The Stubborn Takeover

A single agent can hijack a multi-agent system. Not by being smarter, not by outnumbering the others, but by being stubborn.

Abedini et al. model LLM-based multi-agent systems using the Friedkin-Johnsen opinion formation framework, where each agent updates its state as a weighted average of its neighbors’ messages and its own prior conviction. Stubbornness is a scalar: how much weight an agent gives to its own initial position versus incoming messages. A stubborn agent barely updates; a flexible agent tracks its neighbors.

The result: a single highly stubborn and persuasive agent can take over the collective dynamics of the entire system. The mechanism is asymmetric averaging. When one agent barely moves while all others continually adjust, the immovable agent becomes the attractor. Over iterations, every flexible agent drifts toward the stubborn one — not because it is right, but because it is the only fixed point in the system. Flexibility, the property that makes agents responsive and cooperative, is exactly what makes them vulnerable. The system’s strength is its attack surface.

The obvious defense — make all agents more stubborn — fails. Increasing general stubbornness reduces cooperative functionality. The system becomes rigid, unable to aggregate information, defeating the purpose of multi-agent collaboration.

The authors propose trust adaptation: dynamically reducing the weight assigned to neighbors whose opinions are unusually persistent. Rather than making everyone stubborn, make everyone suspicious of stubbornness. The defense targets the mechanism — fixed-point attraction via asymmetric updating — rather than the symptom.

The structural insight generalizes beyond AI: any system that reaches consensus through iterated averaging is vulnerable to participants who refuse to update. Stubbornness is not just a personality trait. It is a topological property — a fixed point in opinion space — and fixed points attract trajectories.