Research map / Vision-language and multimodal / CLIP and zero-shot
Beyond Heuristic Prompting: A Concept-Guided Bayesian Framework for Zero-Shot Image Recognition
Hui Liu, Kecheng Chen, Jialiang Wang, Xianming Liu, Wenya Wang, Haoliang Li
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 adaptive soft-trim likelihood (ablation) | CGBC Prior (50 prompts) · 11 datasets avg | top-1 accuracy: 68.56 -> 69.25 | Improved |
| add adaptive soft-trim likelihood (ablation) | CGBC Prior (16 prompts) · 11 datasets avg | top-1 accuracy: 68.35 -> 68.95 | Improved |
| use sigmoid soft-trim instead of hard-trim (ablation) | CGBC · 11 datasets avg, 50 prompts | top-1 accuracy: 69.07 -> 69.25 | Improved |
| use sigmoid soft-trim instead of confidence-based selection (ablation) | CGBC · 11 datasets avg, 50 prompts | top-1 accuracy: 68.61 -> 69.25 | Improved |
| compose atomic concepts with 'or' instead of embedding averaging (ablation) | CGBC · 11 datasets avg, 50 prompts | top-1 accuracy: 68.2 -> 69.3 | Improved |
| use larger concept generator (Turbo vs Mini vs Nano) (ablation) | CGBC · 5-dataset avg | top-1 accuracy: 59.88 -> 60.17 -> 60.99 | Improved |
| increase hard-negative neighborhood from 5 to 20 (ablation) | CGBC (GPT-4.1 Mini) · 5-dataset avg | top-1 accuracy: 59.84 -> 60.32 | Improved |
| increase hard-negative neighborhood from 20 to 30 (ablation) | CGBC (GPT-4.1 Mini) · 5-dataset avg | top-1 accuracy: 60.32 -> 60.04 | Hurt |
| raise outlier threshold lambda from 1.5 to 2.5 (ablation) | CGBC · 11 datasets avg | top-1 accuracy: 69.07 -> 69.25 | Mixed |
| use CLIP logit scale for sigmoid slope k instead of 0.25 (ablation) | CGBC · 11 datasets avg | top-1 accuracy: 69.11 -> 69.25 | No clear effect |
| swap GPT-4.1 turbo for Gemini 2.5 Flash Lite concept generator (ablation) | CGBC · 5-dataset avg | top-1 accuracy: 61.0 -> 59.9 | No clear effect |
| add multi-view entropy-filtered augmentation (CGBC + View) | CGBC (1, 50) · ImageNet variants OOD average | top-1 accuracy: 58.95 -> 63.36 | Improved |
Closest papers by meaning
- Visual and Semantic Prompt Collaboration for Generalized Zero-Shot Learning (CVPR 2025 · similarity 0.89)
- Understanding prompt engineering may not require rethinking generalization (ICLR 2024 · similarity 0.89)
- TPR: Topology-Preserving Reservoirs for Generalized Zero-Shot Learning (NeurIPS 2024 · similarity 0.89)
- Interpreting and Analysing CLIP's Zero-Shot Image Classification via Mutual Knowledge (NeurIPS 2024 · similarity 0.89)
- PromptKD: Unsupervised Prompt Distillation for Vision-Language Models (CVPR 2024 · similarity 0.88)
- Explanatory Instructions: Towards Unified Vision Tasks Understanding and Zero-shot Generalization (ICML 2025 · 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.
