Gradient Similarity in Data Attribution Tracks Surface Form Over Task Content

Gradient-based data attribution methods prioritize answer format over task relevance, with gradient alignment dropping from 0.4 to near 0.0 when only answer formats change. This effect persists across model scales and through post-training, suggesting gradient similarity reflects surface patterns rather than semantic skill acquisition.