Measured model probabilities · 16 held-out prompts

Which part of an attention head rescues an answer?

The conversation retains an incorrect earlier answer. Replace the selected head's complete message, its value content, or its attention weights with the corresponding computation from a matched correct-history conversation.

Scroll horizontally to inspect the full plot.

Hover over a pair, or focus it with the keyboard, to inspect a prompt.

Qwen2.5-0.5B-Instruct; zero-indexed layer 16, head 7, selected on separate discovery prompts. Probabilities are conditional on the two allowed answer tokens, not unrestricted generation. The same sixteen held-out prompts are present in all three intervention views. Lines show individual prompts; dashed horizontal marks show means. Values are measured model probabilities, not sampled error frequencies or confidence intervals. Download measured probabilities (CSV).