Research map / Image generation / Autoregressive and latent generation
SUB: Benchmarking CBM Generalization via Synthetic Attribute Substitutions
Jessica Bader, Leander Girrbach, Stephan Alaniz, Zeynep Akata
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 |
|---|---|---|---|
| evaluate CBMs on novel single-attribute substitutions (SUB) instead of CUB test classes | Concept Bottleneck Models · SUB vs CUB | concept accuracy (%): 96.7 (CUB test) -> 40.8 (SUB S+), independent CBM | Hurt |
| use Concept Embedding Model instead of standard CBM | Concept Bottleneck Models · SUB (S+ accuracy) | S+ accuracy (%): 40.8 -> 45.7 | Improved |
| use soft concept labels instead of hard labels (ablation) | CBM (independent) · SUB (S+ accuracy) | S+ accuracy (%): 40.8 -> 12.9 | Hurt |
| use per-image concept labels instead of class-level attribute vectors (ablation) | CBM (independent) · SUB (S+ accuracy) | S+ accuracy (%): 40.8 -> 32.8 | Hurt |
| train concepts individually (per-concept CBM) (ablation) | CBM · SUB (S+ accuracy) | S+ accuracy (%): 40.8 -> 0.39 | Hurt |
| train CBM jointly instead of independently (ablation) | CBM · SUB (S+ accuracy) | S+ accuracy (%): 40.8 -> 34.3 | Hurt |
| scale up to large pretrained VLM backbones (CLIP, SigLIP, EVA-CLIP) | Concept recognition · SUB VLMs (S+ accuracy) | S+ accuracy (%): 9.3 (random) -> 46.8 (EVA-CLIP) vs human 92.4 | Mixed |
| add Tied Diffusion Guidance to FLUX prompting for attribute substitution (ablation) | FLUX.1-dev · SUB generation | qualitative attribute fidelity | Improved |
| use SUB human-validated labels instead of CUB image-level labels | Benchmark annotation · Label correctness | attribute label agreement (%): 57.50 -> 98.90 | Improved |
Closest papers by meaning
- VLG-CBM: Training Concept Bottleneck Models with Vision-Language Guidance (NeurIPS 2024 · similarity 0.88)
- Bayesian Concept Bottleneck Models with LLM Priors (NeurIPS 2025 · similarity 0.87)
- Stochastic Concept Bottleneck Models (NeurIPS 2024 · similarity 0.87)
- Beyond Concept Bottleneck Models: How to Make Black Boxes Intervenable? (NeurIPS 2024 · similarity 0.87)
- Energy-Based Concept Bottleneck Models: Unifying Prediction, Concept Intervention, and Probabilistic Interpretations (ICLR 2024 · similarity 0.87)
- DCBM: Data-Efficient Visual Concept Bottleneck Models (ICML 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.
