Research map / Label-efficient and robust learning / OOD and anomaly detection
Avoiding Leakage Poisoning: Concept Interventions Under Distribution Shifts
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.
Shows that leakage bypasses in state-of-the-art concept models cause "leakage poisoning" under OOD inputs and proposes MixCEM, which gates residual concept embeddings by concept entropy to keep interventions effective for both ID and OOD samples.
Problem
Concept-based models with leakage bypasses (e.g., CEMs) fail to improve task accuracy when concepts are intervened on for out-of-distribution inputs, undermining the usefulness of human-in-the-loop corrections under distribution shift.
Approach
MixCEM mixes a global, sample-agnostic concept embedding with a contextual residual embedding, scaling the residual by (1 - entropy of the concept prediction) so leakage is suppressed when inputs are OOD, and adds a prior-error loss and random dropout of residuals.
Key results
- On AwA2-Incomplete, MixCEM's OOD intervention accuracy improves by up to approximately 48% over CEM and 41% over IntCEM.
- MixCEM is the only completeness-agnostic baseline whose OOD intervention curves remain near or above the Bayes classifier in most tasks (all except CIFAR10 and a short segment of CelebA).
- MixCEM attains task accuracy competitive with the best CM in several tasks, e.g., 76.54% on CUB vs CEM 76.67%, and 74.54% on CUB-Incomplete vs CEM 74.42%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MixCEM instead of CEM | Concept-based models · OOD concept interventions across tasks | task accuracy: up to approx. 48% improvement in AwA2-Incomplete over CEM | Helped |
| Use MixCEM instead of IntCEM | Concept-based models · OOD concept interventions across tasks | task accuracy: up to approx. 41% improvement in AwA2-Incomplete over IntCEM | Helped |
| Use MixCEM instead of competing baselines | Concept-based models · ID concept interventions | task accuracy | Helped |
| Use MixCEM instead of CEM | CUB-Incomplete · CUB-Incomplete, 5% test noise, no interventions | task accuracy: 24.08% vs 29.69% | Helped |
| Use MixCEM instead of IntCEM | TravelingBirds-Incomplete · TravelingBirds-Incomplete, OOD test without spurious correlation, all concepts intervened | task accuracy: more than 10% points | Helped |
| Use MixCEM instead of ProbCBM | TravelingBirds · TravelingBirds, OOD test without spurious correlation, no interventions | task accuracy | Hurt |
Surprising findings
- CEM's accuracy with all concepts intervened on is significantly worse for OOD samples than for ID samples, which is not observed for Vanilla CBMs.
- MixCEM remains highly intervenable even when the concept loss weight is zero, provided the prior-error weight is positive.
Limitations
- MixCEM requires more parameters (O(km) for global embeddings) and more hyperparameters than CEMs.
- When a concept goes OOD to the point an expert cannot interpret it, MixCEM may not fully block leakage poisoning.
- In incomplete tasks, intervened MixCEMs do not always recover full ID performance on OOD inputs.
Experimental setup
Datasets: CUB CUB-Incomplete AwA2 AwA2-Incomplete CelebA CIFAR10 TravelingBirds
Benchmarks: CUB CUB-Incomplete AwA2 AwA2-Incomplete CelebA CIFAR10 TravelingBirds
Metrics: task accuracy mean concept ROC-AUC intervention accuracy curves
Baselines: DNN Vanilla CBM Hybrid CBM ProbCBM P-CBM Residual P-CBM CEM IntCEM
Builds on: Concept Embedding Models Concept Bottleneck Models Intervention-aware CEMs Platt scaling Monte Carlo Dropout
Closest papers by meaning
- An Analysis of Concept Bottleneck Models: Measuring, Understanding, and Mitigating the Impact of Noisy Annotations (NeurIPS 2025 · similarity 0.84)
- Causally Reliable Concept Bottleneck Models (NeurIPS 2025 · similarity 0.84)
- Learning to Intervene on Concept Bottlenecks (ICML 2024 · similarity 0.84)
- Interpretable Concept-Based Memory Reasoning (NeurIPS 2024 · similarity 0.84)
- FaCT: Faithful Concept Traces for Explaining Neural Network Decisions (NeurIPS 2025 · similarity 0.83)
- A theoretical design of concept sets: improving the predictability of concept bottleneck models (NeurIPS 2024 · similarity 0.83)
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.