LLMs Cannot Perform Abductive 'Jump' in Physics, New Paper Argues

A recent preprint available on arXiv (2608.14397) contends that Large Language Models (LLMs) are incapable of executing the abductive 'Jump' that resulted in Einstein's equivalence principle, despite their proficiency in both induction and deduction. The authors reference Zahavy [2026] to explain this shortcoming as stemming from a lack of embodied simulation. Conversely, Zheng-Xin [2026] and Farmer [2026] challenge the necessity of embodiment for abduction, highlighting alternative pathways to General Relativity and forms of abduction that do not depend on sensorimotor grounding. The study examines Max Planck's 1900 resolution of the blackbody radiation issue, where Planck's postulate E = hν emerged from a mathematical implication of classical theory, rather than embodied simulation. The authors argue that this postulate's acceptance required a connection between epistemic error and physical cost, a coupling that LLMs do not possess, preventing them from making the 'Jump'.