Member-only story
Reproducing Google DeepMind IMO 2025 Results Using Gemini Pro: A Prompting Strategy Analysis
What if we could reproduce the IMO 2025 results using Gemini Pro with just prompting strategies no web tools or sophisticated RL methods.
I think we’ve all heard the news about how Google DeepMind conquered the 66th International Mathematical Olympiad (IMO) in Australia using its Gemini Deep Think AI model. The system solved 5 out of 6 problems, earning 35 out of 45 points and securing a gold medal. This sent a massive shockwave through the worlds of mathematics and AI. The event featured six notoriously difficult problems in algebra, combinatorics, geometry, and number theory.
They used a steroid version of Gemini 2.5 Pro operating in “Deep Think” mode, which adds enhanced reasoning capabilities such as parallel thinking, multi‑step reinforcement learning, and exposure to a curated corpus of high‑quality mathematical proofs.
But… what if we could reproduce these results using the standard Gemini Pro without the fancy multi-step reinforcement learning by relying solely on prompting strategies? We’ll explore Chain-of-Thought (CoT), Role-Based, and a Hybrid approach that combines both CoT and Role-Based prompting. For the sake of simplicity, we’ll…
