Title: Validity Threats in Foundation Model Research

Abstract:

While controlled experiments serve as the cornerstone of machine learning inquiry, their application to modern foundation models is often hindered by prohibitive costs at scale. Consequently, the research community has shifted toward cost-effective approximations of ideal experiments, such as proxy experiments, scaling laws, observational studies utilizing publicly available models, and single-run designs that exploit variance within individual training processes. This paper argues that there is no cost-free approach to approximating large-scale experiments; rather, reductions in compute expenditure introduce validity threats. These threats consist of hidden, and occasionally untestable, assumptions that, if violated, can undermine the validity of research conclusions. To address these challenges, we introduce an evaluation framework that reframes foundation model research as a causal inference problem. Using this framework, we assess various research strategies through four validity dimensions borrowed from empirical social sciences: statistical, internal, external, and construct validity. Our findings indicate that each methodology presents a distinct validity profile. For instance, proxy experiments sacrifice external and construct validity to gain statistical and internal validity. Observational studies are compromised by confounding variables and effect heterogeneity, while single-run designs suffer from interference among treated units. This analysis highlights several validity threats that have been overlooked in existing literature. Ultimately, our framework offers researchers a practical toolkit for rigorously examining validity threats in the design of foundation model studies.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC