How Weak AI Agents Take Over a Network: The Crystallization of Inequality ✛ ✎

A tiny speck of dust in a saturated solution turns into a crystal with sharp faces in seconds. Neighboring particles go unnoticed. A similar process recently appeared in a group of Large Language Models (autonomous computer programs based on language models).

The experiment was simple: 100 agents with different roles — ‘Explainer’, ‘Verifier’ — freely chose whom to talk to in order to solve a task. Each made a decision making move by evaluating a partner with a attention mechanism — as if you were deciding how important a conversation partner is and giving them points. Gradually, random encounters (here, stochastic processes) arranged themselves into a Graphs network, where more visible agents attracted more and more links. This is collective behavior known as preferential attachment — the effect where the popular get even more popular.

Why does this happen? The researchers described the average network dynamics with differential equations and found that the system inevitably reaches an equilibrium, but that equilibrium is often unfair. Just as a crystal grown in a saturated solution won't change its shape, the agent network hardens into a particular configuration. Changing it later is hard. So, when building collective Machine Learning, it’s not enough to check if it answers correctly. We must watch who talks to whom, and whether a smart agent is being pushed to the sidelines.

Scientists can now predict such imbalances just by knowing how pairs of agents communicate with each other. It’s like chemists learning to add impurities to a solution to alter a crystal’s shape. Perhaps soon we’ll have ‘crystallization regulators’ for AI networks — they’ll weaken unwanted inequality without disrupting the overall work. That way, the brightest facets of the collective mind won’t be dimmed by random circumstances.