The Mean is the Mirage: Entropy-Adaptive Model Merging under Heterogeneous Domain Shifts in Medical Imaging

Entropy minimization (EM) is the dominant objective for test-time adaptation, yet its failure mode, model collapse, remains poorly understood. In this work, we show that distribution shifts can cause feature clusters corresponding to distinct classes in the model's representation space to merge, while the decision boundary remains fixed. This induces a systematic skew in the predicted class distribution, referred to as prediction bias. Prediction bias refers to a shift in the predicted class distribution, with some classes overrepresented and others suppressed. We show that entropy minimization amplifies this prediction bias by tightening the existing clusters, reinforcing the incorrect groupings until all predictions collapse to a trivial solution. Next, to demonstrate the significance of prediction bias and mitigate it, we further propose Distribution Shift Bias Reduction (DSBR), a bias-correcting objective that specifically targets this failure mode by equalizing the contribution of each predicted class to the unsupervised entropy minimization loss. To study this failure mode, we design suitable adaptation settings using four medical-imaging datasets and additionally evaluate on ImageNet-C. We find that DSBR consistently stabilizes test-time adaptation, prevents model collapse, and matches or outperforms state-of-the-art methods. Moreover, DSBR operates solely at test-time.

Federated learning enables collaborative model training across medical institutions without sharing raw data, but its performance is often limited by domain heterogeneity across clients. Existing approaches to address this challenge fall into two main paradigms: model-side personalization, which adapts model parameters to each client, and data-side harmonization, which reduces inter-client variation at the input level. Despite their widespread use, these strategies have not been systematically compared. In this work, we conduct a comprehensive study across six medical imaging settings-colon polyp, skin lesion, and breast tumor segmentation, and tuberculosis CXR, brain tumor, and breast tumor classification-covering diverse types of domain shift. We evaluate a broad set of state-of-the-art harmonization and personalization methods under a unified framework. Our results reveal a conditional trade-off driven by the nature of heterogeneity: harmonization is more effective when variation is primarily appearance-based (e.g., CXR classification), while personalization performs better when differences are structural (e.g., colon polyp segmentation). When inter-client variation is limited, both strategies perform similarly. These findings demonstrate that the effectiveness of adaptation in federated medical imaging depends on the type and magnitude of domain shift rather than the strategy alone. We provide practical guidelines for selecting between harmonization and personalization and highlight directions for future hybrid approaches that combine both paradigms. Code is available at https://github.com/ChamaniS/WhenToAdapt.

Large vision-language models (LVLMs) can be adapted to specialized medical imaging tasks via parameter-efficient fine-tuning approaches such as low-rank adaptation (LoRA), leading to a growing ecosystem of expert models tailored to specific imaging modalities and clinical scenarios. However, deploying multiple expert LVLMs in practice incurs substantial computational and operational overhead. Model merging provides a promising solution by consolidating multiple experts into a single model without retraining, yet it remains largely unexplored in the medical domain. In this work, we present the first systematic study of model merging for medical LVLMs. We introduce MergeMedBench, a comprehensive benchmark spanning eight imaging modalities and diverse clinical task types, comprising 16 LoRA fine-tuned models built upon two mainstream architectures. We conduct an extensive evaluation of existing merging methods and further propose winner-take-all, a simple and hyperparameter-free approach that retains only the most dominant parameters across expert models. By preserving the critical parameters that govern model behavior and discarding weaker ones, our method avoids the information dilution inherent in averaging- or alignment-based strategies. Despite its simplicity, winner-take-all consistently outperforms existing approaches, offering both a new perspective on LoRA merging and a strong practical baseline for future research.