Gemini 3 Pro solves IMO 2025 P6 with some prompting (no hints or tools involved) (old.reddit.com)

Gemini 3 Pro — Google’s latest large language model — was reported to have solved IMO 2025 Problem 6 using only prompt engineering (no external hints, calculators, or auxiliary tools). Problem 6 is traditionally the hardest, most creative contest problem and a benchmark for deep mathematical reasoning; a model producing a correct solution without tool assistance suggests a notable step forward in long-range symbolic reasoning and the ability to chain multiple nontrivial inferences. Technically, the success appears to hinge on careful prompting to elicit stepwise, proof-oriented reasoning (akin to chain-of-thought/few‑shot strategies) rather than any external theorem-proving backend. The result underscores both opportunities and caution: such models can accelerate discovery, draft complex proofs, and assist mathematicians, but outputs still require formal verification because LLMs can hallucinate plausible yet incorrect steps. Practically, this pushes the community to update benchmarks (hard contest problems, formal proof corpora), invest in automated proof checkers and verification pipelines, and study prompt robustness and failure modes to responsibly integrate LLMs into mathematical research and education.