Field Medalist Uses AI to Crack PhD‑Level Math Problems – Implications for Future Researchers

Timothy Gowers, a Cambridge professor and Fields Medalist, posted that he fed several open problems from Melvyn Nathanson’s work in additive number theory to GPT‑5.5 Pro. The model spent about 17 minutes producing a construction that turned an exponential bound into a polynomial one, and then formatted the argument as a pre‑print in a further two minutes. Afterward Gowers gave the model a related MIT student result (Isaac Rajagopal) and the model not only improved it but also introduced a previously unseen mathematical construction, all within two hours and with virtually no mathematical prompting from Gowers.

Gowers argues that these results raise urgent questions for the mathematics community. He notes that current journals and arXiv do not accept AI‑generated content, suggesting the need for a dedicated platform where AI‑produced results can be archived and verified by human mathematicians. He also warns that if AI can solve “medium‑difficulty” open problems in hours, traditional PhD‑training methods—assigning such problems to novices—may become ineffective, effectively raising the entry barrier for new researchers.

He offers two nuanced views: (1) PhD students can and should use AI as a collaborative partner, shifting the goal from “prove what AI cannot” to “co‑solve what neither can alone”; (2) the impact may be limited to problem‑oriented fields like combinatorial number theory, while areas that rely on aesthetic judgment may remain resistant to AI assistance.

In parallel, DeepMind’s recent paper “AI Co‑Mathematician: Accelerating Mathematicians with Agentic AI” proposes a workflow‑level solution. The system introduces a coordinating agent that decomposes a research task, dispatches parallel work streams (literature review, computation, search), and retains all failed paths for later analysis. A demo shows a mathematician uploading a paper and receiving guided clarification before any solving begins, emphasizing that asking the right question is more critical than the answer itself.

The paper reports that the system achieved a 48 % score on the FrontierMath Tier 4 benchmark—the highest among existing AI systems—demonstrating that a unified orchestration layer can substantially improve AI‑assisted mathematics. However, the authors acknowledge that the system is still in limited release and that broader adoption will require further development of rigorous verification and uncertainty‑handling mechanisms.

Overall, the article combines Gowers’s experiential evidence with DeepMind’s architectural proposal to illustrate both the immediate capabilities of large language models in solving advanced mathematical problems and the longer‑term challenges of integrating AI into the full research workflow.