Research map / Label-efficient and robust learning / Recognition and long-tail
On the Faithfulness of Post-Hoc Concept Bottleneck Models
Laines Schmalwasser, Jan Blunk, Niklas Penzel, Julia Niebling, Joachim Denzler
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 |
|---|---|---|---|
| train the concept projection on out-of-distribution data instead of in-domain data (ablation) | PCBM · Elements (post-hoc CBM) | downstream accuracy with oracle h*: 1.00 -> 0.21 | Hurt |
| train the concept projection on out-of-distribution data instead of in-domain data (ablation) | PCBM · Elements (post-hoc CBM) | concept direction cosine similarity: 0.96 -> 0.47 | Hurt |
| train the concept projection on out-of-distribution data instead of in-domain data (ablation) | PCBM · Elements (post-hoc CBM) | downstream accuracy with learned h: 1.00 -> 1.00 | No clear effect |
| train on near-domain P_near^forall instead of in-domain data (ablation) | PCBM · Elements (post-hoc CBM) | concept accuracy on P_task: 1.00 -> 0.70 | Hurt |
| use CLIP surrogate labels instead of 25% random-noise labels (ablation) | LFCBM · Elements (post-hoc CBM) | concept accuracy on P_task: 0.98 -> 0.56 | Hurt |
| use DINOv3 instead of CLIP for surrogate labels | LFCBM · Elements (post-hoc CBM) | downstream accuracy with oracle h*: 0.54 -> 0.78 | Improved |
| use Grounding DINO surrogate labels instead of 25% noise (ablation) | VLG-CBM · Elements (post-hoc CBM) | downstream accuracy with oracle h*: 0.98 -> 0.04 | Hurt |
| use learned classifier h instead of oracle h* | LFCBM with CLIP labels · Elements (post-hoc CBM) | downstream accuracy: 0.54 -> 0.99 | Mixed |
| increase number of random concept projections K (ablation) | post-hoc CBM with random projections · Elements (post-hoc CBM) | downstream accuracy (CIFAR-100) | Improved |
| replace random projections with meaningful concept sources | post-hoc CBM · Elements (post-hoc CBM) | downstream accuracy (CIFAR-100) | No clear effect |
| use Grounding DINO instead of CLIP or DINOv3 for surrogate labels | surrogate labeling · Elements (post-hoc CBM) | mean absolute error-activation correlation: 0.152 -> 0.142 | Improved |
Closest papers by meaning
- Sample-efficient Learning of Concepts with Theoretical Guarantees: from Data to Concepts without Interventions (NeurIPS 2025 · similarity 0.90)
- An Analysis of Concept Bottleneck Models: Measuring, Understanding, and Mitigating the Impact of Noisy Annotations (NeurIPS 2025 · similarity 0.90)
- Interpretable but Fragile? Robustness of Concept Bottlenecks under Geometric-Semantic Perturbations (NeurIPS 2026 · similarity 0.89)
- Coarse-to-Fine Concept Bottleneck Models (NeurIPS 2024 · similarity 0.89)
- Energy-Based Concept Bottleneck Models: Unifying Prediction, Concept Intervention, and Probabilistic Interpretations (ICLR 2024 · similarity 0.89)
- Hyperbolic Concept Bottleneck Models (NeurIPS 2026 · similarity 0.89)
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.
