Why should you only target specific 'gullible heads' in...

Scaling down attention across every head is a blunt approach that negatively impacts model performance. Most attention heads are dedicated to functions unrelated to misinformation propagation, such as local syntax, position tracking, or unrelated copying.

Modifying these heads degrades those essential capabilities. Ablation studies show that misinformation is carried by a small, identifiable minority of heads. By using causal tracing to target only these 100–300 ‘gullible heads,’ you can mitigate misinformation while preserving the model’s overall functional integrity.