Cross-Model CKA: Low-Frequency Smoothing Bias in EVO2/FAESM–RFdiffusion3 Alignment

Cheng Ding
ShanghaiTech University
dingcheng2024@shanghaitech.edu.cn

This repository contains the complete analysis pipeline for the paper:

"Cross-Model CKA Alignment Between EVO2 and RFdiffusion3 Reflects Low-Frequency Smoothing Bias Rather Than Structural Convergence"

Repository Structure

├── paper/                     # Manuscript (PDF, DOCX, source)
├── scripts/                   # Analysis pipeline
│   ├── core/                  # Position mapping, CKA library
│   ├── scan/                  # Full head-pair CKA scan
│   ├── controls/              # Max-null, Pearson, segment analysis
│   ├── diagnostics/           # Frame-shift, position controls, residual scans
│   └── report/                # Report generation
├── data/                      # Processed results
│   ├── matrices/              # CKA matrices
│   ├── nulls/                 # Null distributions
│   └── figures/               # Publication figures
├── LICENSE
└── README.md

Key Results

Control	EVO2×RF3	FAESM×RF3	FAESM×EVO2
Raw CKA	0.851	0.870	0.101
Cubic residual	0.004	0.272	0.001
Fourier20 residual	0.000	0.024	0.000
Delta excess (obs−shuf)	−0.005	−0.005	—
Retrieval top-1	1.0%	0.0%	—

Citation

@article{ding2025crossmodelcka,
  title={Cross-Model CKA Alignment Between EVO2 and RFdiffusion3 Reflects
         Low-Frequency Smoothing Bias Rather Than Structural Convergence},
  author={Ding, Cheng},
  year={2025},
  eprint={arXiv:XXXX.XXXXX}
}

License

MIT License. See LICENSE for details.