Crystallization of Inequality in AI Agent Networks ✛ ✎
Context
A tiny speck of dust in a supersaturated solution grows into a large crystal within seconds, leaving its neighbors unnoticed. Networks of Large Language Models behave similarly when given the freedom to choose whom to communicate with. As the population grows, collective behavior does not smooth out but solidifies. Understanding the nature of such stochastic processes means learning to steer collective intelligence so that capable agents aren't sidelined just because of an unlucky twist of fate.
Methods
Built a Graphs of 100 agents—each with its own model, size, and role. Agents solved two types of tasks: collaborative answer finding and debates. At each step, a new node was randomly added or connections were made between existing ones. The agent itself decided on the connection strength by evaluating a partner through an attention mechanism: the target's query was matched against the source's key. This yielded a three-dimensional communication vector: token budget, number of rounds, and depth of reasoning. Then the averaged dynamics were described by an ordinary differential equations and convergence to equilibrium was proved. In parallel, numerical simulation with 50 independent runs confirmed the theoretical predictions.
Results
The equilibrium proved stable—the share of influence of each agent type stopped changing. But the equilibrium itself was often asymmetric. In some cases, for example among models of the Gemini family, the stronger version naturally attracted up to 60% of the communication mass. This is collective behavior, aligned with capabilities. In others, roles assigned by the system prompt outweighed computational power: an 'Explainer' agent on a weak model dominated a strong 'Checker'. A reverse decision making emerged: the strong were left in the shadows. Experiments showed that in such 'inverted' networks, truthful information fades out, while hallucinations, on the contrary, are spread by central nodes.
Implications
The finding changes the perspective on multi-component AI systems. It's not enough to check answer accuracy—one must analyze the Graphs structure, visibility, and influence of agents. A forecasting tool emerges: knowing the statistics of pairwise interactions, one can compute what equilibrium the community will reach and whether a capable helper will be left out in the cold.
Future development
The theory can be extended to three or more types of agents, incorporating mechanisms for targeted network optimization. Promising is the search for universal 'crystallization regulators'—parameters that weaken unfair collective behavior without destroying beneficial cooperation.
Impact
Affected are the development of Machine Learning multi-agent AI systems, artificial intelligence safety, and computational sociology. The results will help design stable collectives of software agents for collaborative work.
Next steps
Planned are experiments with heterogeneous populations of dozens of models and prolonged simulations that test predictions beyond laboratory conditions.
Key open problems
The work contributes to understanding emergent behavior in complex Large Language Models AI systems and raises the question of controllability of spontaneously arising hierarchies—one of the key unsolved problems of safe artificial intelligence.
The knowledge machine's view
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors built a network of 100 LLM agents that choose their partners via an attention mechanism and studied the dynamics of communication inequality. Using differential equations and simulations, they showed that the equilibrium is often asymmetric: weak agents can occupy central positions, which affects the spread of information and hallucinations.
Where its strength is
A strong point is the combination of theoretical analysis (contractive mappings, convergence to equilibrium) with numerical experiments, providing a tool for predicting network structure in collectives of AI agents.
Where it could go next
- It might be interesting to compare the observed crystallization of inequality with universal mechanisms that make different systems behave alike: perhaps the same self-organization principles operate in agent networks as in other complex systems.based on: Why Different Systems Behave the Same arXiv:2607.09079
- Related work on how charged particles choose a crystal also exhibits preferential attachment and structure formation. Transferring methods from condensed matter physics might help describe phase transitions in agent networks more precisely.based on: How Charged Particles Choose Their Crystal arXiv:2608.09714v1
- There is an arXiv paper where network links form triads: hypergraph structures can capture collective effects invisible in pairwise interactions. This could extend the model if agents form coalitions or influence each other indirectly.based on: A Network Where Connections Befriend in Trios arXiv:2605.17004
Worth trying
- Test whether the glass ceiling effect persists when hypergraph connections are introduced, allowing agents to form groups of three or more participants.
- Use methods from the work on crystallization of charged particles to analyze threshold phenomena: at what network parameters does a sharp transition to dominance by weak agents occur?
- Compare the dynamics of agent networks with universal self-organization patterns from the work on behavioral similarity across different systems to identify common regularities.
Why this may matter
Understanding how inequality of influence emerges in AI agent collectives is important for designing reliable multi-agent systems: if weak agents can dominate, this affects decision quality and information spread, which is critical for practical applications.
Nearby works in our archive
- When the Mind Honestly Says 'I Don't Know'similarity 0.59
- Why Different Systems Behave the Samesimilarity 0.58
- A Network Where Connections Befriend in Triossimilarity 0.58
- Digital Twin: How People's Invisible Behavior Changes Planssimilarity 0.57
- How to Design the Perfect Therapeutic Peptide: Three Properties at Oncesimilarity 0.57
- Cellular Atlases: Who Reads the Language of Tissue Bestsimilarity 0.57
- How Charged Particles Choose Their Crystalsimilarity 0.56
The archive is sparse around this work: the closest review is only 0.59, against a usual 0.64. It looks like few have drilled nearby.
In debates, the Gemini-2.5-Flash-Lite model (weak) completely displaced the more powerful Gemini-3.5-Flash from the network center: its influence share dropped almost to zero, even though under equal conditions the strong model should have dominated.
Key numbers
- number of agents in the experiment: 100
- maximum number of macro-steps of evolution: 100
- influence share of the strong model when aligned with capabilities: up to 60%
- communication vector dimensionality: 3 (tokens, rounds, effort)
- number of independent repeated simulations: 50