Convergent Representations (The Platonic Representation hypothesis)

Because we share a single physical reality, the internal representations that large-scale models learn may converge toward a unified representation space.

Huh2024platonic is the seminal paper that explicitly proposed this hypothesis of convergence (“The Platonic Representation Hypothesis”).

Edamadaka2025universally argues that this convergence occurs for representations of “matter” across science foundation models.

Groger2026aristotelian pushes back, showing that common similarity metrics are confounded by network scale; after null-calibration, global convergence largely disappears but local neighborhood structure survives — an “Aristotelian” rather than “Platonic” view.