Taisa Kushner

Andrew Lauziere, Jonathan Daugherty, Taisa Kushner

As large language models (LLMs) continue to improve and see further integration into software systems, so does the need to understand the conditions in which they will perform. We contribute a statistical framework for understanding the impact of specific prompt features on LLM performance. The approach extends previous explainable artificial intelligence (XAI) methods specifically to inspect LLMs by fitting regression models relating portions of the prompt to LLM evaluation. We apply our method to compare how two open-source models, Mistral-7B and GPT-OSS-20B, leverage the prompt to perform a simple arithmetic problem. Regression models of individual prompt portions explain 72% and 77% of variation in model performances, respectively. We find misinformation in the form of incorrect example query-answer pairs impedes both models from solving the arithmetic query, though positive examples do not find significant variability in the impact of positive and negative instructions - these prompts have contradictory effects on model performance. The framework serves as a tool for decision makers in critical scenarios to gain granular insight into how the prompt influences an LLM to solve a task.

Valentina Roquemen-Echeverri, Taisa Kushner, Peter G. Jacobs et al.

Simulating glucose dynamics in individuals with type 1 diabetes (T1D) is critical for developing personalized treatments and supporting data-driven clinical decisions. Existing models often miss key physiological aspects and are difficult to individualize. Here, we introduce physiologically-constrained neural network (NN) digital twins to simulate glucose dynamics in T1D. To ensure interpretability and physiological consistency, we first build a population-level NN state-space model aligned with a set of ordinary differential equations (ODEs) describing glucose regulation. This model is formally verified to conform to known T1D dynamics. Digital twins are then created by augmenting the population model with individual-specific models, which include personal data, such as glucose management and contextual information, capturing both inter- and intra-individual variability. We validate our approach using real-world data from the T1D Exercise Initiative study. Two weeks of data per participant were split into 5-hour sequences and simulated glucose profiles were compared to observed ones. Clinically relevant outcomes were used to assess similarity via paired equivalence t-tests with predefined clinical equivalence margins. Across 394 digital twins, glucose outcomes were equivalent between simulated and observed data: time in range (70-180 mg/dL) was 75.1$\pm$21.2% (simulated) vs. 74.4$\pm$15.4% (real; P<0.001); time below range (<70 mg/dL) 2.5$\pm$5.2% vs. 3.0$\pm$3.3% (P=0.022); and time above range (>180 mg/dL) 22.4$\pm$22.0% vs. 22.6$\pm$15.9% (P<0.001). Our framework can incorporate unmodeled factors like sleep and activity while preserving key dynamics. This approach enables personalized in silico testing of treatments, supports insulin optimization, and integrates physics-based and data-driven modeling. Code: https://github.com/mosqueralopez/T1DSim_AI

Michael Robinson, Sourya Dey, Taisa Kushner

This article presents a general and flexible method for prompting a large language model (LLM) to reveal its (hidden) token input embedding up to homeomorphism. Moreover, this article provides strong theoretical justification -- a mathematical proof for generic LLMs -- for why this method should be expected to work. With this method in hand, we demonstrate its effectiveness by recovering the token subspace of Llemma-7B. The results of this paper apply not only to LLMs but also to general nonlinear autoregressive processes.