TechPrufer
Get started
DeepMind
DeepMind logoDeepMind · Feature trail

Gemini Deep Think Reasoning Engine

7 posts · Jul 2025 – Mar 2026

What this trail covers

Gemini Deep Think continues to advance its abstract problem-solving capabilities, achieving gold-medal level performance in both the International Mathematical Olympiad and the International Collegiate Programming Contest (ICPC) World Finals. This year, an advanced version of Gemini Deep Think solved five out of the six IMO problems perfectly, earning 35 total points, and achieving gold-medal level performance. The model operated end-to-end in natural language, producing rigorous mathematical proofs directly from the official problem descriptions within the competition time limit. This version of Gemini was trained on novel reinforcement learning techniques that leverage multi-step reasoning, problem-solving, and theorem-proving data, and was provided with access to a curated corpus of high-quality solutions to mathematics problems, along with hints and tips on how to approach IMO problems. A version of this Deep Think model will be made available to trusted testers and then rolled out to Google AI Ultra subscribers.

All DeepMind trails
  1. AlphaGo at 10: How AI Innovation Is Paving the Path to AGIThis post highlights the evolution of Gemini's reasoning capabilities, specifically mentioning the 'Deep Think' mode achieving gold-medal standard performance at the 2025 IMO and its application to complex scientific and engineering challenges. It draws a direct lineage from AlphaGo's search and planning techniques, emphasizing the use of deep neural networks, advanced search, and reinforcement learning as foundational elements for advanced reasoning.
  2. Gemini 3.1 Pro: A smarter model for your most complex tasksThis post announces the release of Gemini 3.1 Pro, an upgraded core intelligence model that enhances reasoning capabilities. It builds upon the Gemini 3 Deep Think model, achieving a verified score of 77.1% on the ARC-AGI-2 benchmark. The release includes availability for developers via the Gemini API (Google AI Studio, Gemini CLI, Antigravity, Android Studio) and for enterprises via Vertex AI and Gemini Enterprise. Consumer access is through the Gemini app and NotebookLM. The post highlights the model's application in complex problem-solving and advanced reasoning.
  3. Gemini 3 Deep Think: Advancing science, research and engineeringThis post details the latest upgrade to Gemini 3 Deep Think, highlighting its enhanced performance on rigorous academic benchmarks like Humanity's Last Exam (48.4% without tools), ARC-AGI-2 (84.6%), Codeforces Elo (3455), and the International Math Olympiad 2025 (gold-medal level). It also showcases advancements in scientific domains, achieving gold medal-level results on the International Physics Olympiad and Chemistry Olympiad (written sections) and scoring 50.5% on CMT-Benchmark for theoretical physics. The post emphasizes the practical application of Deep Think in interpreting complex data and modeling physical systems through code, with availability extended via the Gemini API for select researchers, engineers, and enterprises.
  4. Gemini Deep Think: Redefining the Future of Scientific ResearchThis post details the application of Gemini Deep Think mode, powered by an agent named Aletheia, to solve professional research problems in mathematics, physics, and computer science. It introduces a math research agent with a natural language verifier and Google Search integration for iterative solution generation and revision. The agent has achieved autonomous solutions to four open questions on Bloom's Erdős Conjectures database and contributed intermediate propositions to further papers. In physics and computer science, the 'Advisor' model with 'Vibe-Proving' cycles and 'balanced prompting' techniques are detailed. Specific breakthroughs include solving 'Max-Cut' and 'Steiner Tree' problems by applying concepts from continuous mathematics, refuting a decade-old conjecture in online submodular optimization with a combinatorial counterexample, explaining an adaptive penalty in ML optimization, extending a 'Revelation Principle' for auctioning AI generation tokens to continuous real numbers using topology and order theory, and finding a novel solution to gravitational radiation integrals from cosmic strings using Gegenbauer polynomials. The post also discusses a taxonomy for classifying AI-assisted mathematics research and highlights the value of AI as a high-level scientific collaborator.
  5. A new era of intelligence with Gemini 3Introduces Gemini 3 Deep Think, an enhanced reasoning mode that significantly improves performance on complex problem-solving benchmarks like Humanity's Last Exam and GPQA Diamond, and achieves new state-of-the-art results on ARC-AGI-2 with code execution.
  7. Gemini achieves gold-medal level at the International Collegiate Programming Contest World FinalsThis post details Gemini 2.5 Deep Think's gold-medal performance at the 2025 ICPC World Finals, solving 10 out of 12 problems. It highlights the model's ability to tackle complex algorithmic problems, including a novel solution to 'Problem C' using dynamic programming and nested ternary searches. The post also describes the underlying technical advances in pretraining, post-training, reinforcement learning, multi-step reasoning, and parallel thinking, including the use of multiple agents to propose, execute, and iterate on solutions.
  8. Advanced version of Gemini with Deep Think officially achieves gold-medal standard at the International Mathematical OlympiadThis post details the technical advancements in Gemini Deep Think that enabled it to achieve gold-medal performance at the International Mathematical Olympiad. Key contributions include: operating end-to-end in natural language without requiring translation to formal languages, achieving this within the competition time limit, and incorporating parallel thinking to explore multiple solutions simultaneously. The model was further trained using novel reinforcement learning techniques for multi-step reasoning and theorem-proving, and was augmented with a curated corpus of high-quality mathematical solutions and problem-solving hints.