Mathematicians Say LLMs Lack Creative Intuition

Prominent mathematicians Timothy Gowers and Peter Sarnak have highlighted the fundamental limitations of large language models in advanced mathematics. While acknowledging that these AI systems possess impressive calculation skills, the scholars argue that current models lack the creative intuition required for genuine mathematical discovery. Gowers points out that while LLMs are adept at combining established methodologies and exploring multiple search paths, they struggle to identify the most productive routes within vast search spaces.

Sarnak echoes this sentiment, noting that AI can successfully derive results from existing theories but fails when tasked with developing the novel abstractions that underpin major mathematical proofs. This perspective aligns with research from Google DeepMind scientist Tom Zahavy. In his paper titled "LLMs Can't Jump," Zahavy attributes this limitation to a bottleneck he calls "manipulative abduction," which is the capacity to formulate entirely new foundational assumptions that have no prior linguistic precedent.

For AI practitioners and researchers, these insights clarify the boundaries of current generative AI. While LLMs can serve as powerful assistants for verifying steps or automating tedious calculations, they cannot yet replace human mathematicians in formulating groundbreaking theories. To overcome these hurdles, some researchers are looking toward world models as a potential path forward to help AI transcend the limits of its training data. This ongoing assessment feeds into a larger industry debate over whether LLMs are truly becoming more versatile or simply getting better at memorizing benchmarks and navigating familiar problem spaces.