Research map / Detection and segmentation / CNN features
Formalizing the Binding Problem
Lianghuan Huang, Yihao Li, Saeed Salehi, Yingshan Chang, Ansh Soni, Konrad Kording
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 |
|---|---|---|---|
| use full spatial tokens with attention probe instead of [CLS] token (ablation) | DINOv2-Large probes · ColorShape (test) | normalized binding information: 48.5% -> 92.2% | Improved |
| use quadratic instead of linear probe on [CLS] (ablation) | DINOv2-Large [CLS] object probe · ColorShape (test) | error reduction: 37.4% -> 63.0% | Improved |
| use DNN instead of quadratic probe on [CLS] (ablation) | DINOv2-Large [CLS] object probe · ColorShape (test) | error reduction: 63.0% -> 64.4% | No clear effect |
| share color and shape weights across object probes (parameter reuse) (ablation) | quadratic object probe on [CLS] · ColorShape (test) | error reduction: 63.0% -> 57.7% | Hurt |
| increase number of colors and shapes from 1 to 7 (ablation) | CLIP ViT-L/14 (224px) · ColorShape | beta*_O,F | Hurt |
| reduce occlusion by raising camera elevation (ablation) | DINOv2-Large · CLEVR (4 colors, 3 shapes) | beta*_O,F: 45.0% -> 58.7% | Improved |
| increase CLIP input resolution from 224px to 336px (ablation) | CLIP ViT-L/14 · ColorShape (7x7) | beta*_O,F: 47.7% -> 56.4% | Improved |
| natural features (Visual Genome) instead of synthetic datasets | ViT models · VG:Color, VG:TopAttr | beta*_O,F: DINOv2 ColorShape 41.8% vs VG:Color 41.8% | No clear effect |
Closest papers by meaning
- Does Object Binding Naturally Emerge in Large Pretrained Vision Transformers? (NeurIPS 2025 · similarity 0.91)
- Understanding the Limits of Vision Language Models Through the Lens of the Binding Problem (NeurIPS 2024 · similarity 0.89)
- Beyond Bag-of-Words: Diagnosing Compositional Binding Failures in Vision-Language Models (NeurIPS 2026 · similarity 0.89)
- Visual symbolic mechanisms: Emergent symbol processing in Vision Language Models (ICLR 2026 · similarity 0.88)
- Binding Visual Features Point by Point (NeurIPS 2026 · similarity 0.87)
- Register and [CLS] tokens induce a decoupling of local and global features in large ViTs (NeurIPS 2025 · similarity 0.87)
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.
