Loading...
Loading...
Mathematician Timothy Gowers reports that ChatGPT 5.5 Pro produced PhD-level mathematical research in about an hour with minimal human input, prompting him to raise his assessment of LLM capabilities. He tested the model on open problems from Mel Nathanson’s paper on additive number theory and found LLMs increasingly able to spot overlooked simple arguments or assemble existing literature to solve research-level problems. Gowers argues this raises the bar for what counts as a sufficiently hard p
LLMs demonstrating the ability to produce publishable mathematical research changes expectations for research assistance, doctoral training, and verification workflows. Tech professionals must reassess tooling, evaluation, and risk controls when models can autonomously generate research-level outputs.
Dossier last updated: 2026-05-11 16:17:13
Field Medalist Timothy Gowers tested ChatGPT 5.5 Pro and the model produced publishable-level results on additive number theory problems within minutes, prompting alarm about doctoral training. Using only high-level prompts, Gowers fed problems from Mel Nathanson’s paper to GPT-4o-class ChatGPT 5.5 Pro; the model produced a quadratic upper-bound construction (improving on an exponential bound) in ~17 minutes and combined results into a LaTeX preprint in under an hour. It later generated novel k-dissociated-set ideas that extended MIT student Isaac Rajagopal’s work, iterating to stronger bounds with minimal human math input. Gowers warns this raises authorship, publication, and PhD training challenges; Terence Tao suggests human “digestion” of proofs remains a key value. The story matters for AI-assisted research workflows, publication norms, and graduate education in math and related fields.
A recent experience with ChatGPT 5.5 Pro
Mathematician Timothy Gowers reports that ChatGPT 5.5 Pro produced PhD-level mathematical research in about an hour with minimal human input, prompting him to raise his assessment of LLM capabilities. He tested the model on open problems from Mel Nathanson’s paper on additive number theory and found LLMs increasingly able to spot overlooked simple arguments or assemble existing literature to solve research-level problems. Gowers argues this raises the bar for what counts as a sufficiently hard problem: researchers must now expect LLMs to solve easier open questions. The piece highlights the implications for mathematical research practices and problem selection as powerful generative models become research collaborators.
<a href="https://twitter.com/wtgowers/status/2052830948685676605" rel="nofollow">https://twitter.com/wtgowers/status/2052830948685676605</a><p><a href="https://xcancel.com/wtgowers/status/2052830948685676605" rel="nofollow">https://xcancel.com/wtgowers/status/2052830948685676605</a>