NVIDIA Team Achieves Gold-Medal Performance on International Mathematical Olympiad
SAN FRANCISCO – September 13, 2026 – Researchers at NVIDIA have developed an open-source AI system that achieved gold-medal performance on the 2026 International Mathematical Olympiad (IMO), scoring 30 out of 42 points using only natural language reasoning without formal proof assistants or external tools.
The team, led by Ivan Moshkov, Stephen Ge, George Armstrong, Wei Du, Sadegh Mahdavi, and Igor Gitman, published their findings on arXiv on September 9, 2026, and released the complete technical stack to enable reproducibility.
Natural Language Reasoning Without Formal Tools
The Nemotron IMO pipeline uses NVIDIA's general-availability Nemotron 3 Ultra model as a foundation. Through supervised fine-tuning and reinforcement learning, the researchers created two specialist checkpoints trained specifically on mathematical reasoning. These three checkpoints work together in an iterative generation, verification, and refinement loop that operates entirely within the model's context window.
Unlike previous approaches to AI mathematical reasoning that rely heavily on formal proof assistants like Lean—where a language model generates code that a compiler verifies—the Nemotron system restricts itself to natural language proofs. The model itself serves as the sole arbiter of proof validity during both generation and verification.
This approach demonstrates that sufficiently capable language models can internalize both problem-solving and verification functions that were previously outsourced to symbolic systems.
Three-Checkpoint Architecture
The system employs a division of labor across three model checkpoints:
- Base model: The general-availability Nemotron 3 Ultra
- Specialist checkpoint 1: Fine-tuned for specific mathematical domains
- Specialist checkpoint 2: Additional specialization for harder problem classes
A separate high-compute stage selects the final submission for each problem, choosing the highest-quality proof from multiple candidate generations.
Open Release Enables Reproducibility
In a move unusual for gold-medal-level competition results, the researchers released their complete technical stack. This includes:
- The two specialist model checkpoints
- Full training data
- Training and inference code
- All submitted IMO solutions
- A new 200-problem benchmark called Nemotron-IMO-Bench
The open release addresses a persistent gap in mathematical AI research where competition problems are often reused and potentially contaminated in training data. The Nemotron-IMO-Bench provides a standardized, uncontaminated evaluation surface for future systems.
Implications for Agent Development
The architecture demonstrates several properties characteristic of autonomous agent systems. The iterative search process—generating candidate proofs, evaluating them against internal verification criteria, and refining failures—operates without human intervention during the competition run.
For agent platform developers, the release provides a concrete template for converting general-purpose foundation models into domain-specialist agents through post-training. The combination of supervised fine-tuning and reinforcement learning to produce specialist checkpoints, followed by an orchestrated inference pipeline, generalizes well beyond mathematics.
Performance Context
The 30-out-of-42 score crossed the gold-medal threshold—a bar typically set to distinguish the top-performing contestants in a given year. While the paper does not provide direct score comparisons with other systems at IMO 2026, this represents one of the strongest publicly documented performances by an AI system on olympiad-level mathematics under competition conditions.
Looking Forward
The Nemotron IMO result provides the research community with both a reference implementation for building domain-specialist agents and a new benchmark for measuring reasoning capabilities. External teams can now replicate the training pipeline, evaluate the specialist checkpoints on the benchmark, and adapt the test-time compute framework for other domains.
Sources
Roman Bietin
Research Contributor · Foresight Institute of Research and Translation
Roman Bietin contributes to FIRAT's mission of generating evidence-based research and translating scientific breakthroughs into sustainable African development.
