DeepMind Paper: LLMs Could Derive Relativity But Fail to Invent It From Data

rohanpaul_ai · x · 2026-07-30

A Google DeepMind position paper examines the limits of LLMs in scientific discovery by separating the process into induction, deduction, and abduction.

Using Einstein's development of General Relativity as a case study, the authors note that Newtonian gravity had almost no empirical error signal (inertial and gravitational mass agreed to 10⁻⁹). A compression-driven system relying on fitting data would find little gradient to rebuild spacetime. Einstein’s breakthrough relied on conceptual conflicts and thought experiments—an abductive jump from experience to a new explanatory premise.

The paper suggests that while LLMs can derive consequences from supplied axioms, they struggle with true invention. The proposed direction involves action-controllable world models, allowing agents to intervene in physically consistent simulations and translate that experience into candidate axioms.

Original post →

More from AGI Musings

AGI Musings channel →