Research map / Vision-language and multimodal / CLIP and zero-shot
Improving Vision and Language Concepts Understanding with Multimodal Counterfactual Samples
Chengen Lai, Shengli Song, Sitong Yan, Guangneng Hu
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 |
|---|---|---|---|
| put counterfactual samples in the same mini-batch as factuals instead of random sampling (ablation) | COMO · VL-Checklist average | accuracy (%): 70.95 -> 76.93 | Improved |
| fine-tune on factual COCO only (ablation) | CLIP · Winoground / VL-Checklist | average score: 70.67 -> 70.84 (VL-Checklist avg) | No clear effect |
| add counterfactual text to counterfactual image training (ablation) | COMO · VL-Checklist average | accuracy (%): 75.07 -> 76.93 | Improved |
| add counterfactual image to counterfactual text training (ablation) | COMO · VL-Checklist average | accuracy (%): 74.80 -> 76.93 | Improved |
| use weighted hard-negative contrastive loss instead of unweighted (ablation) | COMO pretrained · VL-Checklist average | accuracy (%): 76.08 -> 76.93 | Improved |
| use CLIP pretrained initialization instead of training from scratch (ablation) | COMO · VL-Checklist average | accuracy (%): 68.07 -> 76.93 | Improved |
| increase number of generated images per caption up to about 8 (ablation) | COMO · VL-Checklist | accuracy | Improved |
| train COMO on 10% of data in low-resource setting (ablation) | COMO · VL-Checklist | relative improvement over CLIP: 2.24% at 10% training | Improved |
| COMO fine-tuning vs CLIP on 21 zero-shot classification | COMO · 21 zero-shot classification datasets | average accuracy (%): 56.37 -> 55.87 | Hurt |
Closest papers by meaning
- CF-VLM:CounterFactual Vision-Language Fine-tuning (NeurIPS 2025 · similarity 0.90)
- LoMo: Local Modality Substitution for Deeper Vision-Language Fusion (NeurIPS 2026 · similarity 0.89)
- Seeing the Image: Prioritizing Visual Correlation by Contrastive Alignment (NeurIPS 2024 · similarity 0.88)
- Scaling Test-Time Robustness of Vision-Language Models via Self-Critical Inference Framework (CVPR 2026 · similarity 0.88)
- Contrasting Intra-Modal and Ranking Cross-Modal Hard Negatives to Enhance Visio-Linguistic Compositional Understanding (CVPR 2024 · similarity 0.88)
- Remodeling Semantic Relationships in Vision-Language Fine-Tuning (AAAI 2026 · similarity 0.88)
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.
