Google DeepMind says Gemini Deep Think is moving from Olympiad benchmarks into math, physics, and CS research

Google DeepMind's February 11, 2026 post on Gemini Deep Think marks a notable shift from benchmark bragging to research workflow claims. The company says that, under the direction of expert mathematicians and scientists, Gemini Deep Think is now helping solve professional research problems across mathematics, physics, and computer science. That matters because the target is no longer isolated contest questions, but open-ended research work with literature review, error checking, and iterative revision.

Google DeepMind frames the update as a continuation of the advanced Gemini Deep Think system that achieved Gold-medal standard at the International Mathematics Olympiad in 2025 and later posted similar results at the International Collegiate Programming Contest. The February 2026 announcement goes further by tying those capabilities to two new papers and to a math research agent internally codenamed Aletheia. According to the company, Aletheia combines Gemini Deep Think with a natural-language verifier, iterative revision loops, Google Search, and web browsing so it can test candidate solutions, reject flawed ones, and avoid spurious citations or computational mistakes.

The evaluation details are ambitious. Google DeepMind says Gemini Deep Think scored up to 90% on IMO-ProofBench Advanced as inference-time compute scaled, and that the scaling trend continued into PhD-level exercises on its internal FutureMath Basic benchmark. More importantly, the company is pointing to research outputs rather than just scores. It highlighted an autonomous paper on eigenweights in arithmetic geometry, a human-AI collaboration paper on independent sets, and a semi-autonomous review of 700 open problems in Bloom's Erdos Conjectures database that reportedly produced autonomous solutions to four open questions.

The computer science and physics examples are similarly concrete. Google DeepMind says Gemini Deep Think contributed to 18 research problems spanning algorithms, combinatorial optimization, information theory, economics, and physics. The examples it chose are notable: progress on Max-Cut and Steiner Tree by importing tools from continuous mathematics, a counterexample that refuted a decade-old intuition in online submodular optimization, an explanation for an adaptive penalty effect in ML optimization, an extension of an auction-theory result for AI token allocation, and a new approach to singular integrals in cosmic-string radiation calculations.

There are still obvious caveats. These are Google DeepMind's characterizations of joint work, not a claim that AI has independently replaced researchers, and the company explicitly says it is not claiming the highest levels of breakthrough novelty in its proposed taxonomy for AI-assisted math. Still, the announcement is high-signal because it reframes frontier models as scientific collaborators with verification loops, browsing, and domain-expert oversight. If the results hold up through peer review and wider reproduction, Gemini Deep Think could become one of the clearest examples yet of an AI system moving from contest performance into real research production.