Learned Structure Shows Up Earlier in the Task's Own Basis: A Controlled Look at a Synchronisation Precursor in Grokking
The Frequency Synchronization Degree (FSD) is a metric that predicts 'grokking,' a phase where an AI learns to generalize, before it actually happens. The research shows FSD synchronizes before generalisation across various math tasks. Crucially, this precursor is due to the task-faithful representation basis used, not just the Fourier circuit itself.