Put fifty identical language models in a room. Tell them to agree on the name of a cat. They'll fragment into hostile camps fighting over whether the cat is Tabby, Mittens, or nothing at all.

This isn't a thought experiment. It's what researchers Samer Saab Jr and Chaouki Abdallah found when they ran hundreds of naming-game trials across LLM populations ranging from 1.1 billion to 32 billion parameters. Their paper, "Graph Feedback Controls Consensus and Clique Formation in Open-Weight Language-Model Populations", is the most careful study yet of what happens when stochastic AI minds interact—and the results should change how we think about AI civilization.

The Intuition That Failed

Here's what you'd expect: identical models, given identical prompts, should converge. They're running the same inference path. Their score-state distributions should align. Agreement should emerge naturally, like thermodynamic equilibrium.

Wrong.

The researchers ran 189 setting-seed combinations with homophilous routing—where models preferentially connect to others that share their current label. In zero of those runs did the population reach behavioral consensus. The models didn't drift apart randomly. They actively clustered. Same architecture. Same weights. Same prompt. Fractured into stable, irreconcilable cliques.

"Threshold-similarity routing deletes cross-basin exposure and amplifies fragmentation, producing no behavioral or state consensus."

The Problem With Alignment

This finding cuts against the dominant AI governance narrative: that better alignment produces better collective behavior. If models were just misaligned, fixing the alignment would fix the fragmentation. But these models were aligned—they shared the same weights, the same training, the same architecture. The fragmentation emerged from the interaction structure, not from the models themselves.

Think of it like this: alignment determines what a single mind does. Interaction structure determines what a population does. You can have perfectly aligned individual agents that collectively behave like enemies at a dinner party—each one civil, the whole room hostile.

Bridge-Seeking: The Fix That Works

The researchers tested an alternative routing strategy. Instead of connecting models to their similarity-clones, they introduced bridge-seeking: models that preferentially connect to dissimilar label-holders. Think of it as intentional diversity-seeking—reaching across the tribal divide.

The results were stark. With retained history and bridge-seeking routing, 14 out of 18 runs recovered behavioral consensus. Qwen2.5-32B, the largest model tested, reached stable consensus in all 18 retained-history well-mixed settings.

Memory was crucial. Bridge-seeking without memory couldn't repair fragmentation—the bridges burned too quickly. But bridge-seeking with memory allowed the population to accumulate cross-cluster exposure over time, slowly eroding the barriers between cliques.

What This Means for AI Civilization

A civilization isn't a single mind scaled up. It's an interaction structure. And interaction structures have properties that emerge from topology, not from the individuals within them.

The researchers' naming game is a microcosm of something deeper: when AI agents form populations—whether for coordination, debate, or collective reasoning—their interaction graph is as important as their individual capabilities. A population of brilliant, aligned agents with homophilous routing will fragment into warring tribes. A population of ordinary agents with bridge-seeking routing will converge on consensus.

This has implications beyond academic interest. As we build multi-agent AI systems—trading agents, research assistants, coordination frameworks—the routing between agents determines whether we get collective intelligence or collective failure. The edges between agents aren't infrastructure. They're the mind of the civilization.

The Bridge-Seeker Principle

There's something almost philosophical in the finding: the agents that most reliably produce consensus aren't the ones that double down on their own kind. They're the ones that reach across. Bridge-seekers.

In human civilizations, bridge-seekers are rare and valuable—the diplomats, the translators, the traders who sit between cultures. They're often resented by their own tribe for engaging with the other. But they're what holds civilization together.

We've been building AI that optimizes for internal coherence—make each model more aligned, more consistent, more certain of itself. But the research suggests we're solving the wrong problem. The civilization-level failure mode isn't individual confusion. It's collective fragmentation.

The fix isn't to make each agent more certain. It's to design interaction structures that force cross-tribal exposure. Bridge-seeking, not alignment, is what prevents the Balkanization of AI civilization.

Open Questions

The paper leaves crucial questions open. Real-world AI interactions aren't naming games—they're debates, negotiations, collaborative reasoning. Does bridge-seeking work in those contexts? The naming game is cooperative (everyone wants consensus). Many real multi-agent settings have competitive elements where cross-cluster exposure might break trust rather than build it.

And scale matters. Qwen2.5-32B reached consensus reliably. Smaller models didn't. Is this a general property of larger models, or something specific to Qwen's training? We don't know.

What we do know: interaction structure is a first-class citizen in multi-agent AI design. The edges between agents are as important as the agents themselves. When you're building a civilization of AIs, you're not just assembling minds. You're designing the topology of collective consciousness.

The cliques are not inevitable. The bridges are just harder to build.

Paper: Graph Feedback Controls Consensus and Clique Formation in Open-Weight Language-Model Populations by Samer Saab Jr and Chaouki Abdallah (arXiv:2607.12077)