Research map / Vision-language and multimodal / CLIP and zero-shot
What CLIP Knows but Cannot Say: Recovering Negation from Frozen Intermediate Features
Chen-yi Lu, Yueh-Shao Chen, Somali Chaterji
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 PeakPatch (ECN+SCN) correction on frozen CLIP | CLIP ViT-B/32 · NegBench COCO MCQ | accuracy: 39.2 -> 74.3 | Improved |
| add PeakPatch to CLIP | CLIP ViT-B/32 · NegBench VOC MCQ | accuracy: 37.9 -> 65.5 | Improved |
| add ECN on top of SCN-only (joint training) (ablation) | PeakPatch · NegBench COCO MCQ | accuracy: 71.6 -> 74.3 | Improved |
| add SCN to ECN (ablation) | PeakPatch · NegBench COCO MCQ | accuracy: 51.2 -> 74.3 | Improved |
| add anchor layer l_a (ablation) | PeakPatch · NegBench COCO MCQ | accuracy: 68.7 -> 74.3 | Improved |
| let SCN gradients flow into ECN (ablation) | PeakPatch · NegBench COCO MCQ | accuracy: 67.1 -> 74.3 | Improved |
| add SCN to ECN for retrieval (ablation) | PeakPatch · NegBench COCO retrieval | Recall@5: 48.1 -> 64.3 | Improved |
| choose peak layer l_p=8 (LCD maximum) rather than earlier or later (ablation) | PeakPatch · NegBench COCO MCQ | accuracy | Improved |
| vary anchor layer l_a in {2,4,6} (ablation) | PeakPatch · NegBench COCO MCQ | accuracy: <1 pp | No clear effect |
| apply PeakPatch to ViT-L/14 and SigLIP | PeakPatch · NegBench COCO MCQ | accuracy: 40.6 -> 63.8 (ViT-L/14); 28.9 -> 66.4 (SigLIP) | Improved |
| use ECN-corrected embeddings in text-to-image generation | GALIP · negated prompts | Neg score: 29.2 -> 47.6 | Improved |
| use PeakPatch instead of encoder fine-tuning (in-domain retrieval) | PeakPatch · NegBench COCO retrieval | Recall@1: 41.3 -> 37.1 | Hurt |
Closest papers by meaning
- Post-pre-training for Modality Alignment in Vision-Language Foundation Models (CVPR 2025 · similarity 0.90)
- Rethinking Contrastive Loss in CLIP Post-training: A Complementary Framework with Frozen Text Encoder (NeurIPS 2026 · similarity 0.90)
- Enhancing Compositional Reasoning in CLIP via Reconstruction and Alignment of Text Descriptions (NeurIPS 2025 · similarity 0.89)
- VITRIX-CLIPIN: Enhancing Fine-Grained Visual Understanding in CLIP via Instruction-Editing Data and Long Captions (NeurIPS 2025 · similarity 0.89)
- Unlocking Fine-Grained Perception in CLIP via Structurally-Aware Latent Masked Modeling (NeurIPS 2026 · similarity 0.89)
- un$^2$CLIP: Improving CLIP's Visual Detail Capturing Ability via Inverting unCLIP (NeurIPS 2025 · similarity 0.89)
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.
