What is Universality in LLMs? How to Find Universal Neurons

Universality in LLMs proposes that independently trained neural networks can converge on similar internal mechanisms and features. An experiment is described using two small, independently trained transformer models to identify these "universal neurons." The method involves calculating the Pearson correlation between the MLP activations of both models and comparing this to a baseline of randomly rotated activations to find "excess correlation." Neurons with high excess correlation are flagged as universal, providing evidence that different models can learn similar features despite separate training processes.