What does correctness of outputs tell us about reasoning?

Google DeepMind claims that an advanced version of Gemini Deep Think solved five of the six problems at the 2025 International Mathematical Olympiad, earning 35 of 42 points, a gold-medal score. The result is the lab's own claim, but the excerpt says the answers were graded by IMO coordinators "using the same criteria as for student solutions," and that the IMO "confirmed that our submitted answers are complete and correct solutions." That is the verified part: the proofs were checked by the competition's own graders. The only comprehension-like statement is that the graders found the solutions "clear, precise and most of them easy to follow." That describes readability. The excerpt does not say the reasoning was explained or understood in human terms.

The excerpt credits the result to a reasoning mode that "simultaneously explore[s] and combine[s] multiple possible solutions before giving a final answer, rather than pursuing a single, linear chain of thought." The model was also trained with novel reinforcement learning on multi-step reasoning and theorem-proving data, given a curated corpus of solutions, and given general hints in its instructions. The excerpt contrasts this with 2024, when AlphaGeometry and AlphaProof needed problems translated into Lean and two to three days of computation. This year the model worked end-to-end in natural language within the 4.5-hour limit.

The parallel-exploration mechanism is a different axis from the result in Does more thinking time always improve reasoning accuracy?, which concerns what happens when a single chain is made longer. The excerpt attributes its gains to breadth of exploration, not length, and gives no measurement that would test the non-monotonic finding. The certification also bears on Do automated benchmarks hide what frontier AI systems can really do?. An expert-graded competition is a stronger check than an automatically graded benchmark, but it still stops at the output. The accuracy of a result says little about how it was reached, which is the gap Can we measure how deeply a model actually reasons? tries to close from inside the model.

The excerpt does not establish how the system reached its proofs, and the lab says so itself: the IMO's "review does not extend to validating our system, processes, or underlying model." Five correct proofs on one competition's problems show that the outputs were correct on that occasion. They do not show that the training data, the hints or the parallel procedure produced the capability, or that it transfers beyond olympiad problems. At the strength the evidence allows, "correct" holds for these proofs, while "understood" does not yet hold for the system that produced them, and the public claim should be read at that strength.