Establishing Construct Validity in LLM Capability Benchmarks Requires Nomological Networks
This addresses a foundational methodological issue for researchers in AI and machine learning, offering a framework to improve the validity of capability assessments in LLMs.
The paper critiques the attribution of human-like capabilities to large language models based on benchmark performance, arguing that establishing construct validity requires adopting the nomological account from psychometrics to properly link theoretical capabilities to empirical measurements.
Recent work in machine learning increasingly attributes human-like capabilities such as reasoning or theory of mind to large language models (LLMs) on the basis of benchmark performance. This paper examines this practice through the lens of construct validity, understood as the problem of linking theoretical capabilities to their empirical measurements. It contrasts three influential frameworks: the nomological account developed by Cronbach and Meehl, the inferential account proposed by Messick and refined by Kane, and Borsboom's causal account. I argue that the nomological account provides the most suitable foundation for current LLM capability research. It avoids the strong ontological commitments of the causal account while offering a more substantive framework for articulating construct meaning than the inferential account. I explore the conceptual implications of adopting the nomological account for LLM research through a concrete case: the assessment of reasoning capabilities in LLMs.