LLMs Can't Jump (philsci-archive.pitt.edu)

In a recent preprint, researcher Tom Zahavy presents the argument that large language models (LLMs) like those used in generative AI are fundamentally limited in their ability to make innovative scientific discoveries. Drawing on Albert Einstein's conceptual framework of discovery, Zahavy posits that while LLMs excel in statistical pattern recognition (induction) and formal proof (deduction), they lack the capacity for 'abduction'—the critical leap needed to formulate novel hypotheses from sparse observational data. This deficiency poses a significant challenge for the future of AI in scientific invention, as the ability to generate new ideas is essential for advancing knowledge. Zahavy's paper uses Einstein’s formulation of General Relativity as a case study, highlighting that the popular view of creativity as a mere extension of data compression fails to encompass the complex nature of discovery. The research identifies the translation of simulations into formal axioms as a critical bottleneck for AI, suggesting that the development of physically consistent, multimodal world models could provide the necessary sensory grounding to facilitate this 'jump.' This work encourages a reevaluation of how AI might be integrated into scientific methodologies, emphasizing the need for advancements that enable true creative reasoning rather than merely mimicking existing knowledge.