Do Existing Preconditioners Improve Biomedical Tabular Foundation Learning? An Empirical Study on TabPFN Optimization

Abstract

Tabular foundation models have recently shown strong potential for structured biomedical data analysis. Among them, TabPFN has emerged as an effective approach for low-data tabular classification tasks. However, the impact of optimization and preconditioning strategies on biomedical fine-tuning remains largely unexplored. In this work, we present a comprehensive empirical investigation of five AdamW-based preconditioning strategies for fine-tuning TabPFN v2.5 on 59 biomedical datasets spanning Alzheimer's disease, breast cancer, schizophrenia, significant memory concern (SMC), KEEL biomedical datasets, and UCI biomedical benchmarks. The evaluation considers predictive performance, computational efficiency, and statistical significance analysis. Experimental results demonstrate that the original AdamW optimizer consistently achieves the best overall performance and statistical ranking, while existing curvature-aware preconditioners fail to provide reliable improvements across diverse biomedical learning scenarios. The findings suggest that generic preconditioning approaches may not adequately capture the optimization characteristics of biomedical tabular learning, motivating the development of biomedical-aware preconditioners specifically tailored for healthcare-oriented tabular foundation models.