Machine learning systems exhibit diverse failure modes: unfairness toward protected groups, brittleness to spurious correlations, poor performance on minority sub-populations, which are typically studied in isolation by distinct research communities. We propose a unifying theoretical framework that characterizes when different bias mechanisms produce quantitatively equivalent effects on model performance. By formalizing biases as violations of conditional independence through information-theoretic measures, we prove formal equivalence conditions relating spurious correlations, subpopulation shift, class imbalance, and fairness violations. Our theory predicts that a spurious correlation of strength $α$ produces equivalent worst-group accuracy degradation as a sub-population imbalance ratio $r \approx (1+α)/(1-α)$ under feature overlap assumptions. Empirical validation in six datasets and three architectures confirms that predicted equivalences hold within the accuracy of the worst group 3\%, enabling the principled transfer of debiasing methods across problem domains. This work bridges the literature on fairness, robustness, and distribution shifts under a common perspective.

Shaping inclusive representations that embrace diversity and ensure fair participation and reflections of values is at the core of many conversation-based models. However, many existing methods rely on surface inclusion using mention of user demographics or behavioral attributes of social groups. Such methods overlook the nuanced, implicit expression of opinion embedded in conversations. Furthermore, the over-reliance on overt cues can exacerbate misalignment and reinforce harmful or stereotypical representations in model outputs. Thus, we took a step back and recognized that equitable inclusion needs to account for the implicit expression of opinion and use the stance of responses to validate the normative alignment. This study aims to evaluate how opinions are represented in NLP or computational models by introducing an alignment evaluation framework that foregrounds implicit, often overlooked conversations and evaluates the normative social views and discourse. Our approach models the stance of responses as a proxy for the underlying opinion, enabling a considerate and reflective representation of diverse social viewpoints. We evaluate the framework using both (i) positive-unlabeled (PU) online learning with base classifiers, and (ii) instruction-tuned language models to assess post-training alignment. Through this, we provide a principled and structured lens on how implicit opinions are (mis)represented and offer a pathway toward more inclusive model behavior.

This paper proposes a MedGemma-based framework for automatic abnormality detection in musculoskeletal radiographs. Departing from conventional autoencoder and neural network pipelines, the proposed method leverages the MedGemma foundation model, incorporating a SigLIP-derived vision encoder pretrained on diverse medical imaging modalities. Preprocessed X-ray images are encoded into high-dimensional embeddings using the MedGemma vision backbone, which are subsequently passed through a lightweight multilayer perceptron for binary classification. Experimental assessment reveals that the MedGemma-driven classifier exhibits strong performance, exceeding conventional convolutional and autoencoder-based metrics. Additionally, the model leverages MedGemma's transfer learning capabilities, enhancing generalization and optimizing feature engineering. The integration of a modern medical foundation model not only enhances representation learning but also facilitates modular training strategies such as selective encoder block unfreezing for efficient domain adaptation. The findings suggest that MedGemma-powered classification systems can advance clinical radiograph triage by providing scalable and accurate abnormality detection, with potential for broader applications in automated medical image analysis. Keywords: Google MedGemma, MURA, Medical Image, Classification.

The proliferation of harmful online content--e.g., toxicity, spam, and negative sentiment--demands robust and adaptable moderation systems. However, prevailing moderation systems are centralized and task-specific, offering limited transparency and neglecting diverse user preferences--an approach ill-suited for privacy-sensitive or decentralized environments. We propose a novel framework that leverages in-context learning (ICL) with foundation models to unify the detection of toxicity, spam, and negative sentiment across binary, multi-class, and multi-label settings. Crucially, our approach enables lightweight personalization, allowing users to easily block new categories, unblock existing ones, or extend detection to semantic variations through simple prompt-based interventions--all without model retraining. Extensive experiments on public benchmarks (TextDetox, UCI SMS, SST2) and a new, annotated Mastodon dataset reveal that: (i) foundation models achieve strong cross-task generalization, often matching or surpassing task-specific fine-tuned models; (ii) effective personalization is achievable with as few as one user-provided example or definition; and (iii) augmenting prompts with label definitions or rationales significantly enhances robustness to noisy, real-world data. Our work demonstrates a definitive shift beyond one-size-fits-all moderation, establishing ICL as a practical, privacy-preserving, and highly adaptable pathway for the next generation of user-centric content safety systems. To foster reproducibility and facilitate future research, we publicly release our code on GitHub and the annotated Mastodon dataset on Hugging Face.

We introduce a methodology for analyzing neural networks through the lens of layer-wise Hessian matrices. The local Hessian of each functional block (layer) is defined as the matrix of second derivatives of a scalar function with respect to the parameters of that layer. This concept provides a formal tool for characterizing the local geometry of the parameter space. We show that the spectral properties of local Hessians, such as the distribution of eigenvalues, reveal quantitative patterns associated with overfitting, underparameterization, and expressivity in neural network architectures. We conduct an extensive empirical study involving 111 experiments across 37 datasets. The results demonstrate consistent structural regularities in the evolution of local Hessians during training and highlight correlations between their spectra and generalization performance. These findings establish a foundation for using local geometric analysis to guide the diagnosis and design of deep neural networks. The proposed framework connects optimization geometry with functional behavior and offers practical insight for improving network architectures and training stability.

Realizing personalized medicine at scale calls for methods that distill insights from longitudinal patient journeys, which can be viewed as a sequence of medical events. Foundation models pretrained on large-scale medical event data represent a promising direction for scaling real-world evidence generation and generalizing to diverse downstream tasks. Using Epic Cosmos, a dataset with medical events from de-identified longitudinal health records for 16.3 billion encounters over 300 million unique patient records from 310 health systems, we introduce the Curiosity models, a family of decoder-only transformer models pretrained on 118 million patients representing 115 billion discrete medical events (151 billion tokens). We present the largest scaling-law study of medical event data, establishing a methodology for pretraining and revealing power-law scaling relationships for compute, tokens, and model size. Consequently, we pretrained a series of compute-optimal models with up to 1 billion parameters. Conditioned on a patient's real-world history, Curiosity autoregressively predicts the next medical event to simulate patient health timelines. We studied 78 real-world tasks, including diagnosis prediction, disease prognosis, and healthcare operations. Remarkably for a foundation model with generic pretraining and simulation-based inference, Curiosity generally outperformed or matched task-specific supervised models on these tasks, without requiring task-specific fine-tuning or few-shot examples. Curiosity's predictive power consistently improves as the model and pretraining scale. Our results show that Curiosity, a generative medical event foundation model, can effectively capture complex clinical dynamics, providing an extensible and generalizable framework to support clinical decision-making, streamline healthcare operations, and improve patient outcomes.

Reliability and failure detection of large language models (LLMs) is critical for their deployment in high-stakes, multi-step reasoning tasks. Prior work explores confidence estimation for self-evaluating LLM-scorer systems, with confidence scorers estimating the likelihood of errors in LLM responses. However, most methods focus on single-step outputs and overlook the challenges of multi-step reasoning. In this work, we extend self-evaluation techniques to multi-step tasks, testing two intuitive approaches: holistic scoring and step-by-step scoring. Using two multi-step benchmark datasets, we show that stepwise evaluation generally outperforms holistic scoring in detecting potential errors, with up to 15% relative increase in AUC-ROC. Our findings demonstrate that self-evaluating LLM systems provide meaningful confidence estimates in complex reasoning, improving their trustworthiness and providing a practical framework for failure detection.

In autonomous driving and robotics, ensuring road safety and reliable decision-making critically depends on out-of-distribution (OOD) segmentation. While numerous methods have been proposed to detect anomalous objects on the road, leveraging the vision-language space-which provides rich linguistic knowledge-remains an underexplored field. We hypothesize that incorporating these linguistic cues can be especially beneficial in the complex contexts found in real-world autonomous driving scenarios. To this end, we present a novel approach that trains a Text-Driven OOD Segmentation model to learn a semantically diverse set of objects in the vision-language space. Concretely, our approach combines a vision-language model's encoder with a transformer decoder, employs Distance-Based OOD prompts located at varying semantic distances from in-distribution (ID) classes, and utilizes OOD Semantic Augmentation for OOD representations. By aligning visual and textual information, our approach effectively generalizes to unseen objects and provides robust OOD segmentation in diverse driving environments. We conduct extensive experiments on publicly available OOD segmentation datasets such as Fishyscapes, Segment-Me-If-You-Can, and Road Anomaly datasets, demonstrating that our approach achieves state-of-the-art performance across both pixel-level and object-level evaluations. This result underscores the potential of vision-language-based OOD segmentation to bolster the safety and reliability of future autonomous driving systems.

Deep learning-based video surveillance increasingly demands privacy-preserving architectures with low computational and environmental overhead. Federated learning preserves privacy but deploying large vision-language models (VLMs) introduces major energy and sustainability challenges. We compare three strategies for federated violence detection under realistic non-IID splits on the RWF-2000 and RLVS datasets: zero-shot inference with pretrained VLMs, LoRA-based fine-tuning of LLaVA-NeXT-Video-7B, and personalized federated learning of a 65.8M-parameter 3D CNN. All methods exceed 90% accuracy in binary violence detection. The 3D CNN achieves superior calibration (ROC AUC 92.59%) at roughly half the energy cost (240 Wh vs. 570 Wh) of federated LoRA, while VLMs provide richer multimodal reasoning. Hierarchical category grouping (based on semantic similarity and class exclusion) boosts VLM multiclass accuracy from 65.31% to 81% on the UCF-Crime dataset. To our knowledge, this is the first comparative simulation study of LoRA-tuned VLMs and personalized CNNs for federated violence detection, with explicit energy and CO2e quantification. Our results inform hybrid deployment strategies that default to efficient CNNs for routine inference and selectively engage VLMs for complex contextual reasoning.

Automatic speech recognition (ASR) for low-resource languages such as Taiwanese Hokkien is difficult due to the scarcity of annotated data. However, direct fine-tuning on Han-character transcriptions often fails to capture detailed phonetic and tonal cues, while training only on roman-ization lacks lexical and syntactic coverage. In addition, prior studies have rarely explored staged strategies that integrate both annotation types. To address this gap, we present CLiFT-ASR, a cross-lingual fine-tuning framework that builds on Mandarin HuBERT models and progressively adapts them to Taiwanese Hokkien. The framework employs a two-stage process in which it first learns acoustic and tonal representations from phonetic Tai-lo annotations and then captures vocabulary and syntax from Han-character transcriptions. This progressive adaptation enables effective alignment between speech sounds and orthographic structures. Experiments on the TAT-MOE corpus demonstrate that CLiFT-ASR achieves a 24.88% relative reduction in character error rate (CER) compared with strong baselines. The results indicate that CLiFT-ASR provides an effective and parameter-efficient solution for Taiwanese Hokkien ASR and that it has potential to benefit other low-resource language scenarios.