📊 statistics

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

This paper identifies a critical "innovation window" in generative models defined by the temporal gap between learning abstract rules ( $\tau_{\mathrm{rule}}$ ) and memorizing training data ( $\tau_{\mathrm{mem}}$ ), demonstrating that this window's existence and duration depend on dataset size, rule complexity, and model capacity across both diffusion and autoregressive architectures.

Original authors: Binxu Wang, Emma Lucia Byrnes Finn, Bingbin Liu

Published 2026-05-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Binxu Wang, Emma Lucia Byrnes Finn, Bingbin Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot to draw pictures based on a strict set of rules, like "every picture must have an even number of red dots." You give the robot a finite stack of example drawings to study.

This paper investigates a fascinating tug-of-war inside the robot's brain as it learns. The authors discovered that the robot doesn't learn all at once; instead, it operates on two different clocks that tick at different speeds.

Here is the story of those two clocks and the "magic window" between them.

The Two Clocks

Clock 1: The "Rule Learner" (τrule)
This clock measures when the robot finally understands the logic of the game.

What happens: At first, the robot is just guessing. Then, suddenly, it starts drawing pictures that follow the "even number of red dots" rule perfectly. It has learned the abstract concept.
What slows it down: The more complex the rule, the slower this clock ticks. If the rule involves checking groups of 10 dots instead of just 2, the robot takes much longer to figure it out. It's like trying to learn a complex chess strategy versus learning how a pawn moves.

Clock 2: The "Photocopier" (τmem)
This clock measures when the robot starts just copying the exact pictures it saw in its training stack.

What happens: Eventually, the robot stops trying to be creative and starts memorizing the specific examples you gave it. If you ask for a new picture, it just pulls an old one from its memory.
What controls it: Surprisingly, this clock doesn't care much about how hard the rule is. Instead, it is driven by how many examples you gave the robot. The bigger the stack of training pictures, the longer it takes for the robot to start copying them. It's like a student who needs to read a whole library before they start memorizing specific books word-for-word.

The "Innovation Window"

Here is the most exciting part: The Gap.

There is a specific period of time between Clock 1 and Clock 2.

Before Clock 1: The robot is confused and making mistakes.
Between Clock 1 and Clock 2 (The Innovation Window): The robot understands the rules perfectly and can create new, valid pictures that it has never seen before. It is being genuinely creative.
After Clock 2: The robot stops being creative and just starts photocopying the training data.

The size of this "Innovation Window" depends on your setup:

If the rule is very hard: The robot takes so long to learn the rule (Clock 1 is slow) that it starts memorizing the examples (Clock 2) before it ever figures out the logic. The window closes, and the robot never gets to be creative.
If you give it more data: The robot takes longer to memorize (Clock 2 slows down), but it learns the rule at the same speed. This widens the window, giving the robot more time to be creative before it starts copying.

How the Robot's Brain Changes

The authors looked inside the robot's "brain" (specifically the mathematical landscape it creates) to see what was happening:

Early Stage: The robot learns to make dots that look like valid pixels (getting close to the "Boolean cube").
Rule Learning: The robot builds a "valley" in its brain for all the correct answers. Any picture that follows the rules gets pulled into this valley.
Memorization: Later, the robot digs a deeper valley specifically for the exact pictures it saw in training. These specific pictures become "sticky" attractors. If the robot starts near a training picture, it gets sucked into copying that exact image, even if there are other valid pictures nearby.

Does this apply to other robots?

Yes! The authors tested this on two different types of AI:

Diffusion Models (like image generators): They follow this two-clock pattern.
Autoregressive Models (like text generators): They also follow this pattern, though they learn and memorize slightly faster.

They also tested this on other puzzles, like Sudoku and counting rules. The same two clocks appeared: first the robot learns the logic, then it starts memorizing.

The Bottom Line

Generative AI isn't just a magic box that gets smarter over time. It goes through distinct phases. It first learns the rules, then it has a brief moment of true creativity, and finally, it settles into memorization.

If you want an AI to be innovative, you need to make sure the "Rule Learning" clock finishes before the "Memorization" clock starts. If the rules are too hard, the robot will give up on understanding them and just start copying your examples instead.

Technical Summary: The Two Clocks and the Innovation Window

Problem Statement

Generative models trained on finite datasets face a fundamental tension: their training objectives (e.g., score-matching in diffusion models or next-token prediction in autoregressive models) converge toward the empirical training distribution rather than the underlying population distribution. While realism can be achieved through memorization, it remains difficult to determine whether a model has genuinely learned the latent rules defining a distribution. In natural data, distinguishing between rule-based generalization and exact memorization is challenging. This paper addresses this by asking: When do generative models learn the true rules of a distribution, and when do they begin to reproduce training samples?

Methodology

The authors employ a controlled synthetic framework to precisely measure rule satisfaction, novelty, and memorization.

Task: The primary testbed is group-level parity on binary images. Each image is divided into groups of size $G$ , where each group must satisfy an even-parity constraint. The parameter $G$ serves as a direct knob for rule complexity (interaction order).
Models: The study utilizes Diffusion Transformers (DiT) and Autoregressive Transformers (GPT) trained on datasets of varying sizes ( $N$ ) and model capacities.
Metrics:
- Rule Accuracy: The fraction of generated samples satisfying the parity constraint.
- Memorization Ratio: The fraction of generated samples (or groups) that exactly match training data.
- Innovation Window: The temporal interval $[\tau_{rule}, \tau_{mem}]$ where generations are rule-valid but not memorized.
Dynamical Analysis: The authors dissect training dynamics using:
- Per-sample state tracking (invalid, rule-violating, valid-novel, valid-memorized).
- Denoising Score Matching (DSM) loss analysis across different noise scales ( $\sigma$ ).
- Visualization of learned vector fields and attractor basins on slices of the Boolean hypercube.

Key Contributions and Results

1. The Two Clocks of Generative Training

The paper identifies two distinct timescales governing generative training:

$\tau_{rule}$ (Rule-Learning Time): The step at which generated samples first become rule-valid.
$\tau_{mem}$ (Memorization Time): The step at which models begin to reproduce training samples.

These two clocks are governed by different factors:

$\tau_{rule}$ increases monotonically with rule complexity ( $G$ ) (delaying by roughly an order of magnitude per unit increase in $G$ ) and decreases with model capacity. It is largely invariant to dataset size $N$ .
$\tau_{mem}$ is approximately invariant to rule complexity but scales nearly linearly with dataset size ( $N$ ) ( $\tau_{mem} \approx 35 N^{1.14}$ for DiT). It is also accelerated by larger model capacity.

2. The Innovation Window

The interval $[\tau_{rule}, \tau_{mem}]$ is defined as the innovation window, during which the model generates rule-valid but non-memorized (novel) samples.

Widening/Narrowing: The window widens with larger datasets (which delay memorization) and narrows with higher rule complexity (which delays rule learning).
Closure: For high complexity ( $G \ge 9$ ) or small datasets, $\tau_{rule} \ge \tau_{mem}$ , causing the innovation window to vanish entirely. In these cases, the model appears to learn the rule only after it has already begun memorizing, meaning the observed accuracy is driven by retrieval rather than generalization.

3. Mechanistic Insights via Vector Fields

Dissecting the learned score of DiT models reveals a corresponding evolution in the optimization landscape:

Early Phase: The model learns to generate on the Boolean hypercube, turning vertices into attractors.
Around $\tau_{rule}$ : Attractors for rule-valid samples expand substantially, suppressing rule-violating corners. The basins of valid samples (both training and held-out) expand symmetrically.
Around $\tau_{mem}$ : The basins of specific training samples expand further, invading nearby held-out valid regions. The learned vector field becomes a poorer denoiser for held-out samples, pointing toward memorized training points even when noised samples are closer to valid novel points.
Noise Scale Dependence: These transitions occur at specific intermediate noise scales ( $\sigma \in [0.2, 2.0]$ ), corresponding to the geometric distances between data points on the Boolean hypercube.

4. Generalization to Other Architectures and Tasks

Autoregressive Models (GPT): GPT models exhibit the same two-clock structure. However, they learn and memorize faster than DiTs of comparable size, resulting in a shorter innovation window. The transition to memorization in GPT is more sudden.
Other Rules: The dynamics extend to other binary rules (e.g., exact- $K$ $K$ , row- $K$ $K$ ) and combinatorial puzzles (Latin squares, Sudoku).
- Complexity Form Matters: While high-degree parity is difficult, global count constraints (exact- $K$ ) are learned rapidly, suggesting that the form of the constraint (multiplicative interactions vs. additive sums) determines difficulty more than globality alone.
- Nested Constraints: Tasks like Sudoku show anisotropic learning (rows learned before columns) and delayed validity compared to simple parity.

Significance and Claims

The paper claims to provide a unified and predictive account of when and how generative models exhibit genuine innovation versus memorization.

Dynamical Explanation: It moves beyond static evaluation to explain why models fail to learn complex rules: the innovation window closes when the time required to learn the rule exceeds the time required to memorize the data.
Architectural Limitations: The results suggest that current diffusion architectures struggle with high-order multiplicative interactions (high $G$ parity), which may underlie difficulties in learning abstract reasoning rules (e.g., RAVEN matrices) or physical conservation laws.
Pathways Forward: The authors suggest that improving rule learning for complex constraints may require architectural modifications (e.g., global broadcasting, structured multiplicative interactions) or curriculum learning strategies, rather than simply scaling data or model size.

The work isolates a fundamental limitation of current generative models: their inductive biases are not well-aligned with high-order global rules, leading to a reliance on memorization when the "innovation window" is closed.