Research map / Architectures and efficiency / Attention and transformers
Align Once to Explain: Feature Alignment for Scalable B-cosification of Foundational Vision Transformers
Raphael Maser, Siddhartha Gairola, Sukrut Rao, Bernt Schiele
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace B-cosification with ALOE feature alignment | SigLIP2 ViT-B/16 B-cos · ImageNet-1k linear probe | top-1 accuracy (%): 75.84 -> 83.67 | Improved |
| replace B-cosification with ALOE | DINOv3 ViT-B/16 B-cos · k-NN average | accuracy (%): 50.77 -> 82.63 | Improved |
| replace B-cosification with ALOE | SigLIP2 ViT-B/16 B-cos · ImageNet-1k zero-shot | top-1 accuracy (%): 61.01 -> 77.20 | Improved |
| use B-cos ALOE attributions instead of post-hoc AttnLRP | SigLIP2 ViT-B/16 · GridPG localization | Grid-PG score (%): 54.4 -> 84.2 | Improved |
| add layer-wise alignment at {2L/3, L} to global-only alignment (ablation) | ALOE SigLIP2 · 10-dataset linear probe average | accuracy (%): 75.51 -> 85.24 | Improved |
| supervise all layers instead of three depths (ablation) | ALOE SigLIP2 · 10-dataset linear probe average | accuracy (%): 85.42 -> 84.93 | No clear effect |
| reduce alignment data from 100% to 1% of YFCC15M (ablation) | ALOE SigLIP2 · ImageNet-1k linear probe | top-1 accuracy (%): 83.80 -> 83.33 | No clear effect |
| ALOE vs teacher on ImageNet-1k linear probe | ALOE SigLIP2 ViT-B/16 · ImageNet-1k linear probe | top-1 accuracy (%): 84.24 -> 83.67 | Hurt |
| cosine alignment loss vs MSE/InfoNCE (ablation) | ALOE · linear probe | accuracy | Mixed |
Closest papers by meaning
- B-Cos Alignment for Inherently Interpretable CNNs and Vision Transformers (TPAMI 2024 · similarity 0.89)
- B-cosification: Transforming Deep Neural Networks to be Inherently Interpretable (NeurIPS 2024 · similarity 0.88)
- Beyond Scalars: Concept-Based Alignment Analysis in Vision Transformers (NeurIPS 2025 · similarity 0.86)
- Not Too Generative, Not Too Discriminative: The Human Alignment Sweet Spot (NeurIPS 2026 · similarity 0.85)
- $\Delta \mathrm{Energy}$: Optimizing Energy Change During Vision-Language Alignment Improves both OOD Detection and OOD Generalization (NeurIPS 2025 · similarity 0.85)
- Topology-Aware Representation Alignment for Semi-Supervised Vision-Language Learning (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
