prh-convergence-in-distance-measure-not-accuracy
IN premise — summaries/2026/08/24/huh-2024-prh-s0-abstract.md
Created 2026-08-24T17:10:56+00:00
The Platonic Representation Hypothesis posits that neural network representations converge specifically in how models measure distance between datapoints as they scale, not merely in task accuracy, across architectures, objectives, and data modalities
Summary
When different neural networks are trained at scale, even with different architectures, loss functions, and data types, they don't merely end up with similar task accuracy; they converge on how they measure the geometric distance between data points in their internal space. This implies there is a fairly universal structure to how information should be organized in a representation, independent of specific model design choices, which gives a deeper and more principled target for evaluating or guiding network training beyond just benchmark scores.