Keith Feldman

Gavin Levinson, Keith Feldman

As large language models (LLMs) continue to improve at completing discrete tasks, they are being integrated into increasingly complex and diverse real-world systems. However, task-level success alone does not establish a model's fit for use in practice. In applied, high-stakes settings, LLM effectiveness is driven by a wider array of sociotechnical determinants that extend beyond conventional performance measures. Although a growing set of metrics capture many of these considerations, they are rarely organized in a way that supports consistent evaluation, leaving no unified taxonomy for assessing and comparing LLM utility across use cases. To address this gap, we introduce the Language Model Utility Taxonomy (LUX), a comprehensive framework that structures utility evaluation across four domains: performance, interaction, operations, and governance. Within each domain, LUX is organized hierarchically into thematically aligned dimensions and components, each grounded in metrics that enable quantitative comparison and alignment of model selection with intended use. In addition, an external dynamic web tool is provided to support exploration of the framework by connecting each component to a repository of relevant metrics (factors) for applied evaluation.

Steven J. Krieg, Nitesh V. Chawla, Keith Feldman

The widespread application of machine learning techniques to biomedical data has produced many new insights into disease progression and improving clinical care. Inspired by the flexibility and interpretability of graphs (networks), as well as the potency of sequence models like transformers and higher-order networks (HONs), we propose a method that identifies combinations of risk factors for a given outcome and accurately encodes these higher-order relationships in a graph. Using historical data from 913,475 type 2 diabetes (T2D) patients, we found that, compared to other approaches, the proposed networks encode significantly more information about the progression of T2D toward a variety of outcomes. We additionally demonstrate how structural information from the proposed graph can be used to augment the performance of transformer-based models on predictive tasks, especially when the data are noisy. By increasing the order, or memory, of the graph, we show how the proposed method illuminates key risk factors while successfully ignoring noisy elements, which facilitates analysis that is simultaneously accurate and interpretable.