Mirror-Neuron Patterns in AI Alignment (arXiv v2)
Posted November 5 2025
The revised version of our paper on mirror-neuron-like activation patterns in artificial neural networks is now live on arXiv. It presents evidence that cooperative learning can produce shared self-other neural representations - an analog of affective empathy - in artificial agents. It presents a formal framework for the emergence of mirror neuron patterns in AI systems, and it introduces the Checkpoint Mirror Neuron Index (CMNI), a metric to measure such patterns across training stages. These shared representations point to a possible basis for AI alignment, where systems learn to understand others' states as part of their own.