more reliably through collaboration when communication balances the added coordination costs.

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

The paper presents MECoBench to examine how multimodal large language models act as embodied agents that must cooperate in visually grounded settings. Experiments across multiple models, real-world tasks, cooperation structures, and collaboration modes show that joint work raises task completion rates, yet only when the performance lift exceeds the overhead of coordinating actions. Communication turns out to be necessary for those lifts, while the most effective mode varies with team size and individual model strength. Collaboration also makes agents more stable when initial information contains noise or when exploration is constrained. The benchmark therefore supplies a controlled way to measure the practical limits of multi-agent embodied cooperation.

Core claim

MECoBench supplies a platform with diverse real-world tasks, two cooperation structures, and three collaboration modes. Systematic tests across MLLMs establish that collaboration generally raises embodied task completion rates provided collaborative gains exceed coordination complexity. Communication is required to realize those gains, and the optimal collaboration mode depends on team size and model capability. Collaboration further increases robustness when priors are noisy or exploration is limited.

What carries the argument

MECoBench, the benchmark platform that systematically varies tasks, cooperation structures, and collaboration modes to isolate when and how multimodal agents benefit from joint work.

If this is right

  • Communication protocols become a necessary design element for any multi-agent embodied system that hopes to exceed single-agent performance.
  • Choice of collaboration mode must be tuned to both the number of agents and the capability level of the underlying models.
  • Joint operation can be used to offset uncertainty in starting conditions or incomplete environmental knowledge.
  • Performance gains scale with the ability to exchange information without incurring prohibitive coordination overhead.
  • The same patterns hold across different MLLMs, suggesting the findings are not tied to one model family.

Where Pith is reading between the lines

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

  • Designers of future robotic teams could prioritize lightweight communication channels over raw model scale.
  • The observed dependence on team size points toward possible scaling rules for when adding agents stops helping.
  • The benchmark could be extended to test whether the same trade-offs appear in longer-horizon or more open-ended tasks.
  • Results imply that single-agent baselines may systematically underestimate what is achievable once coordination is solved.

Load-bearing premise

The tasks, structures, and modes selected for MECoBench are representative of the broader space of embodied multi-agent problems.

What would settle it

Re-running the full set of experiments on a fresh collection of embodied tasks outside the current benchmark and observing no consistent improvement from collaboration.

Figures

Figures reproduced from arXiv: 2606.31966 by Jingyi Hu, Jiwen Zhang, Qingyun Liu, Siyuan Wang, Zhongyu Wei.

Figure 1
Figure 1. Figure 1: An illustration of realistic scenarios, where cross-modal multi-agent collaboration significantly im￾proves the efficiency compared with single-agent. on pure language-based multi-agent frame￾work (Schmidgall et al., 2025; Hong et al., 2024) have proved that such cooperation could improve efficiency and overcome individual capability limits. However, the potential of multi-agent collaboration under multimo… view at source ↗
Figure 2
Figure 2. Figure 2: Data construction pipeline of MECoBench. Each task is first grounded from a high-level task into a concrete scene, then set collaboration configuration for parallel or sequential execution. relative gains under noisy priors, showing that multi-agent teams can compensate for mislead￾ing information through communication and dis￾tributed exploration. Available task information further amplifies the benefits … view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of four protocols under three collaboration modes. plate, and then randomly place the selected goal objects at legal initial locations. We define a broad range of surfaces and containers to enable diverse and realistic object distributions (listed in [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overall workflow of evaluation. Agents receive task goals and prior information, perceive the environ￾ment, communicate under different protocols, reason with history and memory, and execute actions with feedback. each agent shares one message with all others be￾fore action; in discussion, agents communicate se￾quentially and can continue for another round when consensus is not reached. The centralized mod… view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of model performance between parallel single-agent and sequential two-agent set￾tings. Dashed lines indicate quartile thresholds. gpt-5-mini gpt-5.4 gemini-3.1-pro qwen3-8b-vl qwen3-32b-vl qwen3-235b-vl qwen3.5-9b qwen3.5-27b gemma4-26b gemma4-31b internvl-3.5-38b internvl-3.5-241b llama-4 glm-4.6v glm-4.6v-flash 10 0 10 20 Percentage Points Closed source Open source SR gain (positive) SR gain (… view at source ↗
Figure 6
Figure 6. Figure 6: Performance change from 1-agent to 2- agent under parallel settings. Bars show the absolute change in SR and CR (percentage points) performance, and Gemma4 series performs partic￾ularly well on collaboration tasks. In contrast, In￾ternVL3.5 performs well individually but struggles to cooperate, revealing a gap between individual execution and group collaboration. Can models benefit from collaboration when … view at source ↗
Figure 7
Figure 7. Figure 7: Team size scaling effect. (a) Performance curve of different team size. (b) SR of Qwen3-32B-VL versus the #objects, with fitted trend lines for different team sizes. balance between collaboration benefits and costs. We further analyze the effect of task complexity using object count as a proxy [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of collaboration modes. (a) Performance profiles in 2-agent setup on parallel and sequential tasks. (b) SR and CR trends with increasing team size under decentralized broadcast and centralized leader-based. (a) (b) [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Comparison between broadcast and shared-memory. (a) Relative changes in effectiveness and efficiency over broadcast of Qwen3-VL. (b) Aver￾age completion progress over steps under broadcast and shared-memory of Qwen3-32B-VL. 5.3 Is Textual Communication Enough? While explicit textual communication is essential, it may be insufficient for embodied collaboration, where agents must share evolving task states … view at source ↗
Figure 11
Figure 11. Figure 11: Comparison in leader-based collaboration with and without visual augmentation. both textual reports and current visual observations to the leader. As shown in [PITH_FULL_IMAGE:figures/full_fig_p008_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: Team size scaling effect without prior lo￾cation information. (a) Performance curve of different team size. (b) SR of Qwen3-32B-VL versus the #ob￾jects, with fitted trend lines for different team sizes [PITH_FULL_IMAGE:figures/full_fig_p008_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Example of task goal content. The goal spec [PITH_FULL_IMAGE:figures/full_fig_p014_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Three-layer sunburst chart illustrating the [PITH_FULL_IMAGE:figures/full_fig_p015_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Data distribution grouped by task type and [PITH_FULL_IMAGE:figures/full_fig_p015_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Example observation image. For all evaluated models, we adopt the recom￾mended generation configurations whenever avail￾able. For closed-source models, we use the official APIs with their default generation settings. For open-source models, we follow the generation con￾figurations specified in the corresponding model cards or vLLM (Kwon et al., 2023) documentation to ensure strong and reliable performance… view at source ↗
Figure 18
Figure 18. Figure 18: Example of action decoding and grounding. [PITH_FULL_IMAGE:figures/full_fig_p017_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Broadcast protocol communication prompt. [PITH_FULL_IMAGE:figures/full_fig_p017_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Discussion protocol prompt. The first round [PITH_FULL_IMAGE:figures/full_fig_p018_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: Centralized mode: worker report prompt. You are an embodied agent in a 3D simulated environment. At this stage, you are acting as the leader for this step. Based on your current observation, memory, dialogue, and the workers' reports, assign exactly ONE next-step instruction to each agent, including yourself. You have access to a combined image that shows every agent's current observation views (your own … view at source ↗
Figure 24
Figure 24. Figure 24: Action and memory update prompt. Normal memory is highlighted in blue and shared memory is highlighted in orange. You are resolving object_ids for an action based on the class name and descriptions. You are given: - An action with object names and descriptions. - A list of visible candidate objects with ids and class names. - A list of holding objects with ids and class names. - Two images of the same vie… view at source ↗
Figure 27
Figure 27. Figure 27: Input block for the act phase. For sequential [PITH_FULL_IMAGE:figures/full_fig_p019_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: Memory update rules block. D Experiment Results In this section, we provide detailed results for the experiments in Sections 4, 5, and 6, as well as some additional experimental results. D.1 Performance of Different Models in Parallel and Sequential Tasks [PITH_FULL_IMAGE:figures/full_fig_p020_28.png] view at source ↗
Figure 29
Figure 29. Figure 29: Shared memory update rules block. ter within the same family. However, scaling tends to amplify existing task-specific patterns rather than change them. Some families remain balanced across parallel and sequential tasks, while others are consistently stronger in single-agent or par￾allel settings, or particularly weak on sequential tasks requiring precise coordination. For exam￾ple, Gemma4 benefits from s… view at source ↗
Figure 30
Figure 30. Figure 30: Task progress over steps for 1 to 5 agents [PITH_FULL_IMAGE:figures/full_fig_p022_30.png] view at source ↗
Figure 31
Figure 31. Figure 31: Comparison of explicit broadcast and implicit shared memory [PITH_FULL_IMAGE:figures/full_fig_p023_31.png] view at source ↗
Figure 33
Figure 33. Figure 33: Task progress over equivalent steps under [PITH_FULL_IMAGE:figures/full_fig_p023_33.png] view at source ↗
Figure 35
Figure 35. Figure 35: Completion rate under different prior￾information settings on parallel tasks. ure 34 shows the team size scaling effect under stronger exploration demands, where prior loca￾tion information is removed. Compared with the full-information setting, Qwen3-32B-VL exhibits a similar but weaker inverted-U-shaped trend, sug￾gesting that moderate team scaling remains helpful. For Qwen3-8B-VL, SR remains low across… view at source ↗
Figure 34
Figure 34. Figure 34: Team size scaling effect without location [PITH_FULL_IMAGE:figures/full_fig_p024_34.png] view at source ↗
Figure 36
Figure 36. Figure 36: Performance by number of objects across task conditions and model groups. [PITH_FULL_IMAGE:figures/full_fig_p025_36.png] view at source ↗
Figure 37
Figure 37. Figure 37: Duplicate grab rate of different models in [PITH_FULL_IMAGE:figures/full_fig_p027_37.png] view at source ↗
read the original abstract

Recent multimodal large language models (MLLMs) have strong potential as embodied agents, but their ability to collaborate in visually grounded environments remains underexplored. To address this gap, we introduce MECoBench, a multimodal embodied cooperation benchmark with an evaluation platform spanning diverse real-world tasks, two cooperation structures, and three collaboration modes. Through extensive experiments across various MLLMs, we summarize three key findings: (i) Collaboration generally improves embodied task completion, but its benefits depend on balancing collaborative gains against coordination complexity. (ii) Communication is essential to collaboration gains, while the best collaboration mode depends on team size and model capability. (iii) Moreover, collaboration improves robustness under noisy priors and exploration conditions. Generally, MECoBench provides a systematic testbed for understanding the mechanisms and limits of multimodal embodied collaboration. Code and dataset are available at https://github.com/q-i-n-g/MECoBench.

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

2 major / 2 minor

Summary. The manuscript introduces MECoBench, a multimodal embodied cooperation benchmark spanning diverse real-world tasks, two cooperation structures, and three collaboration modes. Experiments across multiple MLLMs yield three key findings: (i) collaboration generally improves task completion subject to coordination-complexity trade-offs, (ii) communication is essential while optimal mode depends on team size and model capability, and (iii) collaboration enhances robustness under noisy priors and exploration; the benchmark is positioned as a systematic testbed.

Significance. If the reported patterns prove robust to task selection and model families, MECoBench supplies a needed empirical platform for studying multimodal multi-agent collaboration in grounded settings, with potential to guide future agent design.

major comments (2)
  1. [Abstract] Abstract: the three key findings are phrased as general statements ("Collaboration generally improves", "Communication is essential", "the best collaboration mode depends on team size and model capability") without accompanying analysis or ablations that test invariance across task families, state-space characteristics, or MLLM architectures. This makes the claims load-bearing for the paper's central contribution yet unsupported by evidence of broader applicability.
  2. [Experiments / Results] The manuscript does not report any cross-task meta-analysis or sensitivity checks (e.g., correlation of performance deltas with task metrics such as state dimensionality or required coordination depth) that would substantiate the claimed mechanisms over benchmark-specific artifacts.
minor comments (2)
  1. [§3] Clarify the exact definition and implementation details of the three collaboration modes and two cooperation structures in the main text rather than relying solely on supplementary material.
  2. [§4] Ensure all reported metrics include error bars or statistical significance tests across random seeds and model runs.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback. The comments highlight important opportunities to strengthen the presentation of generality and mechanistic support in our work. We address each point below and commit to revisions that qualify claims appropriately while adding requested analyses where feasible.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the three key findings are phrased as general statements ("Collaboration generally improves", "Communication is essential", "the best collaboration mode depends on team size and model capability") without accompanying analysis or ablations that test invariance across task families, state-space characteristics, or MLLM architectures. This makes the claims load-bearing for the paper's central contribution yet unsupported by evidence of broader applicability.

    Authors: We agree the abstract phrasing is too broad. While the benchmark already includes diverse real-world tasks, two cooperation structures, three collaboration modes, and multiple MLLM families (as detailed in Sections 3 and 4), we did not perform explicit invariance ablations. In revision we will (i) qualify all three findings in the abstract with phrases such as "within the evaluated tasks and models" and (ii) add a dedicated subsection reporting performance deltas stratified by task family and model scale to make the scope explicit. revision: yes

  2. Referee: [Experiments / Results] The manuscript does not report any cross-task meta-analysis or sensitivity checks (e.g., correlation of performance deltas with task metrics such as state dimensionality or required coordination depth) that would substantiate the claimed mechanisms over benchmark-specific artifacts.

    Authors: The referee is correct that no such meta-analysis appears in the current manuscript. To address this, we will add a new analysis subsection that computes Spearman correlations between observed collaboration gains and task-level metrics (state dimensionality, coordination depth, and exploration requirement) across the benchmark tasks. This will be included in the revised Experiments section. revision: yes

Circularity Check

0 steps flagged

Empirical benchmark study with no derivations or fitted parameters

full rationale

The paper introduces MECoBench as an evaluation platform, runs experiments across MLLMs on its tasks/structures/modes, and reports observed patterns as findings. No equations, parameter fitting, predictions, or derivation chains exist that could reduce to inputs by construction. All claims are direct empirical summaries from the benchmark runs, rendering the work self-contained with no circularity.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The benchmark itself rests on domain assumptions about what constitutes representative embodied tasks and collaboration modes; no free parameters or invented physical entities are introduced.

axioms (1)
  • domain assumption The selected tasks and environments sufficiently capture real-world embodied cooperation challenges.
    Invoked when claiming the benchmark provides a systematic testbed for understanding mechanisms and limits.

pith-pipeline@v0.9.1-grok · 5705 in / 1038 out tokens · 30766 ms · 2026-07-01T02:14:46.555877+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

64 extracted references · 12 canonical work pages

  1. [1]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    1972

  2. [2]

    Publications Manual , year = "1983", publisher =

    1983

  3. [3]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    work page doi:10.1145/322234.322243 1981

  4. [4]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of

  5. [5]

    Dan Gusfield , title =. 1997

    1997

  6. [6]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    2015

  7. [7]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =

  8. [8]

    Tenenbaum and Sanja Fidler and Antonio Torralba , booktitle=

    Xavier Puig and Tianmin Shu and Shuang Li and Zilin Wang and Yuan-Hong Liao and Joshua B. Tenenbaum and Sanja Fidler and Antonio Torralba , booktitle=. Watch-And-Help: A Challenge for Social Perception and Human-. 2021 , url=

    2021

  9. [9]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Puig, Xavier and Ra, Kevin and Boben, Marko and Li, Jiaman and Wang, Tingwu and Fidler, Sanja and Torralba, Antonio , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , month =

  10. [10]

    M ulti A gent B ench : Evaluating the Collaboration and Competition of LLM agents

    Zhu, Kunlun and Du, Hongyi and Hong, Zhaochen and Yang, Xiaocheng and Guo, Shuyi and Wang, Zhe and Wang, Zhenhailong and Qian, Cheng and Tang, Xiangru and Ji, Heng and You, Jiaxuan. M ulti A gent B ench : Evaluating the Collaboration and Competition of LLM agents. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Vol...

    work page doi:10.18653/v1/2025.acl-long.421 2025

  11. [11]

    Collab-Overcooked: Benchmarking and Evaluating Large Language Models as Collaborative Agents

    Sun, Haochen and Zhang, Shuwen and Niu, Lujie and Ren, Lei and Xu, Hao and Fu, Hao and Zhao, Fangkun and Yuan, Caixia and Wang, Xiaojie. Collab-Overcooked: Benchmarking and Evaluating Large Language Models as Collaborative Agents. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10.18653/v1/2025.emnlp-main.249

    work page doi:10.18653/v1/2025.emnlp-main.249 2025

  12. [12]

    AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors , url =

    Chen, Weize and Su, Yusheng and Zuo, Jingwei and Yang, Cheng and Yuan, Chenfei and Chan, Chi-Min and Yu, Heyang and Lu, Yaxi and Hung, Yi-Hsin and Qian, Chen and Qin, Yujia and Cong, Xin and Xie, Ruobing and Liu, Zhiyuan and Sun, Maosong and Zhou, Jie , booktitle =. AgentVerse: Facilitating Multi-Agent Collaboration and Exploring Emergent Behaviors , url =

  13. [13]

    Multi-Agent Collaboration via Evolving Orchestration , url =

    Dang, Yufan and Qian, Chen and Luo, Xueheng and Fan, Jingru and Xie, Zihao and Shi, Ruijie and Chen, Weize and Yang, Cheng and Che, Xiaoyin and Tian, Ye and Xiong, Xuantang and Han, Lei and Liu, Zhiyuan and Sun, Maosong , booktitle =. Multi-Agent Collaboration via Evolving Orchestration , url =

  14. [14]

    Forty-second International Conference on Machine Learning , year=

    EmbodiedBench: Comprehensive Benchmarking Multi-modal Large Language Models for Vision-Driven Embodied Agents , author=. Forty-second International Conference on Machine Learning , year=

  15. [15]

    The Thirteenth International Conference on Learning Representations , year=

    VisualAgentBench: Towards Large Multimodal Models as Visual Foundation Agents , author=. The Thirteenth International Conference on Learning Representations , year=

  16. [16]

    2026 , month = mar, howpublished =

    GPT-5.4 Thinking System Card , author =. 2026 , month = mar, howpublished =

    2026

  17. [17]

    2026 , month = feb, howpublished =

    2026

  18. [18]

    2026 , month = apr, howpublished =

    2026

  19. [19]

    AI magazine , volume=

    Multiagent systems , author=. AI magazine , volume=

  20. [20]

    Autonomous Robots , year =

    Stone, Peter and Veloso, Manuela , title =. Autonomous Robots , year =. doi:10.1023/A:1008942012299 , url =

    work page doi:10.1023/a:1008942012299

  21. [21]

    Agent Laboratory: Using LLM Agents as Research Assistants

    Schmidgall, Samuel and Su, Yusheng and Wang, Ze and Sun, Ximeng and Wu, Jialian and Yu, Xiaodong and Liu, Jiang and Moor, Michael and Liu, Zicheng and Barsoum, Emad. Agent Laboratory: Using LLM Agents as Research Assistants. Findings of the Association for Computational Linguistics: EMNLP 2025. 2025. doi:10.18653/v1/2025.findings-emnlp.320

    work page doi:10.18653/v1/2025.findings-emnlp.320 2025

  22. [22]

    Many Heads Are Better Than One: Improved Scientific Idea Generation by A LLM -Based Multi-Agent System

    Su, Haoyang and Chen, Renqi and Tang, Shixiang and Yin, Zhenfei and Zheng, Xinzhe and Li, Jinzhe and Qi, Biqing and Wu, Qi and Li, Hui and Ouyang, Wanli and Torr, Philip and Zhou, Bowen and Dong, Nanqing. Many Heads Are Better Than One: Improved Scientific Idea Generation by A LLM -Based Multi-Agent System. Proceedings of the 63rd Annual Meeting of the As...

    work page doi:10.18653/v1/2025.acl-long.1368 2025

  23. [23]

    MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework , url =

    Hong, Sirui and Zhuge, Mingchen and Chen, Jonathan and Zheng, Xiawu and Cheng, Yuheng and Wang, Jinlin and Zhang, Ceyao and wang, zili and Yau, Steven and Lin, Zijuan and Zhou, Liyang and Ran, Chenyu and Xiao, Lingfeng and Wu, Chenglin and Schmidhuber, J\". MetaGPT: Meta Programming for A Multi-Agent Collaborative Framework , url =. International Conferen...

  24. [24]

    Ashraful and Ali, Mohammed Eunus and Parvez, Md Rizwan

    Islam, Md. Ashraful and Ali, Mohammed Eunus and Parvez, Md Rizwan. C ode S im: Multi-Agent Code Generation and Problem Solving through Simulation-Driven Planning and Debugging. Findings of the Association for Computational Linguistics: NAACL 2025. 2025. doi:10.18653/v1/2025.findings-naacl.285

    work page doi:10.18653/v1/2025.findings-naacl.285 2025

  25. [25]

    AI Hospital: Benchmarking Large Language Models in a Multi-agent Medical Interaction Simulator

    Fan, Zhihao and Wei, Lai and Tang, Jialong and Chen, Wei and Siyuan, Wang and Wei, Zhongyu and Huang, Fei. AI Hospital: Benchmarking Large Language Models in a Multi-agent Medical Interaction Simulator. Proceedings of the 31st International Conference on Computational Linguistics. 2025

    2025

  26. [26]

    npj Digital Medicine , year =

    Chen, Xi and Yi, Huahui and You, Mingke and Liu, WeiZhi and Wang, Li and Li, Hairui and Zhang, Xue and Guo, Yingman and Fan, Lei and Chen, Gang and Lao, Qicheng and Fu, Weili and Li, Kang and Li, Jian , title =. npj Digital Medicine , year =. doi:10.1038/s41746-025-01550-0 , url =

    work page doi:10.1038/s41746-025-01550-0

  27. [27]

    2025 , howpublished =

    AgentSociety: Large-Scale Simulation of LLM-Driven Generative Agents Advances Understanding of Human Behaviors and Society , author =. 2025 , howpublished =

    2025

  28. [28]

    2024 , eprint=

    MineLand: Simulating Large-Scale Multi-Agent Interactions with Limited Multimodal Senses and Physical Needs , author=. 2024 , eprint=

    2024

  29. [29]

    V illager A gent: A Graph-Based Multi-Agent Framework for Coordinating Complex Task Dependencies in M inecraft

    Dong, Yubo and Zhu, Xukun and Pan, Zhengzhe and Zhu, Linchao and Yang, Yi. V illager A gent: A Graph-Based Multi-Agent Framework for Coordinating Complex Task Dependencies in M inecraft. Findings of the Association for Computational Linguistics: ACL 2024. 2024. doi:10.18653/v1/2024.findings-acl.964

    work page doi:10.18653/v1/2024.findings-acl.964 2024

  30. [30]

    Scaling Large Language Model-based Multi-Agent Collaboration , url =

    Qian, Chen and Xie, Zihao and Wang, YiFei and Liu, Wei and Zhu, Kunlun and Xia, Hanchen and Dang, Yufan and Du, Zhuoyun and Chen, Weize and Yang, Cheng and Liu, Zhiyuan and Sun, Maosong , booktitle =. Scaling Large Language Model-based Multi-Agent Collaboration , url =

  31. [31]

    Embodied Agent Interface: Benchmarking LLMs for Embodied Decision Making , url =

    Li, Manling and Zhao, Shiyu and Wang, Qineng and Wang, Kangrui and Zhou, Yu and Srivastava, Sanjana and Gokmen, Cem and Lee, Tony and Li, Li and Zhang, Ruohan and Liu, Weiyu and Liang, Percy and Fei-Fei, Li and Mao, Jiayuan and Wu, Jiajun , booktitle =. Embodied Agent Interface: Benchmarking LLMs for Embodied Decision Making , url =. doi:10.52202/079017-3...

    work page doi:10.52202/079017-3188

  32. [32]

    ProcTHOR: Large-Scale Embodied AI Using Procedural Generation , url =

    Deitke, Matt and VanderBilt, Eli and Herrasti, Alvaro and Weihs, Luca and Ehsani, Kiana and Salvador, Jordi and Han, Winson and Kolve, Eric and Kembhavi, Aniruddha and Mottaghi, Roozbeh , booktitle =. ProcTHOR: Large-Scale Embodied AI Using Procedural Generation , url =

  33. [33]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =

    Savva, Manolis and Kadian, Abhishek and Maksymets, Oleksandr and Zhao, Yili and Wijmans, Erik and Jain, Bhavana and Straub, Julian and Liu, Jia and Koltun, Vladlen and Malik, Jitendra and Parikh, Devi and Batra, Dhruv , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =

  34. [34]

    PARTNR: A Benchmark for Planning and Reasoning in Embodied Multi-agent Tasks , url =

    Chang, Matthew and Chhablani, Gunjan and Clegg, Alexander and Dallaire Cote, Mikael and Desai, Ruta and Hlavac, Michal and Karashchuk, Vladimir and Krantz, Jacob and Mottaghi, Roozbeh and Parashar, Priyam and Patki, Siddharth and Prasad, Ishita and Puig, Xavier and Rai, Akshara and Ramrakhya, Ram and Tran, Daniel and Truong, Joanne and Turner, John and Un...

  35. [35]

    Robotics: Science and Systems , year=

    NaVid: Video-based VLM Plans the Next Step for Vision-and-Language Navigation , author=. Robotics: Science and Systems , year=

  36. [36]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Yang, Yijun and Zhou, Tianyi and Li, Kanxue and Tao, Dapeng and Li, Lusong and Shen, Li and He, Xiaodong and Jiang, Jing and Shi, Yuhui , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2024 , pages =

    2024

  37. [37]

    Nature Machine Intelligence , volume =

    Embodied large language models enable robots to complete complex tasks in unpredictable environments , author =. Nature Machine Intelligence , volume =. 2025 , doi =

    2025

  38. [38]

    2025 , eprint=

    EmbodiedEval: Evaluate Multimodal LLMs as Embodied Agents , author=. 2025 , eprint=

    2025

  39. [39]

    The Twelfth International Conference on Learning Representations , year=

    Large Language Models as Generalizable Policies for Embodied Tasks , author=. The Twelfth International Conference on Learning Representations , year=

  40. [40]

    COMBO: Compositional World Models for Embodied Multi-Agent Cooperation , url =

    Zhang, Hongxin and Wang, Zeyuan and Lyu, Qiushi and Zhang, Zheyuan and Chen, Sunli and Shu, Tianmin and Dariush, Behzad and Lee, Kwonjoon and Du, Yilun and Gan, Chuang , booktitle =. COMBO: Compositional World Models for Embodied Multi-Agent Cooperation , url =

  41. [41]

    Heterogeneous Embodied Multi-Agent Collaboration , year=

    Liu, Xinzhu and Guo, Di and Zhang, Xinyu and Liu, Huaping , journal=. Heterogeneous Embodied Multi-Agent Collaboration , year=

  42. [42]

    Building Cooperative Embodied Agents Modularly with Large Language Models , url =

    Zhang, Hongxin and Du, Weihua and Shan, Jiaming and Zhou, Qinhong and Du, Yilun and Tenenbaum, Joshua B and Shu, Tianmin and Gan, Chuang , booktitle =. Building Cooperative Embodied Agents Modularly with Large Language Models , url =

  43. [43]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Shridhar, Mohit and Thomason, Jesse and Gordon, Daniel and Bisk, Yonatan and Han, Winson and Mottaghi, Roozbeh and Zettlemoyer, Luke and Fox, Dieter , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

  44. [44]

    Science China Information Sciences , year =

    Feng, Zhaohan and Xue, Ruiqi and Yuan, Lei and Yu, Yang and Ding, Ning and Liu, Meiqin and Gao, Bingzhao and Sun, Jian and Zheng, Xinhu and Wang, Gang , title =. Science China Information Sciences , year =. doi:10.1007/s11432-025-4820-4 , url =

    work page doi:10.1007/s11432-025-4820-4

  45. [45]

    2025 , eprint=

    Qwen3-VL Technical Report , author=. 2025 , eprint=

    2025

  46. [46]

    2026 , howpublished =

    Gemma 4 Model Card , author =. 2026 , howpublished =

    2026

  47. [47]

    2025 , eprint=

    InternVL3.5: Advancing Open-Source Multimodal Models in Versatility, Reasoning, and Efficiency , author=. 2025 , eprint=

    2025

  48. [48]

    2025 , howpublished =

    Llama 4: Model Cards and Prompt Formats , author =. 2025 , howpublished =

    2025

  49. [49]

    2025 , eprint=

    GLM-4.5V and GLM-4.1V-Thinking: Towards Versatile Multimodal Reasoning with Scalable Reinforcement Learning , author=. 2025 , eprint=

    2025

  50. [50]

    2025 , howpublished =

    GPT-5 mini Model , author =. 2025 , howpublished =

    2025

  51. [51]

    EmbodiedGPT: Vision-Language Pre-Training via Embodied Chain of Thought , url =

    Mu, Yao and Zhang, Qinglong and Hu, Mengkang and Wang, Wenhai and Ding, Mingyu and Jin, Jun and Wang, Bin and Dai, Jifeng and Qiao, Yu and Luo, Ping , booktitle =. EmbodiedGPT: Vision-Language Pre-Training via Embodied Chain of Thought , url =

  52. [52]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Szot, Andrew and Mazoure, Bogdan and Attia, Omar and Timofeev, Aleksei and Agrawal, Harsh and Hjelm, Devon and Gan, Zhe and Kira, Zsolt and Toshev, Alexander , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2025 , pages =

    2025

  53. [53]

    Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=

    Efficient Memory Management for Large Language Model Serving with PagedAttention , author=. Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles , year=

  54. [54]

    A Cordial Sync: Going Beyond Marginal Policies for Multi-agent Embodied Tasks

    Jain, Unnat and Weihs, Luca and Kolve, Eric and Farhadi, Ali and Lazebnik, Svetlana and Kembhavi, Aniruddha and Schwing, Alexander. A Cordial Sync: Going Beyond Marginal Policies for Multi-agent Embodied Tasks. Computer Vision -- ECCV 2020. 2020

    2020

  55. [55]

    RoCo: Dialectic Multi-Robot Collaboration with Large Language Models , year=

    Mandi, Zhao and Jain, Shreeya and Song, Shuran , booktitle=. RoCo: Dialectic Multi-Robot Collaboration with Large Language Models , year=

  56. [56]

    2025 , eprint=

    Collaborating Action by Action: A Multi-agent LLM Framework for Embodied Reasoning , author=. 2025 , eprint=

    2025

  57. [57]

    2024 , eprint=

    TeamCraft: A Benchmark for Multi-Modal Multi-Agent Systems in Minecraft , author=. 2024 , eprint=

    2024

  58. [58]

    2026 , eprint=

    COOP ^2 : Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems , author=. 2026 , eprint=

    2026

  59. [59]

    VIKI‑R: Coordinating Embodied Multi-Agent Cooperation via Reinforcement Learning , url =

    Kang, Li and Song, Xiufeng and Zhou, Heng and Qin, Yiran and Yang, Jie and Liu, Xiaohong and Torr, Philip and BAI, LEI and Yin, Zhenfei , booktitle =. VIKI‑R: Coordinating Embodied Multi-Agent Cooperation via Reinforcement Learning , url =

  60. [60]

    On the Utility of Learning about Humans for Human-AI Coordination , url =

    Carroll, Micah and Shah, Rohin and Ho, Mark and Griffiths, Tom and Seshia, Sanjit and Abbeel, Pieter and Dragan, Anca , booktitle =. On the Utility of Learning about Humans for Human-AI Coordination , url =

  61. [61]

    LLM -Coordination: Evaluating and Analyzing Multi-agent Coordination Abilities in Large Language Models

    Agashe, Saaket and Fan, Yue and Reyna, Anthony and Wang, Xin Eric. LLM -Coordination: Evaluating and Analyzing Multi-agent Coordination Abilities in Large Language Models. Findings of the Association for Computational Linguistics: NAACL 2025. 2025. doi:10.18653/v1/2025.findings-naacl.448

    work page doi:10.18653/v1/2025.findings-naacl.448 2025

  62. [62]

    SoMi-ToM: Evaluating Multi-Perspective Theory of Mind in Embodied Social Interactions , url =

    Fan, Xianzhe and Zhou, Xuhui and Jin, Chuanyang and Nottingham, Kolby and Zhu, Hao and Sap, Maarten , booktitle =. SoMi-ToM: Evaluating Multi-Perspective Theory of Mind in Embodied Social Interactions , url =

  63. [63]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    AirCopBench: A Benchmark for Multi-drone Collaborative Embodied Perception and Reasoning , author =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2026 , doi =

    2026

  64. [64]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =

    Zhang, Shiduo and Xu, Zhe and Liu, Peiju and Yu, Xiaopeng and Li, Yuan and Gao, Qinghui and Fei, Zhaoye and Yin, Zhangyue and Wu, Zuxuan and Jiang, Yu-Gang and Qiu, Xipeng , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =. 2025 , pages =

    2025