Multilingual vision-language models (VLMs) promise universal image-text retrieval, yet their social biases remain underexplored. We perform the first systematic audit of four public multilingual CLIP variants: M-CLIP, NLLB-CLIP, CAPIVARA-CLIP, and the debiased SigLIP-2, covering ten languages that differ in resource availability and morphological gender marking. Using balanced subsets of FairFace and the PATA stereotype suite in a zero-shot setting, we quantify race and gender bias and measure stereotype amplification. Contrary to the intuition that multilinguality mitigates bias, every model exhibits stronger gender skew than its English-only baseline. CAPIVARA-CLIP shows its largest biases precisely in the low-resource languages it targets, while the shared encoder of NLLB-CLIP and SigLIP-2 transfers English gender stereotypes into gender-neutral languages; loosely coupled encoders largely avoid this leakage. Although SigLIP-2 reduces agency and communion skews, it inherits -- and in caption-sparse contexts (e.g., Xhosa) amplifies -- the English anchor's crime associations. Highly gendered languages consistently magnify all bias types, yet gender-neutral languages remain vulnerable whenever cross-lingual weight sharing imports foreign stereotypes. Aggregated metrics thus mask language-specific hot spots, underscoring the need for fine-grained, language-aware bias evaluation in future multilingual VLM research.

Federated learning (FL) is a new paradigm for training machine learning (ML) models without sharing data. While applying FL in cross-silo scenarios, where organizations collaborate, it is necessary that the FL system is reliable; however, participants can fail due to various reasons (e.g., communication issues or misconfigurations). In order to provide a reliable system, it is necessary to analyze the impact of participant failures. While this problem received attention in cross-device FL where mobile devices with limited resources participate, there is comparatively little research in cross-silo FL. Therefore, we conduct an extensive study for analyzing the impact of participant failures on the model quality in the context of inter-organizational cross-silo FL with few participants. In our study, we focus on analyzing generally influential factors such as the impact of the timing and the data as well as the impact on the evaluation, which is important for deciding, if the model should be deployed. We show that under high skews the evaluation is optimistic and hides the real impact. Furthermore, we demonstrate that the timing impacts the quality of the trained model. Our results offer insights for researchers and software architects aiming to build robust FL systems.

Shortcuts, spurious rules that perform well during training but fail to generalize, present a major challenge to the reliability of deep networks (Geirhos et al., 2020). However, the impact of shortcuts on feature representations remains understudied, obstructing the design of principled shortcut-mitigation methods. To overcome this limitation, we investigate the layer-wise localization of shortcuts in deep models. Our novel experiment design quantifies the layer-wise contribution to accuracy degradation caused by a shortcut-inducing skew by counterfactual training on clean and skewed datasets. We employ our design to study shortcuts on CIFAR-10, Waterbirds, and CelebA datasets across VGG, ResNet, DeiT, and ConvNeXt architectures. We find that shortcut learning is not localized in specific layers but distributed throughout the network. Different network parts play different roles in this process: shallow layers predominantly encode spurious features, while deeper layers predominantly forget core features that are predictive on clean data. We also analyze the differences in localization and describe its principal axes of variation. Finally, our analysis of layer-wise shortcut-mitigation strategies suggests the hardness of designing general methods, supporting dataset- and architecture-specific approaches instead.

Algorithmic trading requires short-term tactical decisions consistent with long-term financial objectives. Reinforcement Learning (RL) has been applied to such problems, but adoption is limited by myopic behaviour and opaque policies. Large Language Models (LLMs) offer complementary strategic reasoning and multi-modal signal interpretation when guided by well-structured prompts. This paper proposes a hybrid framework in which LLMs generate high-level trading strategies to guide RL agents. We evaluate (i) the economic rationale of LLM-generated strategies through expert review, and (ii) the performance of LLM-guided agents against unguided RL baselines using Sharpe Ratio (SR) and Maximum Drawdown (MDD). Empirical results indicate that LLM guidance improves both return and risk metrics relative to standard RL.

This causes traditional evaluation metrics to be biased towards majority classes (e.g.

Section A provides a proof that isometry preserves angles. Section D lists the grid considered for hyper-parameters. T is an isometry iff it preserves inner products. Suppose T is an isometry. Conversely, if T preserves inner products, then nullT (v w),T ( v w) null = null v w,v w null, which implies null T ( v w)null = null v w null, and since T is linear, nullT (v) T ( w) null = null v w null .

Reviewer 3 Thank you for your review.

We investigate how to calculate Kaplan-Meier survival curves across multiple health-care jurisdictions while protecting patient privacy with node-level differential privacy. Each site discloses its curve only once, adding Laplace noise whose scale is determined by the length of the common time grid; the server then averages the noisy curves, so the overall privacy budget remains unchanged. We benchmark four one-shot smoothing techniques: Discrete Cosine Transform, Haar Wavelet shrinkage, adaptive Total-Variation denoising, and a parametric Weibull fit on the NCCTG lung-cancer cohort under five privacy levels and three partition scenarios (uniform, moderately skewed, highly imbalanced). Total-Variation gives the best mean accuracy, whereas the frequency-domain smoothers offer stronger worst-case robustness and the Weibull model shows the most stable behaviour at the strictest privacy setting. Across all methods the released curves keep the empirical log-rank type-I error below fifteen percent for privacy budgets of 0.5 and higher, demonstrating that clinically useful survival information can be shared without iterative training or heavy cryptography.