CONDA: Adaptive Concept Bottleneck for Foundation Models Under Distribution Shifts

Jihye Choi, Jayaram Raghuram, Yixuan Li, Somesh Jha

TL;DR

CONDA (CONcept-based Dynamic Adaptation) is a framework that makes interpretable Concept Bottleneck Models (CBMs) robust to real-world distribution shifts. By dynamically adapting concept vectors and prediction layers using only unlabeled test data, CONDA boosts post-deployment accuracy by up to 28% while ensuring that model explanations remain relevant to the new environment.

What Problem Does This Paper Address?

Foundation Models (FMs) like CLIP are powerful but operate as "black boxes," making them difficult to trust in safety-critical fields like healthcare or autonomous driving. Concept Bottleneck Models (CBMs) attempt to solve this by forcing the model to first predict human-understandable "concepts" (e.g., "wings," "beak") before making a final classification.

However, CBMs are notoriously brittle. If the test data differs from the training data—due to sensor noise (low-level shift) or changes in background/context (concept-level shift)—the pre-defined concepts often fail to generalize. These shifts cause significant performance drops, sometimes making interpretable models perform worse than their "black-box" counterparts.

Performance of feature-based and concept-based models under low-level and concept-level distribution shifts. CONDA significantly recovers the performance of concept-based models.

Methods

The authors propose CONDA, a three-step test-time adaptation procedure that updates the model "on the fly" using unlabeled target data:

  1. Concept-Score Alignment (CSA): This step adjusts the concept vectors to ensure that the distribution of concept scores in the new domain matches the expected distribution from the training data. It uses a Mahalanobis distance-based loss to handle low-level input perturbations like noise.
  2. Linear Probing Adaptation (LPA): Here, the model updates the final "label predictor" layer. This allows the model to re-learn which concepts are actually relevant for classification in the new environment, helping it ignore spurious correlations (like a "water" background always implying a "waterbird").
  3. Residual Concept Bottleneck (RCB): If the original concept set is incomplete for the new domain, CONDA introduces a small set of "residual" concepts. These are learned jointly with a linear predictor to capture missing information and improve accuracy.

To guide these updates without ground-truth labels, CONDA uses pseudo-labeling, creating an ensemble of zero-shot and linear-probing predictions from the underlying foundation model.

Datasets

The framework was evaluated across diverse real-world scenarios:

  • CIFAR10-C & CIFAR100-C: Standard benchmarks for low-level image corruptions (noise, blur, etc.).
  • Waterbirds: A classic dataset for concept-level shift where birds are often seen on "spurious" backgrounds (e.g., landbirds on water).
  • Metashift: Simulates disparate correlations, such as cats appearing on sofas in the source but on shelves in the target.
  • Camelyon17: A medical imaging dataset featuring natural shifts caused by different hospital equipment and lighting.

Key Insights and Ideas

  • Interpretability is Dynamic: A model's explanation is only useful if it's accurate for the current data. CONDA shows that "fixing" the interpretation at test-time actually helps the model's performance.
  • Different Shifts, Different Fixes: The authors found that CSA is crucial for low-level noise, while LPA and RCB are the heavy lifters for semantic shifts like background changes.
  • Residuals Fill the Gaps: Pre-defined concept banks (often generated by LLMs like GPT-3) are rarely perfect. The "Residual" branch allows the model to discover task-relevant features that humans might have missed during initial setup.

Key Results

  • Massive Accuracy Gains: On the Waterbirds dataset, CONDA improved the accuracy of a deployed CBM by 28%, bringing it on par with non-interpretable models.
  • Better Explanations: Qualitative analysis shows that before adaptation, the model might rely on spurious concepts (like "land" for a landbird). After CONDA, the model correctly shifts its weight to more robust, class-relevant concepts (like "beak" or "feathers").
  • Efficiency: Despite performing optimization at test-time, CONDA is fast, taking only a few seconds per batch on standard hardware.

Limitations and Future Work

  • Pseudo-label Dependency: CONDA's success relies on the underlying Foundation Model being "mostly right." If the FM is completely baffled by a new domain, the pseudo-labels will be inaccurate, and adaptation will fail.
  • Hyperparameter Sensitivity: While the authors provide a heuristic for choosing the number of residual concepts, finding the optimal balance between accuracy and interpretability remains a challenge.
  • Future Directions: The authors suggest exploring more advanced pseudo-labeling techniques and developing theoretical bounds to quantify how much a model can realistically adapt to a new domain without source data.

Summary updated

Abstract

Advancements in foundation models (FMs) have led to a paradigm shift in machine learning. The rich, expressive feature representations from these pre-trained, large- scale FMs are leveraged for multiple downstream tasks, usually via lightweight fine-tuning of a shallow fully-connected network following the representation. However, the non-interpretable, black-box nature of this prediction pipeline can be a challenge, especially in critical domains, such as healthcare, finance, and security. In this paper, we explore the potential of Concept Bottleneck Models (CBMs) for transforming complex, non-interpretable foundation models into interpretable decision-making pipelines using high-level concept vectors. Specifically, we focus on the test-time deployment of such an interpretable CBM pipeline “in the wild”, where the distribution of inputs often shifts from the original training distribution. We first identify the potential failure modes of such pipelines under different types of distribution shifts. Then we propose an adaptive concept bottleneck framework to address these failure modes, that dynamically adapts the concept-vector bank and the prediction layer based solely on unlabeled data from the target domain, without access to the source dataset. Empirical evaluations with various real-world distribution shifts show our framework produces concept-based interpretations better aligned with the test data and boosts post-deployment accuracy by up to 28%, aligning CBM performance with that of non-interpretable classification.
BibTeX
@inproceedings{choi2025conda,
  author = {Jihye Choi and Jayaram Raghuram and Yixuan Li and Somesh Jha},
  title = {{CONDA: Adaptive Concept Bottleneck for Foundation Models Under Distribution Shifts}},
  booktitle = {International Conference on Learning Representations},
  year = {2025},
  url = {https://openreview.net/forum?id=8sfc8MwG5v},
}
Figure