Gemini DeepMind AI Solves Complex Problems in Math & Science

Google DeepMind is demonstrating a significant leap in the application of artificial intelligence to fundamental scientific research. A new paper, “Accelerating Research with Gemini,” details how an advanced version of the Gemini model, operating in a “Deep Think” mode, has helped resolve longstanding problems across a diverse range of disciplines, including mathematics, computer science, physics, and economics.

The work builds on Google’s previous AI-driven breakthroughs in mathematical problem-solving, including advancements with AlphaProof, AlphaGeometry 2, FunSearch, and AlphaEvolve. This latest effort moves beyond simply solving existing problems to actively accelerating the research process itself, acting as a “force multiplier” for human intellect, according to the researchers.

Breaking Mathematical Deadlocks

One striking example highlighted in the paper involves progress on two classic computer science problems – “Max-Cut” and the “Steiner Tree.” Both had seen limited progress for years. Gemini, however, bypassed these roadblocks by drawing on concepts from seemingly unrelated areas of continuous mathematics, specifically the Kirszbraun Theorem, measure theory, and the Stone-Weierstrass theorem. This interdisciplinary approach allowed the AI to find novel solutions.

The Gemini model also settled a decade-old conjecture in online submodular optimization. A 2015 theory paper proposed that copying an arriving item in a data stream is always less valuable than moving the original. Despite its intuitive appeal, proving this claim proved elusive for ten years. Gemini engineered a specific three-item counterexample, definitively disproving the conjecture.

Optimizing Machine Learning and Economic Theory

The impact extends beyond pure mathematics. Researchers developed a new technique for automatically tuning a mathematical “penalty” used in machine learning to filter out noise. While the method worked, its underlying principles remained unclear. Gemini analyzed the equations and demonstrated that the technique implicitly generates its own “adaptive penalty” during operation, providing a mathematical explanation for its success.

In the realm of economic theory, Gemini addressed a limitation in a recent ‘Revelation Principle’ for auctioning AI generation tokens. The original theorem only held true when bids were restricted to rational numbers. Extending the theorem to encompass continuous real numbers – a more realistic scenario – invalidated the proof. Gemini leveraged advanced topology and order theory to extend the theorem, making it applicable to real-world auction dynamics.

Solving Problems in Physics

Even the complexities of theoretical physics yielded to Gemini’s capabilities. Calculating gravitational radiation from cosmic strings involves solving intricate integrals containing “singularities.” Gemini discovered a novel solution using Gegenbauer polynomials, effectively absorbing these singularities and transforming an infinite series into a finite sum.

A New Scientific Workflow

The researchers emphasize that these results aren’t simply about AI solving problems independently. Instead, Gemini is being used as a collaborative tool, guided by human experts. The model handles knowledge retrieval and rigorous verification, freeing scientists to focus on conceptual breakthroughs and creative direction. About half of the findings are targeted for presentation at strong conferences, including an acceptance at ICLR ’26, with the remainder slated for future journal submissions.

The team also highlights the value of AI in identifying errors and refuting existing conjectures, demonstrating its role not just as a problem-solver but as a critical evaluator of scientific work. This represents a departure from traditional research, where the focus is often on confirming existing theories.

The Future of Human-AI Collaboration

The project involved a large-scale collaboration across Google, with leadership from Thang Luong and Vahab Mirrokni, and deep technical expertise from Tony Feng and David Woodruff. The authors of the papers involved a broad range of researchers, including mathematicians, physicists, and computer scientists, alongside Google researchers building the agentic reasoning workflows on top of Gemini.

The researchers acknowledge the support of the broader scientific community and express gratitude to the experts who provided feedback and discussions throughout the project. Funding for the project was provided by Quoc Le, Koray Kavukcuoglu, Demis Hassabis, James Manyika, Yossi Matias, and Jeff Dean.

This work represents a fundamental shift in the scientific workflow. As Gemini continues to evolve, it promises to become an increasingly valuable collaborator in the pursuit of scientific progress, handling the tedious aspects of research and allowing human scientists to focus on the most challenging and creative aspects of their work.