Position: AI Evaluations Should be Grounded on a Theory of Capability

Nathanael Jo, Ashia C. Wilson
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:170718-170740, 2026.

Abstract

Evaluations of generative models are now ubiquitous, and their outcomes critically shape public and scientific expectations of AI’s capabilities. Yet skepticism about their reliability continues to grow. How can we know that a reported accuracy genuinely reflects a model’s underlying performance? Although benchmark results are often presented as direct measurements of capability, in practice they are inferences: treating a score as evidence of capability already presupposes a theory of what it means to be capable at a task. We argue that AI evaluations should instead be framed as inference tasks grounded on an explicit theory of capability. While this perspective is standard in fields like psychometrics, it remains underdeveloped in AI evaluation, where core assumptions are often left implicit. As a proof-of-concept, we empirically show that reported performance can depend strongly on the evaluator’s modeling assumptions, underscoring the need for transparent, theory-driven evaluation practices. We conclude by offering an Evaluation Card to help researchers document, justify, and scrutinize the modeling decisions underlying AI evaluations.

Cite this Paper


BibTeX
@InProceedings{pmlr-v306-jo26b, title = {Position: {AI} Evaluations Should be Grounded on a Theory of Capability}, author = {Jo, Nathanael and Wilson, Ashia C.}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {170718--170740}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/jo26b/jo26b.pdf}, url = {https://proceedings.mlr.press/v306/jo26b.html}, abstract = {Evaluations of generative models are now ubiquitous, and their outcomes critically shape public and scientific expectations of AI’s capabilities. Yet skepticism about their reliability continues to grow. How can we know that a reported accuracy genuinely reflects a model’s underlying performance? Although benchmark results are often presented as direct measurements of capability, in practice they are inferences: treating a score as evidence of capability already presupposes a theory of what it means to be capable at a task. We argue that AI evaluations should instead be framed as inference tasks grounded on an explicit theory of capability. While this perspective is standard in fields like psychometrics, it remains underdeveloped in AI evaluation, where core assumptions are often left implicit. As a proof-of-concept, we empirically show that reported performance can depend strongly on the evaluator’s modeling assumptions, underscoring the need for transparent, theory-driven evaluation practices. We conclude by offering an Evaluation Card to help researchers document, justify, and scrutinize the modeling decisions underlying AI evaluations.} }
Endnote
%0 Conference Paper %T Position: AI Evaluations Should be Grounded on a Theory of Capability %A Nathanael Jo %A Ashia C. Wilson %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-jo26b %I PMLR %P 170718--170740 %U https://proceedings.mlr.press/v306/jo26b.html %V 306 %X Evaluations of generative models are now ubiquitous, and their outcomes critically shape public and scientific expectations of AI’s capabilities. Yet skepticism about their reliability continues to grow. How can we know that a reported accuracy genuinely reflects a model’s underlying performance? Although benchmark results are often presented as direct measurements of capability, in practice they are inferences: treating a score as evidence of capability already presupposes a theory of what it means to be capable at a task. We argue that AI evaluations should instead be framed as inference tasks grounded on an explicit theory of capability. While this perspective is standard in fields like psychometrics, it remains underdeveloped in AI evaluation, where core assumptions are often left implicit. As a proof-of-concept, we empirically show that reported performance can depend strongly on the evaluator’s modeling assumptions, underscoring the need for transparent, theory-driven evaluation practices. We conclude by offering an Evaluation Card to help researchers document, justify, and scrutinize the modeling decisions underlying AI evaluations.
APA
Jo, N. & Wilson, A.C.. (2026). Position: AI Evaluations Should be Grounded on a Theory of Capability. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:170718-170740 Available from https://proceedings.mlr.press/v306/jo26b.html.

Related Material