The two clocks and the innovation window: When and how generative models learn rules

TL;DR

This paper investigates the timing of rule learning and memorization in generative models, revealing a two-clock structure that explains when models genuinely innovate versus memorize training data.

Contribution

It introduces the concepts of rule-learning time and memorization time, analyzing their dependence on model capacity, rule complexity, and dataset size across different architectures.

Findings

Rule learning time increases with rule complexity and decreases with model capacity.

Memorization time scales nearly linearly with dataset size and is invariant to rule complexity.

The 'innovation window' between rule learning and memorization can vanish, affecting genuine model innovation.

Abstract

Generative models trained on finite data face a fundamental tension: their score-matching or next-token objective converges to the empirical training distribution rather than the population distribution we seek to learn. Using rule-valid synthetic tasks, we trace this tension across two training timescales: $\tau_{\mathrm{rule}}$, the step at which generations first become rule-valid, and $\tau_{\mathrm{mem}}$, the step at which models begin reproducing training samples. Focusing on parity and extending to other binary rules and combinatorial puzzles, we characterize how these two clocks, $\tau_{\mathrm{rule}}$ and $\tau_{\mathrm{mem}}$, depend on key aspects of the learning setup. Specifically, we show that $\tau_{\mathrm{rule}}$ increases with rule complexity and decreases with model capacity, while $\tau_{\mathrm{mem}}$ is approximately invariant to the rule and scales nearly linearly with dataset size $N$. We define the \emph{innovation window} as the interval $[\tau_{\mathrm{rule}}, \tau_{\mathrm{mem}}]$. This window widens with increasing $N$ and narrows with rule complexity, and may vanish entirely when $\tau_{\mathrm{rule}} \geq \tau_{\mathrm{mem}}$. The same two-clock structure arises in both diffusion (DiT) and autoregressive (GPT) models, with architecture-dependent offsets. Dissecting the learned score of DiT models reveals a corresponding evolution of the optimization landscapes, where rule-valid samples' basins expand substantially around $\tau_{\mathrm{rule}}$, while training samples' basins begin to dominate around $\tau_{\mathrm{mem}}$. Together, these results yield a unified and predictive account of when and how generative models exhibit genuine innovation.