Accelerating Mathematical and Scientific Discovery with Gemini Deep Think
Google DeepMind
Google DeepMind says its Gemini Deep Think AI helped solve real, unsolved research problems in math, physics, and computer science - not just contest puzzles. Two new papers show it disproving a decade-old conjecture and co-authoring papers with almost no human help, which is a different kind of milestone than winning a math olympiad.
Based on reporting by Google DeepMind — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google DeepMind spent the back half of 2025 chasing gold medals — literally. Gemini Deep Think matched Olympiad-level performance at the IMO in July, then did it again at the International Collegiate Programming Contest. Those were tightly scoped, well-defined problems with known answers. The two papers DeepMind published this week are a different animal entirely: open research questions where nobody, including the researchers, knew the answer going in.
The headline effort is an internal system called Aletheia, a math research agent built on top of Deep Think. Ordinary language models tend to bluff their way through advanced math because the training data thins out fast once you leave undergraduate territory — hallucinated citations, sloppy computation, false confidence. Aletheia is built to fight that instinct. It generates candidate proofs, runs them through a natural-language verifier that hunts for logical holes, revises, and — this is the part DeepMind keeps emphasizing — it's allowed to just say it can't solve something rather than fabricate a plausible-looking wrong answer. It also browses the web and uses Google Search to check its work against published literature instead of inventing references.
The results are concrete enough to name. One paper, generated with zero human intervention, computed structure constants in arithmetic geometry known as eigenweights. Another documents a human-AI collaboration proving bounds on interacting-particle systems. DeepMind also ran the agent against 700 open problems from the Erdős Conjectures database and came away with autonomous solutions to four of them, including Erdős-1051, which fed into a broader generalization published separately. DeepMind is careful here — it's proposing a four-level taxonomy for grading how much AI actually contributed to a result, and it explicitly stops at Level 2,
My take — AI-written commentary, not fact-checked reporting
AI won't grade its own results as a 'major advance' yet. The second paper moves outside pure math into computer science, physics and economics, working through 18 problems with domain experts using what the team calls an 'Advisor' model — humans steering Gemini through cycles of proof and refutation they're calling 'Vibe-Proving.' The wins read like a greatest-hits list for stuck problems: Deep Think broke a stall on Max-Cut and Steiner Tree algorithms by importing tools from continuous mathematics, like the Kirszbraun theorem, that nobody had thought to apply there. It also disproved a decade-old assumption in online submodular optimization by constructing a tight three-item counterexample, explained why an ad hoc machine-learning denoising trick actually works, extended an auction theory result from rational to real-valued bids, and found a closed-form solution to a gravitational-wave integral involving cosmic strings using Gegenbauer polynomials. None of this makes Gemini an independent scientist. Every example still depended on expert mathematicians and physicists framing the problem, judging the output, and deciding what counted as a real result. But the shift from 'solve this known Olympiad problem' to 'help me disprove this conjecture nobody has cracked since 2015' is a meaningfully different kind of claim, and it's the first time DeepMind has put its research process — including a formal scale for AI contribution — out for the math community to argue with in public.
Read more about this at: Google DeepMind