Research map / Label-efficient and robust learning / Recognition and long-tail
Addressing Concept Mislabeling in Concept Bottleneck Models Through Preference Optimization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Concept Preference Optimization (CPO), a DPO-derived loss for jointly trained concept bottleneck models that makes them robust to concept mislabeling and noisy concept labels.
Problem
Concept Bottleneck Models assume accurate concept labels, but real datasets and concept-destroying augmentations introduce concept mislabeling that substantially degrades task accuracy, concept AUC and intervention performance.
Approach
Replace the BCE concept loss with a DPO-style preference objective in which empirical concept labels are preferred over concepts sampled from the model's own policy; this yields a loss proportional to BCE only on mismatched samples, is shown to approximate the concept posterior, and has smaller gradient deviation under noise.
Key results
- Under concept label noise, CBM CPO keeps higher task accuracy than CBM BCE, e.g. CUB at p=0.2 reaches 0.793 vs 0.719 for BCE.
- Without added noise, CBM CPO improves CUB task accuracy to 0.800 vs 0.753 for CBM BCE and AwA2 concept AUC 0.971 vs 0.959.
- Confidence-based structured noise on CUB yields CBM CPO task accuracy 0.793 vs 0.733 for CBM BCE.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace BCE with CPO for CBM training | CBM · CUB | task accuracy: 0.753 -> 0.800 | Helped |
| Replace BCE with CPO for CBM training | CBM · CUB | concept AUC: 0.937 -> 0.952 | Helped |
| Replace BCE with CPO for CBM training | CBM · AwA2 | task accuracy: 0.900 -> 0.915 | Helped |
| Replace BCE with CPO for CBM training | CBM · AwA2 | concept AUC: 0.959 -> 0.971 | Helped |
| Replace BCE with CPO for CBM training | CBM · CelebA | task accuracy: 0.283 -> 0.310 | Helped |
| Replace BCE with CPO for CBM training | CBM · CelebA | concept AUC: 0.873 -> 0.857 | Hurt |
| Replace BCE with CPO for CBM training | CBM · CUB with labeler-confidence noise | task accuracy: 0.733 -> 0.793 | Helped |
| Replace BCE with CPO for CBM training | CBM · CUB with labeler-confidence noise | concept AUC: 0.876 -> 0.913 | Helped |
| Replace BCE with CPO for CBM training | CBM · CUB with group noise p=0.1 | task accuracy: 0.741 -> 0.796 | Helped |
| Replace BCE with CPO for CBM training | CBM · CUB with group noise p=0.1 | concept AUC: 0.914 -> 0.941 | Helped |
| Replace BCE with CPO for CBM training | CBM · AwA2 with group noise p=0.1 | task accuracy: 0.879 -> 0.894 | Helped |
| Replace BCE with CPO for CBM training | CBM · AwA2 with group noise p=0.1 | concept AUC: 0.922 -> 0.943 | Helped |
Surprising findings
- CEMs trained with BCE are more resilient to noise than plain CBMs trained with BCE, and CPO-trained CBMs can outperform BCE-trained CEMs.
- Under concept-group structured noise, the paper's own appendix text contains a contradictory statement that CPO is less robust than BCE, while the reported results favor CPO.
Limitations
- Concept noise is simulated by random flipping in most experiments rather than from real annotation errors.
- Theoretical analysis relies on assumptions such as binary concepts, conditional independence, and uniform prior.
- Gains are mixed on some settings, e.g. CelebA concept AUC and CEM results without noise.
Experimental setup
Datasets: CUB AwA2 CelebA
Benchmarks: CUB AwA2 CelebA
Metrics: task accuracy concept AUC intervention accuracy
Baselines: CBM BCE CEM BCE ProbCBM CoopCBM
Builds on: Direct Preference Optimization Concept Bottleneck Models Concept Embedding Models Probabilistic CBMs
Closest papers by meaning
- Measuring What Matters: Synthetic Benchmarks for Concept Bottleneck Models (NeurIPS 2026 · similarity 0.88)
- An Analysis of Concept Bottleneck Models: Measuring, Understanding, and Mitigating the Impact of Noisy Annotations (NeurIPS 2025 · similarity 0.88)
- Sample-efficient Learning of Concepts with Theoretical Guarantees: from Data to Concepts without Interventions (NeurIPS 2025 · similarity 0.87)
- A theoretical design of concept sets: improving the predictability of concept bottleneck models (NeurIPS 2024 · similarity 0.87)
- Energy-Based Concept Bottleneck Models: Unifying Prediction, Concept Intervention, and Probabilistic Interpretations (ICLR 2024 · similarity 0.87)
- Interpretable but Fragile? Robustness of Concept Bottlenecks under Geometric-Semantic Perturbations (NeurIPS 2026 · similarity 0.86)
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.