Skip to content
conftrace_
2026 ICML ICML 2026

Position: Prioritize Identifying Structure, Not Complex Models, for Scientific Discovery

Researchers rarely observe the mechanistic drivers of scientific processes directly. Instead, they observe proxies, which are measurable quantities correlated with one or more of these latent drivers. This position paper argues that mechanistic learning from high-dimensional proxies is generically underdetermined. Many incompatible mechanisms induce essentially the same observational relationships on the support of the data, so predictive success and coherent explanations are insufficient evidence of mechanism discovery. Machine Learning (ML) and Artificial Intelligence (AI) models can therefore generate scientific hypotheses that sound plausible while the underlying mechanism remains unidentified. The problem is particularly acute with large language models (LLMs), which tend to collapse entire equivalence classes of explanations into a single fluent narrative. The paper proposes concrete standards for “mechanistic AI,” norms that are necessary if AI-centered workflows are to support science rather than merely simulate it.

More from ICML 2026

All ICML 2026 papers →