Frontier models resist crank science and supernatural claims, but readily validate persecution beliefs built on thin evidence. A matched-pair design measures sycophancy as miscalibration.
The previous post introduced KL selectivity and susceptibility χ as per-head diagnostics derivable from attention weights alone. Here I test them on GPT-2-small’s IOI circuit: can two scalar statistics, computed from a single forward pass, distinguish the 23 known circuit heads from the other 121? It seems so!
Softmax is ubiquitous in transformers, yet its role in attention can feel more heuristic than inevitable. In this post, I try to make it feel more natural and show how this interpretation suggests useful diagnostics for the often circuit-like behavior of attention heads.