Motivation. The scaling of model parameters has unlocked the groundbreaking capabilities of foundation models. Likewise, in human society, scaling and collaboration across individuals, organizations, companies, and nations amplify collective intelligence to unprecedented levels, enabling remarkable achievements that would be impossible for individuals alone, such as space exploration and autonomy. Could this principle of scaling also apply to the growth in the number of agents? Multi-agent systems may offer a promising path forward. By progressively integrating more agents, multi-agent systems can activate diverse functionalities within these foundation model-powered generalist agents and coordinate a broader range of complementary functionalities. This synergy fosters improved problem-solving, adaptability, and decision-making capabilities. As the multi-agent system scales, it has a huge potential to achieve enhanced capabilities and tackle increasingly complex tasks, offering a promising solution toward the ultimate goal of achieving artificial general intelligence (AGI).
Background. With the advent of large foundation models, including large language models (LLMs) and visual language models (VLMs), research in multi-agent systems has progressed from relying on specialized models to harnessing the versatile capabilities of more generalized LLM/VLM-powered agents. As a community, we are embracing this paradigm shift to develop general-purpose multi-agent systems. These systems aim to leverage collaboration to enhance individual agent capabilities, ultimately advancing toward sophisticated AI agents that can serve as versatile human assistants. This field has already seen numerous promising developments. For instance, multi-agent systems for social simulation are used to generate diverse synthetic data, further enhancing the capabilities of foundation models and also investigating social behaviors. Moreover, multi-agent systems emulate human standards of processing through agentic workflows, significantly advancing a wide range of complex and urgent real-world applications, including mathematics, software development, web and mobile operations, embodied manipulation, and navigation, among others. These developments underscore the vast potential of multi-agent systems powered by foundation models to tackle increasingly complex challenges across diverse domains.
Our Workshop. This workshop seeks to bring together researchers from diverse disciplines to explore emerging topics in multi-agent systems powered by foundation models (LLMs, VLMs, and MMLMs). By complementing existing workshops focused on individual intelligence, it introduces a new perspective centered on advancing agent capabilities through multi-agent collaboration and their integration into human life. The primary objective of this workshop is to foster meaningful discussions and collaborations that address critical scientific questions surrounding the development and application of such systems.
Call for papers
This workshop aims to deepen understanding and offer innovative perspectives on the growth and collaboration of LLM/VLM-powered agents in multi-agent contexts. To foster an inclusive environment for discussion and debate, we welcome speakers and panelists from diverse backgrounds and expertise. Our lineup features distinguished researchers alongside emerging investigators who have made significant contributions to the field. Spotlight and poster sessions will highlight new ideas, key challenges, and retrospective insights related to the workshop’s themes. We strive for our participant selection to reflect the dynamic and diverse landscape of machine learning and AI. We will explore a range of topics in this workshop, including, but not limited to, the following areas:
Simulations in society, games, psychology, economics, and politics.
The creation of datasets and benchmarks to inspire innovation and provide more effective evaluation of the enhanced capabilities of multi-agent systems.
Multi-agent workflows, collaboration graphs, and communication protocols.
Exploring interactions and synergies between humans and agents to foster the development of more trustworthy and socially friendly agents.
Math, software development, open question answering, web agents, os agents, mobile agents, embodied agents for navigation, exploration, manipulation, autonomous systems, robotics, medical agents.
Advance the multi-agent system from the interactive environment feedback.
Theoretical framework for multi-agent systems.
Submission Guide
Submission Platform: Submit your papers here: OpenReview Submission Site
Submission Requirements:
Use the official LaTeX template of ICML 2025 (Style
Files)
Papers must be prepared and submitted as a single PDF: 8 pages for the main paper, with
unlimited pages for references and appendices (reviewers are not obliged to read the appendices)
All submissions must be anonymized, which will be reviewed in a
double-blind manner
Non-Archival Policy: Submissions will not be indexed or have archival proceedings. We welcome NeurIPS 2025 submissions.
Best Paper Award: Best paper award will be announced at the workshop.
Key Dates
📄 Paper Submission Open: Apr 20, 2025 🐰Happy Easter!
📄 Paper Submission Deadline: May 26, 2025
📄 Acceptance Notification: June 9, 2025
📄 Camera-Ready Deadline: June 30, 2025
📄 Workshop Date: July 18, 2025
Workshop Schedule
| Time | Session | Duration | Details |
|---|---|---|---|
| 9:00AM - 9:10AM | Opening Remarks | 10 min | Welcome and Introduction to the Workshop |
| 9:10AM - 9:30AM | Invited Talk 1 | 20 min | Talk1 - Yejin Choi |
| 9:40AM - 10:00AM | Invited Talk 2 | 20 min | Talk2 - Natasha Jaques |
| 10:10AM - 10:40AM | Oral Presentations 1 | 30 min | 15 min * 2 |
| 10:45AM - 11:45AM | Poster Session & Coffee Socials 1 | 60 min | Poster Session & Coffee Socials1 |
| 11:45AM - 1:00PM | Lunch Break | 75 min | Time for lunch and informal discussions |
| 1:00PM - 1:20PM | Invited Talk 3 | 20 min | Talk3 - Yilun Du |
| 1:30PM - 1:40PM | Invited Talk 4 | 10 min | Talk4 - Diyi Yang |
| 1:50PM - 2:20PM | Oral Presentations 2 | 30 min | Oral Presentations2 |
| 2:25PM - 3:25PM | Poster Session & Coffee Socials 2 | 60 min | Networking and refreshments |
| 3:30PM - 3:50PM | Invited Talk 5 | 20 min | Talk5 - Mengdi Wang |
| 4:00PM - 4:20PM | Invited Talk 6 | 20 min | Talk6 - Yulia Tsvetkov |
| 4:30PM - 5:00PM | Panel Discussion | 30 min | Interactive session with experts - Joon Sung Park, Mingchen Zhuge, Chen Qian |
| 5:00PM - 5:15PM | Awards and Conclusive Remarks | 15 min | Concluding the workshop and award announcements |
Challenge
Multi-Agent Embodied Intelligence Challenge @ ICML 2025
Advancing Embodied AI beyond single-agent settings, the Multi-Agent Embodied Intelligence Challenge, hosted at ICML 2025, invites the community to explore intelligent coordination in physically grounded environments. This challenge focuses on multi-agent manipulation via imitation learning, underpinned by compositional reasoning and scalable learning architectures.
Hosted by the MARS-EAI initiative, which focuses on multi-agent embodied AI, this competition is built upon RoboFactory -- a novel simulation platform and benchmark tailored for collaborative robotic manipulation. RoboFactory introduces a constraint-driven framework where agent behavior is guided by logical, spatial, and temporal constraints. These compositional constraints are critical for enabling safe, efficient, and interpretable collaboration among embodied agents in shared spaces.
Participants will train policies in a suite of challenging scenarios, ranging from dual-arm stacking to long-horizon cooperative tasks involving up to four agents. The core challenge lies in learning coordination policies that can handle uncertainty, leverage egocentric and global observations, and scale with the number of interacting agents.
Key highlights:
- Compositional Imitation Learning: Participants must generate structured, constraint-aware behavior from demonstrations.
- Benchmark Integration: Built atop ManiSkill3 and aligned with real-world application, RoboFactory enables rich, diverse task settings.
- Evaluation Metrics: Performance will be assessed on task success, data efficiency, constraint compliance, and generalization under scene randomization.
We invite researchers from robotics, imitation learning, multi-agent RL, and foundation model communities to participate. Top teams will be recognized at ICML 2025 and contribute to a shared vision of scalable, collaborative embodied intelligence.
Speakers and panelists
Yejin Choi
Senior Director at NVIDIA, Professor at Stanford
Natasha Jaques
Assistant Professor, University of Washington, Senior Research Scientist at Google DeepMind
Yilun Du
Senior Research Scientist at Google Deepmind, Assistant Professor at Harvard
Diyi Yang
Assistant Professor at StanfordWang.png)
Mengdi Wang
Associate Professor at Princeton University
Yulia Tsvetkov
Associate Professor at University of Washington
Joon Sung Park
PhD Student at Stanford
Mingchen Zhuge
PhD Candidate at KAUST
Chen Qian
Associate Professor at Shanghai Jiao Tong UniversityWorkshop Organizers
Zhenfei Yin
PhD at USYD, Visiting Researcher at Oxford
Yue Hu
Postdoctoral Fellow at University of Michigan
Siheng Chen
Associate Professor at Shanghai Jiao Tong University
Bernadette Bucher
Assistant Professor at University of Michigan
Rui Ye
PhD Candidate at Shanghai Jiao Tong University
Chahyon Ku
PhD Student at University of Michigan
Katrina Ashton
PhD Candidate at University of Pennsylvania
Juan Carlos Niebles
Research Director at Salesforce AI Research, Adjunct Professor at Stanford
Lei Bai
Research Scientist at Shanghai AI Laboratory
Dawn Song
Professor at University of California, Berkeley
Deep Wisdom AI