OVERVIEW

The field of multi-agent systems emerged in the early 1990s, driven by the vision that the future of AI would involve networks of AI programs - agents - operating autonomously in pursuit of user-delegated goals. A crucial part of this vision was that these agents would need to interact with one-another in pursuit of their delegated goals, which led to much work in the development of agents with social skills: the ability to cooperate, coordinate, and negotiate.

The emergence of Large-Language Models (LLMs) has led to massive renewed interest in multi-agent systems. Most LLMs are fundamentally chatbots: they aren't equipped with the ability to carry out actions, or to interact with other AI systems. There is much interest currently in the possibility of LLM-powered agents, which leverage the general intelligence of LLMs in systems that can act autonomously, exactly as was envisaged in the original dream of multi-agent systems. In addition, there is evidence that structuring a system as a collection of interacting agents, each playing a specific role, can be a useful way to build complex problem solving systems.

This workshop aims to bring together those interested in LLM-based multi-agent systems. The goals of the event are, firstly, to try to understand how the new technology of LLMs can be best used to realise the classic vision of multi-agent systems, and second, what can be learned from classic multi-agent systems in the LLM-agent era. 

SCHEDULE

8.00-8.30 Welcome, coffee

8.30-10.00 Session 1: Challenges and Opportunities

·      Christian Schroeder de Witt Open Challenges in Multi-Agent Security

·      Nicole Nichols, Sella Nevo and Mark Greaves

Achieving A Secure AI Agent Ecosystem: A Map of Open Opportunities and Actions for Advancement

·      Guohao Li and Guohao Li       

CAMEL: Finding the Scaling Law of Agents.

10.00-10.30 BREAK

10.30-12.30 Session 2: Platforms and Paradigms

·      Amit Chopra and Munindar P. Singh

Interaction-Oriented Programming as a Foundation of Agentic AI

·      Sizhe Yuen, Francisco Gomez Medina, Adam Sobey, Yali Du and Ting Su

Intrinsic Memory Agents: Heterogeneous Multi-Agent LLM Systems through Structured Contextual Memory

·      Matteo Baldoni, Cristina Baroglio, Amit Chopra, Elisa Marengo, Roberto Micalizio and Stefano Tedeschi

BSPL and Commitments for Modeling Interaction in Agentic AI

·      Sebastian Stein, Sarah Kiden, Vahid Yazdanpanah, Zhaoxing Li, Adrian Low, Andrew Poile, Behrad Koohy, Beining Zhang, Bruno Arcanjo, Connor Watson, Eike Schneiders, Eu Jin Lim, Ezhilarasi Periyathambi, Fariba Dehghan, Jan Buermann, Jayati Deshmukh, Jim Dilkes, Luke Nicholas, Prokopis Georgiou, Richard Gomer and Wen Gu 

Multiagent Systems Based on Large Language Models: A Citizen-Centric Perspective

12.30-1.00 Invited talk: Fetch.AI

Sana Wajid 

1.00-2.00 LUNCH

2.00-3.00 Session 3: Robotics

·      Zhaoxing Li, Wenbo Wu, Yue Wang, Yanran Xu and Sebastian Stein

HMCF: A Human-in-the-loop Multi-Robot Collaboration Framework Based on Large Language Models

·      Subramanian Ramamoorthy

Defining and developing AI for assistive autonomy

3.00-3.30 BREAK

3.30-5.30 Session 4: Ethical Issues

·      Naman Goel

On the Truthfulness of Surprisingly Likely Responses of Large Language Models

·      Jayati Deshmukh, Zijie Liang, Vahid Yazdanpanah, Sebastian Stein and Sarvapali D. Ramchurn

Serious Games and LLMs for Ethical Preference Elicitation

·      Steffen Backmann, David Guzman Piedrahita, Emanuel Tewolde, Rada Mihalcea, Bernhard Schölkopf and Zhijing Jin

When Ethics and Payoffs Diverge: LLM Agents in Morally Charged Social Dilemmas

·      Ariel Flint Ashery, Luca Maria Aiello and Andrea Baronchelli

Emergent social conventions and collective bias in LLM populations

5.30-6.30 Closing Reception (Drinks & Canapes)

TOPICS OF INTEREST

Include but are not restricted to...

 Revisiting classic multi-agent architectures in light of LLM capabilities 

Architectures for LLM agents

LLMs as agents vs. tools

Emergent behaviours in LLM-powered agent societies

Standards for LLM agents: MCP, A2A, ... 

 Communication and Coordination

   Language as the primary protocol - agents communicating via  natural language vs. symbolic messages

Grounding meaning and shared understanding among LLM agents

Ontologies for LLM-powered agents

 Negotiation, persuasion, and argumentation with LLM agents

Trust, deception, and misinformation in LLM-based agent societies

Emergence of shared mental models/collective intelligence

Dialogue protocols and structured conversations for LLM MAS

 Learning and Adaptation

On-the-fly role learning and specialization among LLM agents

Multi-agent reinforcement learning with language-based policies

Meta-learning in agent collectives using LLMs

 Evaluation and Benchmarking

Benchmarks for multi-agent communication involving LLMs

Token-efficient dialogues

Emergent cooperation / competition / collusion

Robustness and safety of LLM-powered agent collectives

 Ethics, Safety, and Governance

Detecting and mitigating unintended/emergent behaviours

Alignment of multi-agent objectives when agents have LLM-driven autonomy

Social biases and value alignment in agent interactions

Regulatory and societal implications of scalable LLM-based agent swarms

 Applications and Case Studies

Collaborative problem solving (e.g., scientific discovery, design)

Multi-agent simulations for economic or policy modeling

Large-scale digital societies and synthetic populations

Human-agent teams: hybrid collaboration settings

Multi-agent gaming and open-ended virtual worlds

 Game theory and LLM-based multi-agent systems

   Mechanism design for LLM-agent protocols

LLM agents as rational actors

Convergence to game theoretic solutions

 Future Directions

Bridging symbolic and sub-symbolic reasoning in agent collectives

Designing societal-level intelligence: beyond individual agent optimization

LLM-based robotic AI 

ACCEPTED PRESENTATIONS

HOW TO PARTICIPATE

We are seeking expressions of interest for participation by giving a talk or demo. Please submit your expression of interest via EasyChair:

    https://easychair.org/conferences?conf=rethinkmas2025

Note we do *NOT* require original unpublished research, or indeed the submission of technical papers. We are looking for expressions of interest to present research or demos, and the submission process is lightweight.

During the submission, you will be asked to upload a paper: you can upload a paper (published or unpublished), report, presentation, or any other document in the form of a PDF, to support your application. 

You should use the abstract to briefly describe what work you want to present and how it relates to the theme of the event. Indicate whether you wish to present or give a demo or both.

Expressions of interest will be only lightly reviewed, focussing on relevance/interest. Detailed reviews will not be provided. 

DEADLINE FOR EXPRESSIONS OF INTEREST:  FRIDAY 8 AUGUST 2025 AOE

NOTE: A token fee will be charged for attendance; consideration will be given for participants unable to pay the fee.

WHEN AND WHERE

The workshop will be a 1-day event held on 16 Sept 2025 at:

Department of Computer Science

University of Oxford

Oxford

OX1 3QD

United Kingdom

ORGANISATION

• Michael Wooldridge [University of Oxford] mjw@cs.ox.ac.uk *Main Contact*

• Sarit Kraus [Bar Ilan University, Israel]

• Emanuele la Malfa [University of Oxford]

• Samuele Marro [University of Oxford]

SUPPORT

This event is supported by the Schmidt Sciences Foundation under an AI2020 Senior Fellowship awarded to Michael Wooldridge, and by the UKRI under the AI Hub in Generative Models (https://www.genai.ac.uk).