Logit lens, attribution and causal methods

A · 10-second intuition Read intermediate layers through the output head to see what the model is "leaning toward", then use interventions to test what actually matters.