Research map / Architectures and efficiency / RNNs and normalization
What Linear Probes Miss: Multi-View Probing for Weight-Space Learning
Eunwoo Heo, Kyeongkook Seo, Jaejun Yoo
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 |
|---|---|---|---|
| add all four probing branches vs XU only (ablation) | MVProbe · Model Jungle (ResNet) | accuracy: 90.42 -> 92.24 | Improved |
| add all four probing branches vs XU only (ablation) | MVProbe · Model Jungle (DINO) | accuracy: 74.17 -> 78.29 | Improved |
| add second-order branches to first-order (ablation) | MVProbe · Model Jungle (ResNet) | accuracy: 91.24 -> 92.24 | Improved |
| add first-order branches to second-order (ablation) | MVProbe · Model Jungle (SupViT) | accuracy: 92.04 -> 92.33 | Improved |
| add second-order over first-order only (ablation) | MVProbe · Model Jungle (DINO) | accuracy: 76.01 -> 75.82 | Hurt |
| add per-sample standardization of branch features (ablation) | MVProbe · Model Jungle (all layers, average) | accuracy: 65.9 -> 68.8 | Improved |
| add per-sample standardization of branch features (ablation) | MVProbe · Model Jungle (MAE, all layers) | accuracy: 62.1 -> 62.9 | Improved |
| add per-sample standardization of branch features (ablation) | MVProbe · Model Jungle (ResNet, middle layers stage 2) | accuracy: 73.6 -> 78.6 | Improved |
| add third-order branches (6 branches) (ablation) | MVProbe · Model Jungle (DINO) | accuracy: 78.29 -> 77.02 | Hurt |
| add third-order branches (6 branches) (ablation) | MVProbe · Model Jungle (ResNet) | accuracy: 92.24 -> 93.15 | Improved |
| add fourth-order branches (8 branches) (ablation) | MVProbe · Model Jungle (MAE) | accuracy: 81.62 -> 81.16 | Mixed |
| quadruple ProbeX probes (x4) | ProbeX · Model Jungle (MAE) | accuracy: 77.11 -> 77.26 | No clear effect |
Closest papers by meaning
- Deep Linear Probe Generators for Weight Space Learning (ICLR 2025 · similarity 0.87)
- Learning on Model Weights using Tree Experts (CVPR 2025 · similarity 0.86)
- Fantastic Features and Where to Find Them: A Probing Method to combine Features from Multiple Foundation Models (NeurIPS 2025 · similarity 0.86)
- Deep Minds and Shallow Probes (NeurIPS 2026 · similarity 0.84)
- RAPTOR: Ridge-Adaptive Logistic Probes (NeurIPS 2026 · similarity 0.83)
- Detecting High-Stakes Interactions with Activation Probes (NeurIPS 2025 · similarity 0.83)
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.
