A Distributional View for Visual Mechanistic Interpretability: KL-Minimal Soft-Constraint Principle
Guancheng Zhou, Yisi Luo, Zhengfu He, Zhenyu Jin, Xuyang Ge, Wentao Shu, Deyu Meng, Xipeng Qiu · 2026
Read A Distributional View for Visual Mechanistic Interpretability: KL-Minimal Soft-Constraint Principle by Guancheng Zhou, Yisi Luo, Zhengfu He, Zhenyu Jin, Xuyang Ge, Wentao Shu, Deyu Meng, Xipeng Qiu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-17504.txt.
Introduction
Zhou and colleagues establish a distribution view for visual mechanistic reading. The view models how a feature activation shifts the natural image distribution, framing the task as KL-minimal tuning. Prior paradigms carry biases. They may drift from natural images and read as uninterpretable, or fail to activate model traits and read as unfaithful. A KL-minimal soft-constraint principle balances clarity against faithfulness, realized through energy-guided diffusion sampling. Tests on DINOv3 validate the theory and approach.
Catalog record
| CREATOR | Guancheng Zhou, Yisi Luo, Zhengfu He, Zhenyu Jin, Xuyang Ge, Wentao Shu, Deyu Meng, Xipeng Qiu |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 145038 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.17504