Skip to content
GKgkml.dev
← All scenarios
LLM internalsbrutal~25 min

You ablated a head and the model got worse at everything

You are investigating which attention head implements a name-copying behaviour in GPT-2 medium. You zero-ablate head 9.6 and the behaviour degrades substantially: the logit difference between the correct and incorrect name falls by 2.1.

You also notice that on a set of unrelated control prompts, average loss rises by 0.4. And when you ablate a randomly chosen head instead, the target logit difference falls by 0.3 and control loss rises by 0.15.

You are about to write that head 9.6 implements name copying.

What can you legitimately conclude from this evidence, what can you not, and what experiments would you need before that claim is defensible? Be specific about what each additional experiment rules out.

Choose how you want to be interrogated

The same case under two modes is two different exercises. If you are unsure, RCA drill is the one that trains root-cause analysis most directly.

Before you start: you begin locked. The Mentor will give you nothing — not a nudge, not a category — until you state a specific position and the mechanism you think produces it. Four genuine attempts unlock the resolution. There is no shortcut.