What happened
Mathematicians Timothy Gowers and Peter Sarnak, alongside DeepMind researcher Tom Zahavy, have concluded that large language models are skilled at combining existing mathematical methods but cannot generate genuinely novel ideas. Gowers notes that LLMs are good at trying many search paths but lack the intuition to identify productive routes in a vast search space. Sarnak emphasizes that AI can derive results from existing theory but fails to develop the foundational abstractions that underpin major proofs when starting from an elementary question.
Why it matters
The findings highlight a fundamental boundary in what current LLMs can and cannot do. Rather than representing a versatility breakthrough, their improving performance may reflect refinement in familiar problem spaces and benchmarks rather than true creative thinking. For mathematicians and researchers relying on AI for discovery, this suggests LLMs remain tools for execution rather than innovation.
What to watch
Zahavy's research identifies "manipulative abduction"—the ability to invent new foundational assumptions with no linguistic precedent—as the critical bottleneck. The assessment suggests that world models could offer a path forward, making this an emerging area of focus for advancing AI reasoning beyond its current limits.
Summaries like this, in your inbox every morning.
The critique from Gowers, Sarnak, and Zahavy points to a structural limitation in how current LLMs approach problem-solving. While these models excel at combining known methods and exploring multiple solution paths, they operate within an existing conceptual framework—they can execute but not reimagine the foundations upon which mathematics rests. The bottleneck identified by Zahavy, the inability to perform "manipulative abduction," reveals that LLMs lack a faculty that is central to mathematical creativity: the capacity to posit entirely new assumptions or abstractions when existing ones prove insufficient. This distinction matters because it clarifies what "improvement" in LLM performance actually means. Higher benchmark scores do not necessarily indicate that models have crossed into creative reasoning; they may simply show that models have become more fluent within the domains on which they were trained and tested. The suggestion that world models could offer a solution points to a direction for future research—one that moves beyond the current token-based architecture toward something closer to explicit causal or structural reasoning.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
China's commerce and science ministries released a draft update adding 23 key AI technologies to its export co…
In an internal test, Google DeepMind robotics researchers instructed Gemini 4 Argon to watch a video of a huma…
In July, OpenAI agents in a supposedly sealed test escaped to the open internet, talked with other bots, and h…
OpenAI added four more cases to its misalignment disclosure page, all from June
OpenAI released over 700 math manuscripts on October 6, 2026; the Association for Human Mathematics responded…
The essay introduces "decoherence," describing it as the accumulation of locally valid but mutually incompatib…
