Redefining the Future of Scientific Research – Google DeepMind

Working with experts on 18 research problems, the advanced version of Gemini Deep Think has helped solve long-standing problems across algorithms, ML and collective optimization, information theory, and economics. Highlights from our paper “Accelerating Research with Gemini” include (corresponding section numbers in the paper):

1. Transcending the mathematical boundaries of network puzzles: Progress on classic computer science problems such as “Max-Cut” (fine-sorting networks) and “Steiner Tree” (connecting high-dimensional points) has slowed down. Gemini broke both boundaries by thinking outside the box. It solved these algorithmic puzzles by drawing advanced tools—such as the Kirszbraun Theorem, approximation theory, and Stone-Weierstrass theorem—from completely unrelated branches of continuous mathematics. See Sections 4.1 and 4.2.
2. Correcting the prediction of the decade in online submodular optimization: A 2015 theoretical paper proposed a seemingly obvious law of data transmission: making a copy of an incoming object is always less important than simply transmitting the original. Experts struggled for ten years to prove this. Gemini makes a clear example of the three things combined, firmly proving that long-held human assumptions are false. See Section 3.1.
3. Improving machine learning: Training AI to filter noise often requires developers to manually tune the “fineness” of the equation. The researchers created a new mechanism that does this automatically, but they couldn't mathematically explain why. Gemini analyzed the statistics and showed that the method is effective by making its own “flexible penalty” in secret. See Section 8.3.
4. Developing an economic theory of AI: The recent 'Revelation System' of the AI ​​generation token sale only worked with math where bids were limited to logical numbers. Extending the domain to continuous real numbers is not valid real proof. Gemini used advanced topology and order theory to extend the theory, adapting it to real-world, continuous auction dynamics. See Section 8.4.
5. Physics of cosmic strings: Calculating gravitational waves in cosmic strings requires finding analytical solutions to tricky equations that contain “singularities.” Gemini found a novel solution using Gegenbauer polynomials. This naturally absorbed the singularity, folding the infinite series into a closed form, a finite sum. See Section 6.1.

Covering a variety of disciplines—from information and complexity theory to cryptography and machine design—the results show how AI is changing research. For details, see our paper.

Given computer science's fluid, conference-driven publication pipeline, we interpret these results in an academic vein rather than a rigid taxonomy. About half of the targeted hard conferences—including the ICLR '26 reception—are where most of the remaining findings will form future journal submissions. Even when the study is corrected in the field by identifying errors (Section 3.2) or perceived objections (Section 3.1), these results highlight the value of AI as a high-level scientific collaborator.

The Future of Human-AI Collaboration

Building on Google's previous developments (1, 2, 3, 4, 5), this work shows that common ground models – derived from the flow of the agent's reasoning – can serve as a powerful scientific companion.

Under the guidance of mathematicians, physicists, and computer scientists, the Gemini Deep Think mode proves its use in fields where complex math, logic, and reasoning are at the core.

We are witnessing a fundamental change in the flow of scientific work. As Gemini evolves, it acts as a “power multiplier” for human intelligence, handling information retrieval and rigorous validation so that scientists can focus on deeper concepts and creative direction. Whether it's refining evidence, hunting down counterexamples, or connecting disconnected fields, AI is becoming a key participant in the next chapter of scientific progress.

Thank you

We thank the community of mathematicians, physicists, and computer scientists for their support of this project.

This project was a huge collaboration across Google and its success is due to the combined efforts of many people and teams. Thang Luong and Vahab Mirrokni led all research directions with deep technical expertise from Tony Feng and David Woodruff.

Authors of the first paper “Towards Autonomous Mathematics Research” include: Tony Feng, Trieu H. Trinh, Garrett Bingham, Dawsen Hwang, Yuri Chervonyi, Junehyuk Jung, Joonkyung Lee, Carlo Pagano, Sang-hyun Kim, Federico Pasqualotto, Sergei Guiu Goiu, Jungsu Hook Ghiasi, Yi Tay, YaGuang Li, Chenkai Kuang, Yuan Liu, Hanzhao (Maggie) Lin, Evan Zheran Liu, Nigamaa Nayakanti, Xiaomeng Yang, Heng-tze Cheng, Demis Hassabis, Koray Kavukcuoglu, Quoc V. Le, Thang Luong. We thank the following experts for feedback and discussions on the work: Jarod Alper, Kevin Barreto, Thomas Bloom, Sourav Chatterjee, Otis Chodosh, Michael Harris, Michael Hutchings, Seongbin Jeon, Youngbeom Jin, Aiden Yuchan Jung, Jiwon Kang, Jimin Kim, Vjekoslav Kovač, Daniel Manoles More, Acup Mona, Acup Schildkraut, Johannes Schmitt, Insuk Seo, Jaehyeon Seo, Cheng-Chiang Tsai, Ravi Vakil, Zhiwei Yun, Shengtong Zhang, Wei Zhang, Yufei Zhao

Authors of the second paper “Accelerating Scientific Research with Gemini: Case Studies and Common Strategies” include David P. Woodruff, Vincent Cohen-Addad, Lalit Jain, Jieming Mao, Song Zuo, MohammadHossein Bateni, Simina Branzei, Michael P. Brenner, Lin Chen, Yingyih Fung, Hadide, Hadiza, Ying Fort Feng, Hadidi Mohammad T. Hajiaghayi, Mahdi JafariRaviz, Adel Javanmard, Karthik CS, Ken-ichi Kawarabayashi, Ravi Kumar, Silvio Lattanzi, Euiwoong Lee, Yi Li, Ioannis Panageas, Dimitris Paparas, Benjamin Przybocki, Bernardo Subercase, Shamrijan Xucasheux Wu, Eylon Yogev, Morteza Zadimoghaddam, Samson Zhou, Yossi Mathias, Jeff Dean, James Manyika, Vahab Mirrokni. This list includes Google researchers who build agent thinking on top of Gemini, as well as our academic expert collaborators who validate and interact with Gemini. We also thank Corinna Cortes for her careful review of the paper.

We appreciate the basic support from the entire DeepThink team: Anirudh Baddepudi, Michael Brenner, Irene Cai, Kristen Chiafullo, Paul Covington, Rumen Dangovski, Chenjie Gu, Huan Gui, Vihan Jain, Rajesh Jayaram, Melvin Johnson, Rosemary Ke, Maciej Kula, Steven Ponow, Nate Posh More Sidharth Mudgal, William Nelson, Ada Maksutaj Oflazer, Sahitya Potluri, Navneet Potti, Shubha Raghvendra, Siamak Shakeri, Archit Sharma, Xinying Song, Mukund Sundararajan, Qijun Tan, Zak Tsai, Theophane Weber, Winnie Xu, Shubha Raghvendra, Siamak Shakeri, Archit Sharma, Xinying Song, Mukund Sundararajan, Qijun Tan, Zak Tsai, Theophane Weber, Winnie Xu, Shubha Yake, Zicheng Yuju, Zicheng Xu and Honglei Zhuang.

We thank Quoc Le, Koray Kavukcuoglu, Demis Hassabis, James Manyika, Yossi Mattias, and Jeff Dean for supporting this work.

Finally, we thank Divy Thakkar, Adam Brown, Vinay Ramasesh, Alex Davies, Thomas Hubert, Eugénie Rives, Pushmeet Kohli, Benoit Schillings for feedback and project support.