Consensus Within, Competition Between: Decentralized LLM Teams for Perimeter Defense
Summary
- A study evaluated centralized, consensus-based, and independent LLM team architectures for multi-agent perimeter defense under five network conditions.
- The auction-based consensus team achieved the lowest breach rate at 15%, compared to 32% for the central commander and 29% for independent agents.
- Giving low-cost LLMs more authority over bid computation increased breaches significantly, while visual deception via taxidermy bird disguises fooled models until prompt forewarning was applied.
Why it matters
Understanding how LLM team architectures and model capabilities handle communication constraints and adversarial visual deception is critical for deploying reliable autonomous multi-agent defense systems.