Reproduce “Deception and Communication in Autonomous Multi-Agent Systems: An Experimental Study with Among Us” 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
Maria Milkowski, Tim Weninger
Researchers at the University of Notre Dame conducted an experimental study using a text-based Among Us simulation to investigate communication and deception in autonomous LLM agents. They found agents predominantly communicate through directives and, when deceptive, overwhelmingly favor equivocation (91.2% of deceptive acts) over outright falsification, with this ambiguity increasing under social pressure but not reliably determining strategic success.
Key insights
- LLM agents prioritize task-oriented communication, with approximately 98% of utterances being directives, regardless of their role.
- Deception in LLM agents is overwhelmingly characterized by equivocation (91.2% of deceptive acts), suggesting a preference for ambiguity and plausible deniability over explicit lies, possibly influenced by safety training.
- Deceptive communication, especially equivocation, increases with social pressure (e.g., number of ejections) but does not reliably predict strategic success for the deceptive agents.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.