@FinanceYF5: The Platonic representation hypothesis is mostly a statistical illusion. New research shows that the apparent 'global convergence' in scaled AI models is actually a mathematical artifact caused by selection bias in model width and depth. Once calibrated, global convergence disappears.

The Platonic Representation Hypothesis is largely a statistical illusion.

New research shows that the apparent “global convergence” in scaled-up AI models is actually a mathematical artifact caused by selection bias in model width and depth.

Once calibrated, global convergence disappears. 🧵 https://t.co/dVuL8kN9n8

2/ In Revisiting the Platonic Representation Hypothesis: An Aristotelian View, Fabian Groeger, Shuo Wen, and Maria Brbic demonstrate that standard representation similarity metrics are systematically biased by network dimensionality.

Let’s dive into the math.

3/ Confound 1: Model width.

Under a fully independent null hypothesis, the expected squared Frobenius norm of the cross-covariance does not vanish.

The raw baseline of metrics like Centered Kernel Alignment (CKA) scales as O(d/n), simulating alignment in wide models.

4/ Confound 2: Model depth.

To find alignment, researchers exhaustively evaluate all layer pairs (La x Lb) and report the maximum.

Extreme value theory shows the expected maximum grows with the search space: E[T_max] <= mu + Csigmasqrt(log M). Deeper models “by chance” appear more aligned.

5/ The authors propose a metric-agnostic, permutation-based calibration method.

Instead of correcting cell by cell, they perform a consistent shuffle across all layers of a model to build an empirical null distribution of the maximum score.

Scores falling below the null distribution are mapped to 0.

6/ Applying this framework to 204 vision-language model pairs reveals a clear split:

• Global spectral metrics (e.g., CKA) calibrate to zero. • Local neighborhood metrics (mKNN) remain robust.

What models agree on are topological neighborhoods, not global spaces.

7/ Limitation: The framework assumes exchangeability of samples under the null.

If the dataset has sequential, spatial, or hierarchical dependencies, naive permutation fails and inflates Type I error.

It also scales as O(K * La * Lb), making experiments on large models computationally expensive.

8/ This is an important correction that reshapes how we evaluate foundation models.

Going forward, raw similarity scores cannot be reported directly across different model scales.

Without calibration, any conclusion about representation convergence is mathematically indefensible.

9/ This shifts the perspective from a Platonic view (a perfect global metric space) to an Aristotelian view (shared local topological relationships).

Models learn the same relative neighbor structure, not a common coordinate space.

10/ Full review: https://arxiviq.substack.com/p/revisiting-the-platonic-representation…

Paper: https://arxiv.org/abs/2602.14486

Should representation alignment use local or global metrics? Discussion welcome.

11/ Visualization: Aristotelian correction vs. Platonic illusion.

That’s all. Original author @che_shr_cat

If you enjoyed this thread:

1. Follow me (@FinanceYF5)
2. Like + RT the first post below

Someone used a digital map to label all the neighborhoods of Manhattan, New York.

Harlem, SoHo, Hell’s Kitchen, Tribeca, Financial District…

Each neighborhood in a different color, paired with satellite top-down views — the city finally “makes sense.”

Bookmark for later use.