Project summary

I have a paper accepted to the 25th International Conference on Autonomous Agents and Multi-Agent Systems (AAMAS 2026), titled “Beyond Vibe Decision Theory: Asymmetric Manipulation Vulnerabilities in LLM Multi-Agent Coordination.” This paper is extending upon work we previously published at NeurIPS 2025 Workshop on Algorithmic Collective Action: https://openreview.net/forum?id=PXzSyb2Zlv

This work investigates how large language models (LLMs) behave in multi-agent environments, particularly how contextual framing and strategic advice can systematically manipulate coordination outcomes. Through experiments across canonical game-theoretic settings (e.g., Prisoner’s Dilemma, Public Goods), we demonstrate that LLMs exhibit significant and asymmetric vulnerabilities to manipulation, raising important concerns for alignment and robustness in real-world deployments.

This contribution sits at the intersection of multi-agent systems, AI alignment, and behavioral game theory, and is highly relevant to the AAMAS community, where understanding coordination, robustness, and agent interaction is central.

What are this project's goals? How will you achieve them?

The primary goal is to advance the understanding of robust multi-agent coordination in LLM-based systems.

Specifically we:

• Identifies systematic vulnerabilities in multi-agent LLM coordination

• Demonstrates how contextual framing and conflicting instructions affect strategic behavior

• Provides empirical evidence across multiple models and game environments, highlighting architecture-dependent robustness issues

By presenting at AAMAS, I aim to:

• Share these findings with researchers working on multi-agent systems, alignment, and AI safety

• Receive feedback to extend this work toward mechanism design, robustness guarantees, and safer multi-agent deployments

• Explore collaborations on alignment in decentralized or agent-based AI systems

As a researcher at a early career transitional stage, attending AAMAS would really allow me to engage with fellow researchers in the field and position my future research agenda within multi-agent AI and alignment.

This contribution sits at the intersection of multi-agent systems, AI alignment, and behavioral game theory, and is highly relevant to the AAMAS community, where understanding coordination, robustness, and agent interaction is central. I have worked and am working on similar projects related to epsitemic risk in language models.

How will this funding be used?

• Conference registration (student, main conference only): ~$463.24

• Round-trip travel (US --> Cyprus): ~$1,500

• Accommodation: ~$500

The funding will be used to cover the conference fee and travel expenses. Any support would be appreciated, the remaining amount I will work to cover.

Who is on your team? What's your track record on similar projects?

This research project was supported by Pivotal Research who have a strong track record of supporting researchers in tackling important questions about the safety and governance of emerging technology.

What are the most likely causes and outcomes if this project fails?

The primary risk is lack of funding to attend the conference.

If I am unable to attend there is the risk that the paper may be taken out of the proceedings. Also, if it is not presented in person, it could reduce its visibility and impact but more importantly we miss on opportunities for feedback and collaboration

Given that AAMAS is a leading venue for multi-agent systems, absence would represent a missed opportunity for both scientific exchange and career development

How much money have you raised in the last 12 months, and from where?

I did not raise any money, but part of expenses for travel and accommodation I can cover from my salary. Unfortunately, the conference has rejected our application for volunteering and conference fee support and I haven't received funding through my university.