Research map / Vision-language and multimodal / CLIP and zero-shot
Similarity Is Not Logic: Factored Inference for Dual-Encoder Vision-Language Models
Sultan Alshehri, Zhantao Yang, Han Zhang, Marios Savvides
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 LCSE factored score editing | CLIP ViT-B/32 · FACTOR-Bench | accuracy: 58.3 -> 85.5 | Improved |
| add LCSE | SigLIP 2 · FACTOR-Bench | accuracy: 65.0 -> 90.7 | Improved |
| add LCSE | SigLIP 2 · NegBench COCO MCQ | accuracy: 27.2 -> 65.2 | Improved |
| add LCSE | CLIP · COCO retrieval | R@5: 55.3 -> 55.2 | No clear effect |
| stack LCSE on NegCLIP | NegCLIP · FACTOR-Bench | accuracy: 67.2 -> 86.2 | Improved |
| stack LCSE on NegCLIP | NegCLIP · NegBench COCO MCQ | accuracy: 28.7 -> 55.7 | Improved |
| stack LCSE on NegCLIP | NegCLIP · NegBench COCO R@5 | R@5: 64.4 -> 65.0 | No clear effect |
| fine-tune CLIP on negation data (NegationCLIP, CoN-CLIP etc.) | CLIP · COCO retrieval retention | R@5: 55.3 -> 52.3 | Hurt |
| fine-tune on negation data | CLIP variants · polarity test, no X | AUC: 11.2% -> 19.0% | No clear effect |
| apply negation externally as 1-p (factored) | CLIP · polarity test, no X | AUC: 11.2% -> 88.7% | Improved |
| amplify residual component lambda (ablation) | CLIP not-A · COCO val2017 concept pairs | constraint accuracy: 12% -> 66% | Improved |
| amplify residual component lambda (ablation) | CLIP NOR · COCO val2017 concept pairs | constraint accuracy | No clear effect |
Closest papers by meaning
- Revisiting the Role of Language Priors in Vision-Language Models (ICML 2024 · similarity 0.88)
- The Neglected Tails in Vision-Language Models (CVPR 2024 · similarity 0.88)
- Explaining Similarity in Vision-Language Encoders with Weighted Banzhaf Interactions (NeurIPS 2025 · similarity 0.87)
- The Many Senses of Visual Similarity: A Text-Prompted Image Perceptual Metric (NeurIPS 2026 · similarity 0.87)
- Zero-shot Referring Expression Comprehension via Structural Similarity Between Images and Captions (CVPR 2024 · similarity 0.87)
- Natural Language Inference Improves Compositionality in Vision-Language Models (ICLR 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.
