Research map / Vision-language and multimodal / Multimodal LLMs
Know "No" Better: A Data-Driven Approach for Enhancing Negation Awareness in CLIP
Junsung Park, Jungbeom Lee, Jongyoon Song, Sangwon Yu, Dahuin Jung, Sungroh Yoon
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 plausible objects (P1) instead of random objects (Rand-P1) (ablation) | CLIP ViT-B/32 fine-tuning · VALSE (existence) | accuracy: 73.78 -> 80.15 | Improved |
| use plausible objects (P1) instead of random objects (Rand-P1) (ablation) | CLIP ViT-B/16 fine-tuning · VALSE (existence) | accuracy: 76.22 -> 77.53 | Improved |
| use plausible objects (P1) instead of random objects (Rand-P1) (ablation) | CLIP ViT-L/14@336px fine-tuning · VALSE (existence) | accuracy: 76.22 -> 79.78 | Improved |
| add VQA-derived pipeline P2 to P1 (ablation) | CLIP ViT-B/32 fine-tuning · VALSE (existence) | accuracy: 80.15 -> 80.15 | No clear effect |
| add VQA-derived pipeline P2 to P1 (ablation) | CLIP ViT-B/32 fine-tuning · NegRefCOCOg | accuracy: 63.18 -> 64.09 | Improved |
| add VQA-derived pipeline P2 to P1 (ablation) | CLIP ViT-B/16 fine-tuning · VALSE (existence) | accuracy: 77.53 -> 80.52 | Improved |
| add VQA-derived pipeline P2 to P1 (ablation) | CLIP ViT-L/14 fine-tuning · NegRefCOCOg | accuracy: 60.23 -> 62.95 | Improved |
| fine-tune with generated captions instead of original captions (ablation) | CLIP ViT-B/32 · VALSE (existence) | accuracy: 68.35 -> 80.15 | Improved |
| fine-tune with original captions (ablation) | CLIP ViT-B/32 · VALSE (existence) | accuracy: 70.97 -> 68.35 | Hurt |
| fine-tune with generated captions instead of original captions (ablation) | CLIP ViT-L/14@336px · VALSE (existence) | accuracy: 73.97 -> 80.34 | Improved |
| swap SD-1.4 text encoder for NegationCLIP (no retraining) | SD-1.4 · Text-to-image negation prompts (107 prompts) | Neg Score: 0.295 -> 0.449 | Improved |
| swap SDXL-1.0 text encoders for NegationCLIP | SDXL-1.0 · TIFA | TIFA: 0.849 -> 0.802 | Hurt |
Closest papers by meaning
- Vision-Language Models Do Not Understand Negation (CVPR 2025 · similarity 0.90)
- What "Not" to Detect: Negation-Aware VLMs via Structured Reasoning and Token Merging (ICLR 2026 · similarity 0.88)
- UniLiP: Adapting CLIP for Unified Multimodal Understanding, Generation and Editing (ICLR 2026 · similarity 0.88)
- un$^2$CLIP: Improving CLIP's Visual Detail Capturing Ability via Inverting unCLIP (NeurIPS 2025 · similarity 0.87)
- TripletCLIP: Improving Compositional Reasoning of CLIP via Synthetic Vision-Language Negatives (NeurIPS 2024 · similarity 0.86)
- VITRIX-CLIPIN: Enhancing Fine-Grained Visual Understanding in CLIP via Instruction-Editing Data and Long Captions (NeurIPS 2025 · similarity 0.86)
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.
