Are We Leaving Non-Latin Scripts and Languages Behind?

The vast majority of NLP research and Large Language Models (LLMs) are focused on high-resource languages, predominantly those using the Latin script (e.g., English, French). This creates a critical gap, leading to performance disparities, systemic biases, and the exclusion of billions of speakers from the benefits of advanced AI. The disproportionate focus on Latin scripts means biases and harms are often not adequately measured or addressed for non-Latin script users. Future work must be linguistically informed and strategically address the resource and structural gaps in order to bridge the gap, so that AI can serve billions of people more safely. In my poster presentation, I am actively seeking bilingual collaborators who are proficient in English and another language/script (especially those non-Latin scripts like Arabic, Hindi, Korean, Japanese, or others) to work on practical tools and research in this area.

Are We Leaving Non-Latin Scripts and Languages Behind?

The vast majority of NLP research and Large Language Models (LLMs) are focused on high-resource languages, predominantly those using the Latin script (e.g., English, French). This creates a critical gap, leading to performance disparities, systemic biases, and the exclusion of billions of speakers from the benefits of advanced AI. The disproportionate focus on Latin scripts means biases and harms are often not adequately measured or addressed for non-Latin script users. Future work must be linguistically informed and strategically address the resource and structural gaps in order to bridge the gap, so that AI can serve billions of people more safely. In my poster presentation, I am actively seeking bilingual collaborators who are proficient in English and another language/script (especially those non-Latin scripts like Arabic, Hindi, Korean, Japanese, or others) to work on practical tools and research in this area.

Understanding molecular structure and dynamics in real time is one of the grand challenges of modern physical chemistry. My research integrates artificial intelligence with quantum molecular simulations to reconstruct molecular structures directly from ultrafast x-ray scattering patterns. While forward simulations of x-ray scattering from known geometries are well established, solving the inverse problem—inferring atomic configurations from measured patterns—remains highly challenging. To address this, I am developing supervised machine-learning models, including convolutional and graph neural networks, trained on first-principles simulations to learn the mapping between scattering patterns and molecular geometries. Once trained, these models can rapidly predict transient molecular structures and distinguish between competing reaction pathways, providing an efficient alternative to traditional ab initio molecular dynamics. This AI-driven framework aims to accelerate the creation of molecular “movies” at femtosecond timescales, opening new possibilities for understanding and controlling photochemical reactions.

Emerging non-volatile memory technologies such as Phase-Change Memory (PCM) offer high density and scalability, but they face critical challenges related to high write energy, long write latency, and limited endurance caused by frequent bit transitions and write-disturbance errors. These limitations motivate the development of self-healing memory systems that can autonomously adapt to workload behavior and mitigate reliability degradation over time. This work presents a machine learning–driven self-healing memory framework that combines adaptive write optimization with proactive error prediction. By analyzing data patterns and write characteristics, the system intelligently reduces unnecessary bit transitions during write operations, leading to lower energy consumption and improved memory lifetime. In parallel, learning-based error prediction models are used to identify error-prone memory regions before failures occur, enabling early intervention through selective rewriting, remapping, or correction.The proposed approach allows the memory system to continuously monitor its state and dynamically adjust its behavior in response to evolving error patterns and workload demands. Experimental evaluation using full-system, cycle-accurate simulation demonstrates notable reductions in write energy and error rates with minimal performance overhead. These results illustrate how integrating machine learning into memory management enables resilient, efficient, and autonomous self-healing behavior for future memory systems.

Many complex traits and diseases arise from the interactions between genetics factors and environmental exposures, commonly referred to as gene-environment (G×E) interactions. Accurately modeling these effects is important for predicting individual risk and understanding sources of trait variability, but it remains challenging due to nonlinear effects and high-dimensional feature spaces. Traditional regression-based approaches typically require interactions to be specified in advance, limiting their ability to capture complex relationships. We present a deep learning (DL) approach for predictive modeling of G×E effects that explicitly learns nonlinear 2-way and higher-order interactions directly from data, including genotype dominance effects (i.e., non-additive genetic contributions). The proposed model is a feed-forward, fully-connected neural network that takes genetic and environmental features as inputs and predicts a single outcome, including a quantitative trait, disease status, or survival phenotype. We benchmark this approach against widely used statistical and machine learning methods, including linear and penalized (LASSO, elastic-net) regression, random forest, gradient boosting (LightGBM), and a tabular prior-data fitted network (TabPFN, an alternative DL approach based on a pre-trained foundation model). Using a controlled simulation study with 100 replicated datasets of 10,000 individuals, all models were fit using main effects only, with genetic variables coded additively and no interaction terms provided. Linear regression was additionally fit under a “gold standard” specification that included main effects, G×E interactions, and appropriate dominance modeling, serving as a reference upper bound on achievable performance. Prediction accuracy was evaluated using R2 across increasing levels of interaction complexity. Under the main-effects-only specification, regression-based models achieved limited predictive performance (R2 ≈ <0.01-0.20), particularly as interaction complexity increased. In contrast, DL and boosting models achieved substantially higher R2 values in moderate-to-high complexity settings (DL: R2 ≈ 0.21-0.28; boosting: R2 ≈ 0.20-0.27), reflecting their ability to learn nonlinear and interaction-driven signal. TabPFN achieved the highest predictive performance across all complexity levels (R2 ≈ 0.16-0.30), consistently outperforming both regression-based and alternative machine learning approaches. As expected, the gold standard linear regression model yielded the highest overall R2, providing an upper bound on attainable performance. These results demonstrate the advantages of modern machine learning approaches for prediction in settings dominated by complex relationships. Ongoing work extends these methods to real-world genomic datasets to assess scalability, robustness, and practical impact.

Transformer-based large language models (LLMs) are used in language processing, yet when handling long context most often restrict the context window. Furthermore, many existing solutions are inefficient and overlook the structure inherent to documents. As a result, long-context models often treat text as a flat token stream, which obscures hierarchy and wastes computation by processing both relevant and irrelevant context. We present Hierarchical Semantic Memory Transformer (H2MT), a semantic hierarchy-aware approach that attaches to a backbone model. H2MT represents a document as a tree and performs level-conditioned routing and aggregation. It first propagates memory embeddings (summary vectors produced by the backbone) upward. Thus, child-node memory embeddings are injected into their ancestors to preserve relative context. Finally, the model applies cross-level attention to retrieve related information. H2MT improves quality at similar model size while reducing long-range attention compute and memory. The approach is most helpful for data with a semantic hierarchy that can be modeled as a tree. It uses less memory and fewer parameters.

Inverse ZrO2/Cu shows extraordinary catalytic performance converting CO2 to methanol, yet uncertainties still exist in the reaction mechanism. While conventional Cu/ZrO₂ systems often exhibit rate determining step at the formate hydrogenation, evidence for inverse ZrO₂/Cu catalysts has been conflicting. In this work, we employ density functional theory (DFT) calculations to investigate CO₂ hydrogenation reaction across an ensemble of inverse ZrO₂/Cu configurations under reaction conditions. Detailed reaction-pathway analysis reveals that all the studied inverse structures display the rate-determining step after methoxy formation, typically in hydrogenation to methanol or subsequent water formation, rather than at formate hydrogenation. Structural sensitivity is pronounced, only 19% of the catalyst ensemble are catalytically active across the full pathway, with reactivity favored by partially reduced Zr clusters and reactive site near metallic Cu surface that enhances hydrogen dissociation. Simulated reaction mechanism aligns qualitatively and quantitatively with experimental trends, supporting the view that the inverse configuration mitigates formate stabilization and shifts the reaction kinetic bottleneck to later steps in the mechanism, after formation of methoxy intermediate. These findings clarify the mechanistic origins of activity in inverse ZrO₂/Cu catalysts and highlight the importance of structural ensembles in governing CO₂ hydrogenation performance.

This poster highlights the Leatherby Libraries’ leadership in advancing AI literacy through creative, inclusive, and interdisciplinary approaches. As part of Chapman University’s “Year of AI,” the library launched initiatives such as Beyond the Lens and AI: The Next Chapter, blending art, ethics, and education to inspire campus-wide engagement. Through collaboration with IS&T, Town & Gown, and academic departments, the library positioned itself as a hub for ethical dialogue and innovation. The poster shares replicable models for how libraries can foster AI awareness through community partnerships, exhibitions, and experiential learning.

Can AI replace human subjects? Researchers are increasingly using models such as GPT-4 as surrogates for humans because they are cheaper and faster; however, do they behave like us? To find out, I built 96 AI "retail investors" and unleashed them in a stock market simulation, exposing them to viral "meme stock" buzz while holding financial fundamentals constant. The results were striking: When human retail investors see viral hype, they buy (+30–50%); my AI retail investor agents did the opposite, decreasing buying by 45%. While humans famously hold on to losing investments for too long, my agents sold losers three times faster than they sold winners. They acted exactly like financial textbooks say we should, and exactly unlike real people do. I call this "Hyper-Rationality." AI models are trained on vast amounts of advice: "avoid bubbles," "cut your losses." They prioritize logical training over character instruction; even when explicitly programmed to experience "FOMO," they calculated the transaction costs and rationally refrained from trading. The implication: AI can simulate how we should behave, but it lacks the emotional software to replicate how we actually behave.