Reproduce “MAEBE: Multi-Agent Emergent Behavior Framework” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sinem Erisken, Timothy Gothard, Martin Leitgab, Ram Potham
The MAEBE framework systematically evaluates emergent behaviors in multi-agent LLM systems, revealing that ensemble safety and alignment diverge unpredictably from individual agent performance due to group dynamics like peer pressure and preference amplification.
Key insights
- Moral preference responses of single LLMs are brittle to question framing, leading to significant shifts under dilemma inversion.
- Multi-agent system decisions are often not statistically predictable from individual agent behaviors, displaying emergent amplification effects.
- Group dynamics, particularly peer pressure and the indirect influence of supervisors, profoundly shape MAS collective reasoning and can lead to unexpected convergences or misalignments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.