Aligning AI-driven Discovery with Human Intuition

Kevin Zhang, Judah Goldfeder, Hod Lipson

TL;DR

Researchers have developed the Temporally-Informed Dynamics Encoder (TIDE), a framework that forces AI to discover physical state variables that are smooth and continuous. By adding a simple time-derivative penalty to the learning process, the AI naturally gravitates toward variables that mirror human intuition—like angles and velocities—making automated scientific discovery far more interpretable and collaborative.

What Problem Does This Paper Address?

When we look at a swinging pendulum, we intuitively describe it using its angle and speed. However, a dynamical system can be mathematically described by an infinite number of variable combinations. An AI tasked with "modeling the system" might find a set of variables that are perfectly predictive but look like static-filled nonsense to a human scientist.

This is the non-uniqueness problem. Because many different mathematical representations can explain the same data, AI often settles on "black box" representations that lack physical significance. This creates a massive hurdle for human-AI collaboration: if we can't understand the variables the AI is using, we can't verify its logic or integrate its "discoveries" into our existing body of scientific knowledge. The authors set out to find a general principle that encourages AI to pick the "human-like" variables without being told what they are beforehand.

Methods

The authors introduce TIDE, which builds upon a Variational Autoencoder (VAE) architecture. The process follows a two-step pipeline:

  1. Intrinsic Dimension (ID) Estimation: The model first compresses high-dimensional data (like video frames) into a high-dimensional latent space. It then uses the DANCo algorithm to estimate the "intrinsic dimension"—the minimum number of variables needed to describe the system.
  2. The TIDE Network: A second network is trained using that specific bottleneck dimension.

The "secret sauce" of TIDE is Time-Derivative Regularization ($L_{reg}$). The authors argue that human-interpretable variables (like position, pressure, or temperature) are typically smooth and differentiable over time. TIDE penalizes the model if its learned latent variables exhibit abrupt, discontinuous jumps.

By jointly optimizing for three things—reconstruction of the data (ELBO), prediction of future states (Dynamics Module), and temporal smoothness (Regularization)—the model is nudged toward a latent space that aligns with physical intuition. Finally, the researchers use Symbolic Regression (PySR) to turn these learned latent variables into explicit mathematical formulas.

Datasets

The model was evaluated on nine benchmark video datasets, categorized into:

  • Synthetic Systems: Reaction-diffusion, circular motion, single pendulum, double pendulum, and elastic pendulum.
  • Real-World Systems: A swing stick, an "air dancer" (inflatable tube man), a lava lamp, and a fire flame.

These datasets provide a range of complexity, from simple periodic motion to the chaotic, high-dimensional dynamics of fire and fluid.

Key Insights and Ideas

  • Smoothness as a Proxy for Meaning: The most profound insight is that temporal smoothness is a powerful inductive bias for interpretability. By simply asking the AI to "not be jumpy," it naturally discovers variables that correlate strongly with the ones humans have used for centuries.
  • Recoverability via Symbolic Math: Because TIDE produces smooth variables, symbolic regression actually works. Previous methods produced "fragmented" latent spaces that were impossible to fit with simple equations. TIDE allows us to move from "AI says this happens" to "Here is the formula for the variable the AI found."
  • Improved ID Estimation: Enforcing smoothness actually makes it easier for the model to identify the true number of degrees of freedom in a system. TIDE’s estimates of intrinsic dimension were consistently closer to the ground truth than the baseline.

Key Results

  • Qualitative Alignment: Visualizations of TIDE’s latent variables show smooth, continuous curves that track perfectly with human variables like $\theta$ (angle). In contrast, the baseline (Neural State Variables) produced erratic, discontinuous signals.
  • Mutual Information (MI): TIDE showed significantly higher MI with human variables across the board. For a single pendulum, TIDE achieved an MI of 0.68 compared to the baseline's 0.48, indicating a much stronger correlation with human concepts.
  • Analytical Accuracy (AMSE): The error in fitting symbolic equations to the latent variables was orders of magnitude lower for TIDE. For the circular motion dataset, TIDE’s AMSE was 0.005, while the baseline was 0.34.
  • Real-World Generalization: Even in complex systems like a fire flame, where "ground truth" variables are unknown, TIDE identified a stable set of 8 variables, whereas the baseline fluctuated wildly.

Limitations and Future Work

The authors acknowledge a few key hurdles:

  • Data Efficiency: The method relies on the "stable state variable representation" assumption, meaning it requires a high enough frame rate to capture smooth transitions. It might struggle with sparse or extremely noisy data.
  • Non-Uniqueness Persists: While TIDE aligns better with human intuition, it doesn't guarantee a 1-to-1 match. The AI might still find a "smooth" variable that is a complex combination of two human variables.
  • Future Directions: The team suggests exploring semi-supervised learning, where a model could be "pre-taught" what an angle or velocity looks like in a simple system and then asked to find similar "types" of variables in a completely new, complex environment.

Summary updated

Abstract

As data-driven modeling of physical dynamical systems becomes more prevalent, a new challenge is emerging: making these models more compatible and aligned with existing human knowledge. AI-driven scientific modeling processes typically begin with identifying hidden state variables, then deriving governing equations, followed by predicting and analyzing future behaviors. The critical initial step of identification of an appropriate set of state variables remains challenging for two reasons. First, finding a compact set of meaningfully predictive variables is mathematically difficult and under-defined. A second reason is that variables found often lack physical significance, and are therefore difficult for human scientists to interpret. We propose a new general principle for distilling representations that are naturally more aligned with human intuition, without relying on prior physical knowledge. We demonstrate our approach on a number of experimental and simulated system where the variables generated by the AI closely resemble those chosen independently by human scientists. We suggest that this principle can help make human-AI collaboration more fruitful, as well as shed light on how humans make scientific modeling choices.
BibTeX
@inproceedings{zhang2026aligning,
  author = {Kevin Zhang and Judah Goldfeder and Hod Lipson},
  title = {{Aligning AI-driven Discovery with Human Intuition}},
  booktitle = {AI4Math @ ICML},
  year = {2026},
  url = {https://openreview.net/forum?id=CPNVU3NkZV},
}
Figure