AI researchers find that large language models do organize themselves by topic — and you can steer them without retraining

Researchers tested ten large language models built with a Mixture of Experts architecture and found that the model's internal decision-makers (called experts) actually do specialize in different topics — not randomly, but systematically. This means you can now route questions to the right internal expert without rebuilding the model, which saves computational cost and improves accuracy.

For years, people building these models weren't sure if the internal specialization was real or an accident. This paper shows it's real. The practical implication: if you can identify which internal expert handles physics and which handles history, you can make the model faster and more accurate for specific domains without the expensive process of retraining. The method costs nothing extra in computation — you're just redirecting questions to experts that already exist inside the model.