Mech Interp Finding: Double Descent Is a Phase Transition Between Memorization and Generalization

gordic_aleksa · x · 2026-09-15

Aleksa Gordić surfaced a lesser-known finding from Anthropic's mech-interp blogs: double descent can be understood as a phase transition between memorizing individual examples and learning reusable structure.

In small-data regimes, models memorize individual training examples by treating datapoints themselves as "features," assigning them distinct directions in hidden space (the red polytopes in the bottom row of the plot; blue ones are weight features).
As dataset size grows, models transition from datapoint features to generalizing features that recur across many examples — the memorized structure gets destroyed, effectively undoing the initial memorization.
The messy transition between the two strategies produces the double-descent bump in test loss.

Generalization appears to emerge once each underlying feature has appeared multiple times in different combinations (10 occurrences per feature in the simplest experiments); repeated datapoints compete with reusable features for model capacity.

Original post →

Mech Interp Finding: Double Descent Is a Phase Transition Between Memorization and Generalization

More from Models