Gemini Deep Think solved open math problems and got a paper into ICLR

DeepMind's post is unusually specific about what Gemini Deep Think actually did. In mathematics it solved four open problems from the Erdős database on its own, produced a peer-reviewed paper on eigenweights in arithmetic geometry without human intervention, and reached about 90% on the IMO-ProofBench Advanced test as inference compute scaled. In physics and computer science it contributed to 18 collaborative problems, settled a decade-old conjecture in online submodular optimization with a specific counterexample, and had one paper accepted to ICLR 2026.

The method matters as much as the results. The work runs through a math research agent called Aletheia that loops generation, verification, and revision with web search to catch hallucinations, and the human collaboration is structured as an advisor relationship with deliberate steps to reduce confirmation bias. The post also tracks improvement over time, from the July 2025 IMO gold-medal version to the January 2026 one. Read the full account on DeepMind's blog.

Why it matters

If you do research, the takeaway is the verification loop, not the headline solves. The results that held up came with a generate-verify-revise structure and human advisors checking bias, which is the part you would have to copy to trust output like this in your own field.