Research map / Vision-language and multimodal / CLIP and zero-shot
Flexible Concept Bottleneck Model
Xingbo Du, Qiantong Dou, Lei Fan, Rui Zhang
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 sparsemax instead of hard truncation to 30 concepts (ablation) | FCBM, ViT-L/14 · CIFAR100 zero-shot with DeepSeek-V3 concepts | accuracy (%): 23.61 -> 62.27 | Improved |
| add sparsemax instead of hard truncation to 30 concepts (ablation) | FCBM, ResNet50 · CIFAR10 zero-shot with DeepSeek-V3 concepts | accuracy (%): 54.00 -> 75.32 | Improved |
| add sparsemax instead of hard truncation to 30 concepts (ablation) | FCBM, ViT-L/14 · ImageNet zero-shot with DeepSeek-V3 concepts | accuracy (%): 15.07 -> 51.70 | Improved |
| add learnable temperature instead of default unlearnable temperature (ablation) | FCBM, ViT-L/14 · CIFAR100 zero-shot with DeepSeek-V3 concepts | accuracy (%): 38.54 -> 62.27 | Improved |
| add learnable temperature instead of default unlearnable temperature (ablation) | FCBM, ResNet50 · CIFAR10 trained concepts | accuracy (%): 81.27 -> 85.59 | Improved |
| add learnable temperature instead of default unlearnable temperature (ablation) | FCBM, ResNet50 · CUB trained concepts | accuracy (%): 29.45 -> 63.92 | Improved |
| add sparsemax instead of hard truncation (ablation) | FCBM, ResNet50 · Places365 trained concepts | accuracy (%): 49.85 -> 49.13 | Hurt |
| add sparsemax instead of hard truncation (ablation) | FCBM, ViT-L/14 · CIFAR10 trained concepts | accuracy (%): 97.27 -> 97.21 | No clear effect |
| replace the full concept pool with GPT-4o-generated concepts without fine-tuning (ablation) | FCBM, ViT-L/14 · CIFAR10 | accuracy (%): 97.21 -> 94.19 | Hurt |
| replace the full concept pool with DeepSeek-V3 concepts without fine-tuning (ablation) | FCBM, ResNet50 · CUB | accuracy (%): 63.92 -> 18.07 | Hurt |
| use FCBM hypernetwork-sparsemax instead of linear layer (LF-CBM) | ViT-L/14 backbone · CIFAR100 (NEC about 30) | accuracy (%): 81.98 -> 83.63 | Improved |
| use FCBM instead of LF-CBM | ResNet50 backbone · CUB (NEC about 30) | accuracy (%): 56.91 -> 63.46 | Improved |
Closest papers by meaning
- Hybrid Concept Bottleneck Models (CVPR 2025 · similarity 0.93)
- Concepts Worth Having: Refining VLM-Guided Concept Bottleneck Models with Minimal Annotations (NeurIPS 2026 · similarity 0.92)
- VLG-CBM: Training Concept Bottleneck Models with Vision-Language Guidance (NeurIPS 2024 · similarity 0.92)
- Energy-Based Concept Bottleneck Models: Unifying Prediction, Concept Intervention, and Probabilistic Interpretations (ICLR 2024 · similarity 0.91)
- Coarse-to-Fine Concept Bottleneck Models (NeurIPS 2024 · similarity 0.90)
- Concept-Based Unsupervised Domain Adaptation (ICML 2025 · similarity 0.90)
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.
