Model merging offers a promising avenue for knowledge integration and parallel development without retraining. Yet, existing methods either ignore the geometry of the loss landscape or rely on intractable full-space Hessian approximations. We propose EpiMer, a framework that casts model merging as solving the Fréchet mean on a Riemannian manifold and restricts the computation to a low-rank subspace spanned by the task vectors. With the expected Hessian as the metric, we reveal a connection between local curvature and epistemic uncertainty of the parameters. Our theoretical analysis decomposes the merging error bound into the subspace Fréchet variance and the residual energy, and provides a closed-form characterization of when curvature-aware merging provably outperforms flat-geometry methods. In addition, our framework unifies both curvature-aware methods and recent spectral methods as special cases of the subspace Fréchet mean with different geometric metrics. Merging fine-tuned CLIP-ViT models on eight image classification tasks, Epistemic Merging strictly outperforms the baselines on all three CLIP-ViT backbones at matched rank, improving the across-task average accuracy and worst-task accuracy on every backbone.

Quantum Support Vector Machines face scalability challenges due to high-dimensional quantum states and hardware limitations. We propose an embedding-aware quantum-classical pipeline combining class-balanced k-means distillation with pretrained Vision Transformer embeddings. Our key finding: ViT embeddings uniquely enable quantum advantage, achieving up to 8.02% accuracy improvements over classical SVMs on Fashion-MNIST and 4.42% on MNIST, while CNN features show performance degradation. Using 16-qubit tensor network simulation via cuTensorNet, we provide the first systematic evidence that quantum kernel advantage depends critically on embedding choice, revealing fundamental synergy between transformer attention and quantum feature spaces. This provides a practical pathway for scalable quantum machine learning that leverages modern neural architectures.

We raise a challenging research question in the following: Is there a benchmark that reflects the exact reliance on spurious features of CLIP?

We investigated the reproducibility of FairCLIP, proposed by Luo et al. (2024), for improving the group fairness of CLIP (Radford et al., 2021) by minimizing image-text similarity score disparities across sensitive groups using the Sinkhorn distance. The experimental setup of Luo et al. (2024) was reproduced to primarily investigate the research findings for FairCLIP. The model description by Luo et al. (2024) was found to differ from the original implementation. Therefore, a new implementation, A-FairCLIP, is introduced to examine specific design choices. Furthermore, FairCLIP+ is proposed to extend the FairCLIP objective to include multiple attributes. Additionally, the impact of the distance minimization on FairCLIP's fairness and performance was explored. In alignment with the original authors, CLIP was found to be biased towards certain demographics when applied to zero-shot glaucoma classification using medical scans and clinical notes from the Harvard-FairVLMed dataset. However, the experimental results on two datasets do not support their claim that FairCLIP improves the performance and fairness of CLIP. Although the regularization objective reduces Sinkhorn distances, both the official implementation and the aligned implementation, A-FairCLIP, were not found to improve performance nor fairness in zero-shot glaucoma classification.

In real-world vision-language applications, practitioners increasingly rely on large, pretrained foundation models rather than custom-built solutions, despite limited transparency regarding their training data and processes. While these models achieve impressive performance on general benchmarks, their effectiveness can decline notably under specialized domain shifts, such as unique imaging conditions or environmental variations. In this work, we introduce Deepbench, a framework designed to assess domain-specific robustness of vision-language models (VLMs). Deepbench leverages a large language model (LLM) to generate realistic, context-aware image corruptions tailored to specific deployment domains without requiring labeled data. We evaluate a range of contrastive vision-language architectures and architectural variants across six real-world domains and observe substantial variability in robustness, highlighting the need for targeted, domain-aware evaluation. Deepbench is released as open-source software to support further research into domain-aware robustness assessment.