Revisiting the Platonic Representation Hypothesis: An Aristotelian View

Fabian Gröger* · Shuo Wen* · Maria Brbić

---

The Aristotelian Representation Hypothesis

Neural networks, trained with different objectives on different data and modalities, are converging to shared local neighborhood relationships.

The Platonic Representation Hypothesis suggests that representations from neural networks are converging to a common statistical model of reality. We show that the existing metrics used to measure representational similarity are confounded by network scale: increasing model depth or width can systematically inflate representational similarity scores. To correct these effects, we introduce a permutation-based null-calibration framework that transforms any representational similarity metric into a calibrated score with statistical guarantees. We revisit the Platonic Representation Hypothesis with our calibration framework, which reveals a nuanced picture: the apparent convergence reported by global spectral measures largely disappears after calibration, while local neighborhood similarity retains significant agreement across different modalities.

---

Repository Structure

├── calibrated_similarity/    # Standalone Python package (pip installable)
│   ├── calibration.py        # Core algorithms (Algorithm 1 & 2 from paper)
│   └── __init__.py
├── aristotelian/             # Research code for paper experiments
│   ├── metrics/              # Similarity metrics (CKA, kNN, RSA, CCA, etc.)
│   ├── experiments/          # Experiment utilities
│   ├── prh/                  # PRH replication code
│   ├── utils/                # Utility functions
│   └── style/                # Plotting style
├── scripts/
│   ├── experiments/          # Paper experiment runners
│   └── plots/                # Figure generation
├── tests/                    # Test suite

---

Using the calibration package only

pip install calibrated-similarity

See the package documentation for usage examples.

---

Reproducing Paper Experiments

# Step 1: Run experiments
python -m scripts.experiments.cli --device cuda

# Step 2: Generate all figures
python -m scripts.plots.experiments --sections all

Citation

If you find this work useful, please cite:

@article{groger2026revisiting,
  title   = {Revisiting the Platonic Representation Hypothesis: An Aristotelian View},
  author  = {Gr{\"o}ger, Fabian and Wen, Shuo and Brbi{\'c}, Maria},
  journal = {arXiv preprint},
  year    = {2026},
}

License

MIT License - see LICENSE for details.