Research map / Vision-language and multimodal / CLIP and zero-shot
ChatGPT-Powered Hierarchical Comparisons for Image Classification
Zhiyuan Ren, Yiyang Su, Xiaoming Liu
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 |
|---|---|---|---|
| replace class-name CLIP prompts with hierarchical LLM comparative descriptions | CLIP ViT-L/14 · Describable Textures zero-shot | accuracy: 51.01 -> 58.04 | Improved |
| replace class-name CLIP prompts with hierarchical descriptions | CLIP ResNet-50 · ImageNet zero-shot | accuracy: 54.85 -> 60.63 | Improved |
| replace class-name OpenCLIP prompts with hierarchical descriptions | OpenCLIP ViT-L/14 · Describable Textures zero-shot | accuracy: 51.69 -> 61.79 | Improved |
| replace class-name OpenCLIP prompts with hierarchical descriptions | OpenCLIP · Food101 zero-shot | accuracy: 93.41 -> 92.68 | Hurt |
| replace class-name CLIP prompts with hierarchical descriptions | CLIP ViT-L/14 · CUB zero-shot | accuracy: 63.46 -> 65.45 | Improved |
| replace non-comparative LLM descriptions (classification by description) with hierarchical comparative descriptions | CLIP ViT-L/14 · ImageNet zero-shot | accuracy: 75.36 -> 79.64 | Improved |
| replace non-comparative LLM descriptions with hierarchical comparative descriptions | CLIP ViT-L/14 · mean over six datasets | accuracy: 70.26 -> 71.71 | Improved |
| set score momentum lambda to 0 (no offset) (ablation) | hierarchical scoring · mean over six datasets, ViT-B/32 | accuracy | Hurt |
| use larger group number N (fewer hierarchy levels, ~sixfold speedup) (ablation) | knowledge tree · ImageNet ViT-B/32 | accuracy: 1.64% dip | Hurt |
| increase depth of descriptions used in prediction (ablation) | hierarchical scoring · ImageNet ViT-B/32 | accuracy | Improved |
| vary score reduction tolerance tau from 0 to 1 (ablation) | hierarchical scoring · mean over six datasets, ViT-B/32 | accuracy | No clear effect |
Closest papers by meaning
- Interpreting and Analysing CLIP's Zero-Shot Image Classification via Mutual Knowledge (NeurIPS 2024 · similarity 0.89)
- Follow-Up Differential Descriptions: Language Models Resolve Ambiguities for Image Classification (ICLR 2024 · similarity 0.89)
- Large Language Models are Good Prompt Learners for Low-Shot Image Classification (CVPR 2024 · similarity 0.87)
- Meta CLIP 2: A Worldwide Scaling Recipe (NeurIPS 2025 · similarity 0.87)
- Modeling Caption Diversity in Contrastive Vision-Language Pretraining (ICML 2024 · similarity 0.87)
- SuperCLIP: CLIP with Simple Classification Supervision (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.
