Research map / Label-efficient and robust learning / OOD and anomaly detection
CONDA: Adaptive Concept Bottleneck for Foundation Models 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.
Proposes CONDA, a source-free test-time adaptation framework that adapts concept bottleneck models built on foundation models (concept bank, linear predictor, and residual concepts) using only unlabeled target data to handle low-level and concept-level distribution shifts.
Problem
Concept bottleneck models on foundation model features lose accuracy and robustness under test-time distribution shifts, and source data is unavailable for adaptation.
Approach
Three-stage online adaptation of a CBM using only unlabeled test batches and pseudo-labels from an ensemble of zero-shot and linear-probe predictors: concept-score alignment via Mahalanobis class-aware loss on the concept bank, linear probing adaptation with elastic-net sparsity, and a residual concept bottleneck with added concept vectors regularized for diversity and coherency.
Key results
- CONDA improves target-domain AVG accuracy of the post-hoc CBM on CIFAR10-C from 82.61% to 84.38% and boosts Waterbirds AVG from 32.03% to 60.69% for the tested CBM baseline.
- Adaptation boosts post-deployment accuracy by up to 28% and brings CBM performance on par with or above non-interpretable zero-shot and linear probing baselines, notably in WG accuracy.
- Runtime per batch is a few seconds for CSA, LPA, and RCB combined, with pseudo-labeling dominating cost.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply CONDA adaptation to the CBM | CBM on CIFAR10-C · CIFAR10-C target | AVG accuracy: 82.61 ± 1.65 -> 84.56 ± 3.11 | Helped |
| Apply CONDA adaptation to the CBM | CBM on CIFAR10-C · CIFAR10-C target | WG accuracy: 68.62 ± 2.93 -> 72.88 ± 2.01 | Helped |
| Apply CONDA adaptation to the CBM | CBM on CIFAR100-C · CIFAR100-C target | AVG accuracy: 51.53 ± 0.13 -> 54.79 ± 1.17 | Helped |
| Apply CONDA adaptation to the CBM | CBM on CIFAR100-C · CIFAR100-C target | WG accuracy: 2.80 ± 0.71 -> 6.01 ± 0.22; 30.60 ± 1.42 -> 28.44 ± 0.95 | Mixed |
| Apply CONDA adaptation to the CBM | CBM on Waterbirds · Waterbirds target | AVG accuracy: 32.03 ± 0.58 -> 62.71 ± 0.33 | Helped |
| Apply CONDA adaptation to the CBM | CBM on Waterbirds · Waterbirds target | WG accuracy: 27.80 ± 1.24 -> 44.01 ± 0.60 | Helped |
| Apply CONDA adaptation to the CBM | CBM on Metashift · Metashift target | AVG accuracy: 84.45 ± 1.39 -> 93.90 ± 0.13 | Helped |
| Apply CONDA adaptation to the CBM | CBM on Metashift · Metashift target | WG accuracy: 73.89 ± 3.21 -> 91.77 ± 0.12 | Helped |
| Apply CONDA adaptation to the CBM | CBM on Camelyon17 · Camelyon17 target | AVG accuracy: 89.71 ± 0.65 -> 93.16 ± 0.05; 95.01 ± 0.07 -> 92.54 ± 0.16 | Mixed |
| Apply CONDA adaptation to the CBM | CBM on Camelyon17 · Camelyon17 target | WG accuracy: 85.96 ± 0.88 -> 89.00 ± 0.07 | Helped |
| Use CSA alone instead of all CONDA components (ablation) | CONDA on low-level shifts (CIFAR datasets) · CIFAR10-C and CIFAR100-C target | AVG accuracy | Helped |
| Use CSA alone instead of all CONDA components (ablation) | CONDA on Camelyon17 · Camelyon17 target | AVG accuracy | Helped |
Surprising findings
- Concept-based predictions are not inherently more robust than feature-based predictions under distribution shift; unadapted CBMs can fall far below zero-shot and linear probing on concept-level shifts.
- Using only the CSA step gives the best gains under low-level shifts, exceeding the full pipeline, while LPA alone suffices for concept-level shifts.
Limitations
- Pseudo-labeling depends on foundation model robustness; poor zero-shot/LP quality under severe shifts degrades adaptation.
- Lacks theoretical guarantees on concept sufficiency and adaptation bounds.
- Evaluated only on a limited set of shifts and datasets, with severity level two for CIFAR-C instead of standard level five.
Experimental setup
Datasets: CIFAR10 CIFAR100 CIFAR10-C CIFAR100-C Waterbirds Metashift Camelyon17 ConceptNet
Benchmarks: CIFAR10-C CIFAR100-C Waterbirds Metashift Camelyon17
Metrics: AVG (averaged per-class accuracy) WG (worst-group accuracy)
Baselines: Post-hoc CBM PCBM LaBo zero-shot CLIP linear probing
Builds on: concept bottleneck models post-hoc CBM PCBM CLIP-Dissect class-aware feature alignment
Closest papers by meaning
- A theoretical design of concept sets: improving the predictability of concept bottleneck models (NeurIPS 2024 · similarity 0.88)
- Concept-Based Unsupervised Domain Adaptation (ICML 2025 · similarity 0.88)
- Sample-efficient Learning of Concepts with Theoretical Guarantees: from Data to Concepts without Interventions (NeurIPS 2025 · similarity 0.88)
- Hybrid Concept Bottleneck Models (CVPR 2025 · similarity 0.87)
- Energy-Based Concept Bottleneck Models: Unifying Prediction, Concept Intervention, and Probabilistic Interpretations (ICLR 2024 · similarity 0.87)
- Stochastic Concept Bottleneck Models (NeurIPS 2024 · 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.