Fields Medalists Say LLMs Are Strong Calculators but Weak Creative Thinkers — What That Means

Two of the most prominent mathematicians working on AI-assisted research have drawn a specific line around what language models can and cannot do. Timothy Gowers and Peter Sarnak — both Fields Medal-level figures working at the intersection of mathematics and AI capability evaluation — argue that the creative gap is not a benchmark failure. It’s a structural limitation.

What Happened

The assessment centers on what DeepMind’s Tom Zahavy calls “manipulative abduction” — the ability to invent foundational mathematical assumptions without linguistic precedent. When a mathematician attacks an open problem, a significant part of the work involves developing new conceptual abstractions: new objects, new frameworks, new ways of framing the question that don’t exist anywhere in prior literature. That’s precisely what current language models cannot do.

What they can do is the complementary task. LLMs excel at combining existing mathematical methods, exploring multiple solution pathways efficiently, and applying established theory to derive conclusions. Sarnak specifically notes that when tasks require developing “the fundamental abstractions that underpin major proofs emerging from elementary questions” — the generative step, not the applying step — current models struggle.

Zahavy’s framing: “LLMs lack the intuition to pick productive routes in vast search spaces” when developing genuinely novel abstractions. The issue isn’t the search — it’s the orientation.

Why It Matters

The mathematicians’ assessment matters for agentic AI developers because it clarifies the boundary between tasks where AI coding and research agents are genuinely useful and tasks where they currently aren’t.

Agents are strong at the applying-and-combining layer: retrieving relevant methods from literature, checking whether a standard technique applies, generating variants of known approaches, and exploring computational branches in parallel. They are weak at the inventing layer: identifying that a new conceptual framework is needed and then generating it.

This maps onto software engineering in a pattern that practitioners already recognize. AI coding agents are excellent at implementing known patterns, refactoring to match existing conventions, and testing against specified requirements. They struggle with novel architecture design — problems where the solution approach doesn’t yet exist in the training distribution and needs to be invented.

The creative gap is not infinite or fixed. World models, per Zahavy, might offer a path toward better mathematical intuition. But that’s a research direction, not a current capability.

Source: The Decoder

Content created with AI assistance and reviewed for accuracy.