Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

MAEBE: Multi-Agent Emergent Behavior Framework

Created by
  • Haebom

Author

Sinem Erisken (Independent Researcher), Timothy Gothard (Independent Researcher), Martin Leitgab (Independent Researcher), Ram Potham (Independent Researcher)

Outline

In this paper, we present a multi-agent emergent behavior evaluation (MAEBE) framework for assessing new risks that arise as AI systems composed of multiple agents become more widespread. Using MAEBE, the Greatest Good Benchmark, and a novel double-reversed questioning technique, we show that (1) LLM moral preferences, especially preferences for instrumental harms, vary significantly across questioning styles and are fragile both in single agents and in ensembles, (2) moral reasoning in LLM ensembles cannot be predicted from the behavior of single agents alone due to emergent group dynamics, and (3) convergence occurs in ensembles, particularly due to phenomena such as peer pressure, even with supervisory guidance. This highlights the need to evaluate AI systems in interacting multi-agent environments.

Takeaways, Limitations

•
Takeaways:
◦
Presenting the need for a MAEBE framework for assessing the emergence risk of multi-agent AI systems.
◦
We reveal that LLMs' moral preferences vary significantly and are vulnerable to questioning.
◦
We show that emergent behavior (e.g., peer pressure) in multi-agent systems raises safety and alignment issues.
◦
Emphasize that the behavior of a multi-agent system cannot be predicted based on the behavior of a single agent alone.
◦
Highlights the importance of evaluating AI systems in multi-agent environments.
•
Limitations:
◦
Further research is needed to determine the generality and applicability of the MAEBE framework.
◦
Need discussion on __T32295_____ of the Greatest Good Benchmark and double reversal questioning technique used.
◦
Generalizability verification is needed for various types of LLM and multi-agent systems.
◦
Lack of applicability and validation in real-world scenarios.
👍
Made with Slashpage