*"Mathematics is the language with which God wrote the universe."* — Galileo Galilei
---
🎭 Prologue: A Silent Battle
July 2026. Brisbane, Australia.
Inside the exam hall of the International Mathematical Olympiad (IMO), more than 600 young talents from 107 countries hold their breath. Their enemy is not each other, but six nearly impossible math problems.
Meanwhile, in a data center thousands of kilometers away, another "team" is also sitting the exam. This team neither eats, drinks, nor sleeps, and never feels nervous. It is NVIDIA's Nemotron-3-Ultra — a 550-billion-parameter neural network — together with a group of "assistants."
When the results were announced, the whole world went quiet:
30 points. Out of 42. Gold-medal cutoff: 29.
By a single point, Nemotron crossed the line that human math prodigies have contested for sixty-seven years.
This was the first time AI won gold at the IMO. Not bronze, not silver — gold. And it did not rely on guessing or brute-force memorization. It relied on proofs — rigorous, complete mathematical proofs accepted by human graders.
This is the story of how a machine learned to think like a mathematician.
---
🧮 The IMO: Mount Everest of Human Intellect
Founded in 1959, the IMO is the world's most prestigious high-school math competition: six problems over two days, 7 points each, 42 total.
How hard are these problems?
> Imagine: you spent your entire school career at the top of every math class and national competition. Then you sit in the IMO exam hall, stare at the first — the "easiest" — problem for three hours, and walk away with 1 point.
This is not an exaggeration. Each year, roughly half of contestants score 0 or 1 on Problem 1, and Problem 6 — reserved for the rarest talents — often defeats everyone.
The IMO problems have distinct traits:
- No advanced math allowed: calculus, linear algebra, group theory — all banned. Only secondary-school math: algebra, geometry, number theory, combinatorics.
- Extreme creativity required: solutions often demand a clever auxiliary construction, a hidden number-theoretic property, or an elegant induction.
- Rigorous proofs required: a number or yes/no answer is far from enough. Every logical step must be airtight.
- Three models (GA, RL, SFT) each generate proof attempts.
- Each model uses 8 different strategies (e.g., "lemma-first proof," "compare paths," "seek counterexamples").
- Each strategy attempts 16 times.
- Total: 384 proof attempts.
- Two dedicated verifier models (RL and SFT) independently review each proof.
- Each proof receives 16 scores (2 verifiers × 8 independent judgments).
- Scoring: 0 (fully wrong), 0.5 (partially correct), 1 (fully correct).
- A proof is accepted only if all 16 scores are 1.
- If round 1 yields no perfect proof, the system picks the top 16 "semi-finished" proofs.
- Verifiers provide detailed revision suggestions for each.
- Generators make 192 improvement attempts per semifinalist.
- Improved proofs re-enter verification.
- Three models score it with IMO-style grading (0–7).
- Each receives 48 scores (3 models × 16 judgments).
- Rank by average; ties go to the shorter proof.
- 2.31 billion tokens ≈ reading roughly 15 million pages of text.
- 4,785 GPU-hours ≈ one top GPU running continuously for nearly 200 days.
- SFT performs best in round 1 (70), showing specialized proof training works.
- RL performs best end-to-end (180), showing RL optimizes overall strategy.
- But the ensemble far outperforms any single model (188 vs. 180), showing the three are highly complementary.
- Weights of all three models
- All training data
- All code
- A 200-problem Nemotron-IMO-Bench evaluation set
- NVIDIA Nemotron Team (2026). *An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics*. arXiv:2609.10712.
- OpenAI (2024). *Learning to Reason with LLMs*. OpenAI Blog.
- DeepSeek-AI (2025). *DeepSeek-V4 Technical Report*. arXiv:2505.10047.
- Trinh, T. H., et al. (2024). Solving olympiad geometry without human demonstrations. *Nature*, 625, 476-482.
- Polu, S., & Sutskever, I. (2020). Generative language modeling for automated theorem proving. *arXiv preprint arXiv:2009.03393*.
- Welleck, S., et al. (2022). NaturalProofs: Mathematical theorem proving in natural language. *arXiv preprint*.
That is exactly why the IMO is the ultimate touchstone for AI. Chess? AI crushed humans long ago. Coding? AI writes decent code. But creative mathematical proof demands something close to intuition — the ability to see a path through the fog.
---
🤖 Nemotron: Three "Mathematicians" in One
Nemotron's IMO system is not a lone genius. It combines three "experts":
| Expert | Identity | Strength | |-----|------|------| | GA (General Availability) | Generalist | Broad knowledge base, like a well-read math professor | | SFT (Supervised Fine-Tuning) | Proof specialist | Trained on 410K proof problems, like a solving machine | | RL (Reinforcement Learning) | Strategy expert | RL-optimized, like a competition-savvy coach |
All three share the same base architecture (Nemotron-3-Ultra, 550B parameters) with different "postgraduate training."
SFT: Swimming in an Ocean of Proofs
The SFT training data came from a remarkable pipeline:
1. Collect nearly 16,000 hard problems from AoPS (Art of Problem Solving), the world's largest math competition community. 2. Use DeepSeek-V4-Pro (another powerful AI) to generate proof attempts. 3. Iterate failed proofs up to 3 rounds of "polishing" — analyzing what went wrong and how to fix it. 4. Introduce "verifiers" and "meta-verifiers": judging not only whether proofs are correct, but whether the judges themselves are reliable. 5. Finally curate 414,890 high-quality proof problems for SFT training.
It is like having a new doctor review 410,000 case files, each double-checked by senior experts, before starting practice.
RL: Learning from Mistakes
The RL training is even more interesting. Researchers picked 9,597 problems of "just enough" difficulty — ones the GA model solves 1–3 times out of 4 attempts — and used reinforcement learning to let the model try, fail, and retry.
The key is reward design. Researchers deliberately omitted a "self-analysis reward": the model gains no bonus for "reflecting on its mistakes." This prevents the model from learning to fake thinking to farm rewards.
---
🏗️ Exam Strategy: Not a Student, but a Research Institute
Where traditional AI problem-solving is "one student grinding away," Nemotron operates like "a full research institute collaborating."
Phase 1: Exhaustive Search (up to 8 rounds)
For each problem, Nemotron runs up to 8 rounds of "generate–verify–improve":
Round 1: Full firepower
Verification: strict peer review
Imagine a paper requiring unanimous approval from 16 independent reviewers.
Rounds 2–8: Polishing
Phase 2: Final Selection
After up to 8 rounds, the system picks up to 3 "finalists" from all accepted proofs.
Each finalist undergoes "final judgment":
---
📊 The 30-Point Story
Nemotron's final IMO 2026 score: 30/42, one point above the gold cutoff (29).
| Problem | Score | Round found | Compute used | |-----|------|-------------|-----------| | 1 | 7 | Round 1 | 5.57M tokens | | 2 | 7 | Round 2 | 227M tokens | | 3 | 1 | Round 1 | 106M tokens | | 4 | 7 | Round 1 | 20.4M tokens | | 5 | 7 | Round 1 | 36.4M tokens | | 6 | 1 | Round 2 | 312M tokens |
Highlights and Regrets
Perfectly solved (Problems 1, 2, 4, 5): all four complete proofs were found within the first two rounds, mostly round 1. Problems 1 and 4 needed relatively little compute — Nemotron handles "medium difficulty" with ease.
The drama of Problem 2: the full proof came only in round 2, costing 227 million tokens — 40× Problem 1. This hints Problem 2 required some "twist" or "insight" that brute search alone couldn't reach.
The regrets of Problems 3 and 6: the two hardest problems earned only 1 point each. Notably, Nemotron's internal verifiers overestimated its own performance — it believed it had scored more, but human graders gave just 1 point. At the hardest difficulty, AI self-assessment still has blind spots.
Interestingly, when searching continued after the official 4.5-hour exam, Nemotron found a better solution to Problem 6, unofficially graded 4 points by human judges — suggesting a potential total of 33 with more time.
---
💰 The Cost of Compute: An Expensive Exam
| Metric | When proof found | Full search | |-----|-----------|---------| | Tokens | 707M | 2.31B | | GPU-hours | 1,464 | 4,785 |
For perspective:
At market prices, the compute cost was likely in the tens to hundreds of thousands of dollars.
But that is the point:
> Nemotron didn't win gold by being "smarter," but by "thinking longer" and "verifying more strictly."
It is like giving a student unlimited scratch paper, unlimited attempts, unlimited peer review — and a brain that never tires or loses confidence. Under such conditions, even a non-genius can solve genius-level problems.
---
🔬 Ablation: Three Cobblers Equal One Zhuge Liang
What if only one of the three models were used instead of the ensemble?
| Model | Round-1 score | Accepted proofs | Total | |-----|----------|---------------|------| | GA | 34 | 23 | 162 | | RL | 47 | 23 | 180 | | SFT | 70 | 21 | 165 | | Ensemble (actual) | 91 | 24 | 188 |
Striking results:
Like a research team: some generate ideas, some prove rigorously, some plan strategy. Together they beat any soloist.
---
🌟 Significance: The Birth of the AI Mathematician
Nemotron's IMO gold is more than a competition result. It marks several milestones:
1. From "Pattern Matching" to "Logical Reasoning"
Traditional LLMs (like GPT-4) struggle in math because they mostly "match patterns" — recalling similar problems from training data. But IMO problems are original, with no standard solutions to memorize.
Nemotron's success shows that with specialized training (proof data + RL) and systematic search (generate–verify–improve loops), AI can begin genuine logical reasoning rather than mere pattern matching.
2. The Power of Test-Time Compute
Nemotron's core innovation is not architecture but test-time compute strategy: 384 attempts, 16-way verification, 8 improvement rounds per problem. In a human exam this would be cheating; in AI, it is entirely legal.
> Future AI capability may come not just from training-time "knowledge," but from inference-time "depth of thought."
Just as human mathematicians may ponder a hard problem for weeks, AI can invest massive computation at test time to "think deeply."
3. A Model of Open Science
The Nemotron team made an admirable decision: fully open.
Any researcher can reproduce, verify, and improve the work — precious openness in an increasingly closed AI landscape.
4. The Verification Dilemma and Opportunity
The "self-overestimation" on Problems 3 and 6 reveals a deep issue: how can AI judge whether its own proofs are correct?
At the hardest problems, even AI verifiers develop blind spots — echoing human mathematicians' experience: the hardest errors are not arithmetic mistakes but subtle logical gaps in arguments that "look right."
Solving verification may be the key to stronger AI mathematicians.
---
🌌 Epilogue: When Machines Begin Proving Theorems
Two thousand years ago, Euclid wrote the first mathematical proofs in the *Elements*, opening an epic of human intellect.
Two thousand years later, a machine sat in the IMO's "virtual exam hall" and, through billions of computations and hundreds of attempts, wrote its own first gold-medal proof.
This is not the end. Nemotron's 30 points crossed the gold line but remain far from 42. Problems 3 and 6 still loom like unconquered peaks awaiting the next breakthrough.
But history will remember this moment:
In 2026, artificial intelligence won gold at the International Mathematical Olympiad for the first time.
Not because it was smarter than humans, not because of some mysterious mathematical intuition, but because it learned to search systematically, verify rigorously, and improve relentlessly — precisely how human mathematicians have worked for centuries.
Perhaps the true mathematical genius is not the one who sees the answer at a glance, but the one willing to grope in the dark, learn from errors, and persist through countless failures.
In that sense, Nemotron is not just a machine.
It is a mathematician who never gives up.
---