Localized Conformal Prediction for Image Classification with Vision-Language Models

Conformal predictions have attracted significant attention in the field of uncertainty quantification, mainly because of their strong marginal coverage guarantees. Full conditional guarantee is not an attainable goal, a well known fact in conformal predictions literature. As a result, several approaches have tried to approximate this behavior by adapting the conformal sets of test-time samples according to their similarity to calibration examples. Although the latter has gained traction and shown impressive performances for regression problems, its application to image classification remains under-explored. We conduct an extensive benchmarking on natural image classification tasks with vision-language models (VLMs), using our open source implementation of a recent localized conformal prediction algorithm. We show that straightforward usage of the cosine similarity between test-time and calibration visual features, an intuitive choice for VLMs, is not sufficient to improve over the non-local baselines. In response, we propose a simple non-linear transformation of the cosine similarities, which conserves marginal coverage guarantees and achieves statistically significant mean set sizes reduction. Code is available at https://github.com/cfuchs2023/lcp-vlm/.

No Need for "Learning" to Defer? A Training Free Deferral Framework to Multiple Experts through Conformal Prediction

AI systems often fail to deliver reliable predictions across all inputs, prompting the need for hybrid human-AI decision-making. Existing Learning to Defer (L2D) approaches address this by training deferral models, but these are sensitive to changes in expert composition and require significant retraining if experts change. We propose a training-free, model- and expert-agnostic framework for expert deferral based on conformal prediction. Our method uses the prediction set generated by a conformal predictor to identify label-specific uncertainty and selects the most discriminative expert using a segregativity criterion, measuring how well an expert distinguishes between the remaining plausible labels. Experiments on CIFAR10-H and ImageNet16-H show that our method consistently outperforms both the standalone model and the strongest expert, with accuracies attaining $99.57\pm0.10\%$ and $99.40\pm0.52\%$, while reducing expert workload by up to a factor of $11$. The method remains robust under degraded expert performance and shows a gradual performance drop in low-information settings. These results suggest a scalable, retraining-free alternative to L2D for real-world human-AI collaboration.