Research map / Learning theory and algorithms / Sample complexity

Certified Robust Interpretability via Concept-Space Stability under Interventional Proxies

NeurIPS 2026 · Learning theory and algorithms › Sample complexity · approach: Learning halfspaces under noise

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's abstract only. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a certification framework for concept-space stability (explanation robustness) in concept-bottleneck classifiers, showing that label-only certifiers cannot guarantee it.

Problem

Concept-bottleneck classifiers can keep their label and pixel-space l2 robustness while their concept-level explanation changes under small perturbations; existing certifiers that only see input and label cannot detect this.

Approach

Certify stability of the concept-level attribution (which concepts drive the decision) under interventional proxies, rather than only the predicted label, and show a label-only certifier admits arbitrarily large concept shifts.

Key results

• Abstract states that any certifier seeing only the input and label distribution admits classifiers with arbitrarily large concept-space shifts, so label-only certification is insufficient.

Surprising findings

• Label-level and pixel-space l2 robustness do not imply stability of the concept-level explanation; arbitrarily large concept rewiring is possible under label-only certification.

Limitations

• Only the abstract was available; experimental scope and empirical validation are not described.

Experimental setup

Builds on: randomized smoothing concept bottleneck models

Closest papers by meaning

• Interpretable but Fragile? Robustness of Concept Bottlenecks under Geometric-Semantic Perturbations (NeurIPS 2026 · similarity 0.86)
• Sample-efficient Learning of Concepts with Theoretical Guarantees: from Data to Concepts without Interventions (NeurIPS 2025 · similarity 0.86)
• Probabilistic Stability Guarantees for Feature Attributions (NeurIPS 2025 · similarity 0.85)
• An Analysis of Concept Bottleneck Models: Measuring, Understanding, and Mitigating the Impact of Noisy Annotations (NeurIPS 2025 · similarity 0.85)
• Pixel-level Certified Explanations via Randomized Smoothing (ICML 2025 · similarity 0.84)
• FACE: Faithful Automatic Concept Extraction (NeurIPS 2025 · similarity 0.84)