Preprint

World models converge toward Platonic-like representations.

We test whether DINO-WM predictors built on different frozen visual encoders develop comparable latent geometry after learning the same action-conditioned dynamics.

Read the paper

Why It Matters

A concrete way to study representation alignment

If different world models learn increasingly similar latent geometry, then their internal states may be comparable in a meaningful way. This makes representation alignment a concrete object of study, rather than only a metaphor about similar understanding.

The page keeps the argument simple: different frozen encoders feed the same DINO-WM training setup, and we ask whether the predictor states move toward a shared transition-compatible structure.

Setup

Train the same world model through different frozen visual priors

Conceptual diagram of frozen visual priors feeding a shared DINO-WM predictor. — DINOv2, SigLIP, MAE, and ResNet encoders create different input feature spaces. The predictor architecture and transition objective are held fixed, so the comparison focuses on the learned predictor hidden states.

Diagram showing mutual k-nearest-neighbor alignment between two predictor hidden-state spaces. — m-kNN compares whether two predictors assign the same validation states to similar local neighborhoods. Higher overlap means stronger preservation of local latent topology.

Benchmark

Five environments, one transition-learning lens

The evaluation spans navigation, contact-rich manipulation, and deformable dynamics: PointMaze, PushT, Wall, Granular, and Rope.

Evidence

The core signal is alignment, then the encoder-bias check

Predictor Geometry Aligns

Across training, ViT-based DINO-WM predictors show stronger local neighborhood alignment than the ResNet baseline.

m-kNN similarity trends across environments. — ViT-based predictors exhibit stronger local-neighborhood alignment than the ResNet baseline, with SigLIP and MAE often showing progressive movement toward the reference predictor.

Not Just Encoder Inheritance

Figure 5 and Figure 6 address the main alternative explanation: maybe the predictors only inherit or recover frozen encoder geometry.

Figure 5: comparison of encoder-space and predictor-space alignment. — Figure 5 separates encoder-space similarity from predictor-space recovery. Figure 6 checks whether predictors merely drift back toward their own encoders. Together, they support the interpretation that training reorganizes inherited visual features into transition-compatible hidden states.

Figure 6: within-model encoder-predictor geometry analysis. — Figure 5 separates encoder-space similarity from predictor-space recovery. Figure 6 checks whether predictors merely drift back toward their own encoders. Together, they support the interpretation that training reorganizes inherited visual features into transition-compatible hidden states.

Main Findings

What the paper argues

Predictive consistency acts as an alignment pressure

Under a shared transition objective, strong ViT-based predictors become more geometrically compatible in hidden-state space.

The alignment is not just inherited

Encoder-control analyses show that predictor states transform frozen visual features and later recover cross-model alignment in predictor space.

Architecture sets the boundary

ResNet remains separated from ViT-based predictors, suggesting that Platonic-like convergence depends on architectural and interface compatibility.

Scope

A conservative reading of Platonic-style alignment

This page argues for a convergence tendency, not a universal law. The evidence is strongest for ViT-based world-model predictors trained under comparable visual-action prediction settings.

ResNet remains a boundary case rather than an exception to hide. Stitching provides a secondary functional check, but the main argument here stays centered on predictor geometry and encoder-inheritance controls.

Paper

Read and cite the work

PDF

@misc{prhwm2026,
  title  = {Platonic Representation Hypothesis on World Models},
  author = {Wenhow Li and Chengwei MA and Hui Xiong and Ying-Cong Chen and Lei Zhang},
  year   = {2026},
  note   = {Preprint}
}