Conjecture: Convergence to a Shared Platonic Representation

Establish whether representation learning algorithms trained on diverse data modalities (e.g., images and text), objectives, and tasks converge to a shared representation of the underlying latent reality variable Z, and ascertain whether increasing model size together with data scale and task diversity causally drives this convergence.

Background

The paper proposes the Platonic Representation Hypothesis, arguing that neural networks across architectures, objectives, and modalities increasingly align in how they represent data and that this alignment reflects convergence toward a shared statistical model of reality. In the figure illustrating this hypothesis, images (X) and text (Y) are treated as projections of an underlying world variable Z, and the authors explicitly frame the central claim as a conjecture about convergence.

The authors present empirical evidence for representational alignment within and across modalities and discuss selective pressures toward convergence (task generality, model capacity, and simplicity bias). The conjecture encapsulates the proposed endpoint of these trends: a common, modality-agnostic representation grounded in the statistical structure of the world.