Mathematicians question LLMs’ ability to create new ideas
Mathematicians Timothy Gowers and Peter Sarnak credit large language models with substantial mathematical abilities, but say the systems remain limited when a problem requires genuinely new ideas. Gowers says LLMs can combine established techniques and explore numerous possible approaches. However, he argues that they lack the intuition needed to identify the few promising directions within a very large search space.
Sarnak describes a similar boundary. In his view, AI can obtain results from existing mathematical theory, yet struggles to develop the abstractions that support major proofs when it begins with a basic question. DeepMind researcher Tom Zahavy reached a comparable assessment in his paper “LLMs Can’t Jump,” attributing the difficulty to “manipulative abduction”: inventing foundational assumptions without a prior linguistic precedent.
The article presents world models as a possible way to address this limitation. The assessments are part of a wider debate over whether LLMs are becoming more versatile or mainly improving on benchmarks and familiar types of problems.
Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.