Inside the Language Model: A Primer on Transformers and Their Measurement for Psychologists

Psychologists who study language models increasingly meet results from inside the model: hidden states, layers, probes, lenses, patching. These results are usually written for readers trained in machine learning. This primer explains how a transformer language model is built and how its inner activity is measured, in plain language and with no mathematics beyond school level. Each concept is illustrated, where possible, with an example from our own studies of meaning reorganization. The aim: a reader trained in psychology can follow mechanistic work, judge what it can and cannot show, and enjoy it.