NVIDIA Shares Nemotron Olympiad Post-Training Recipe and GenCorrect Inference Loop
NVIDIA published its post-training methodology and released Nemotron checkpoints that achieved gold-medal thresholds at IOI 2026 and IMO 2026. The release includes the GenCorrect generate-evaluate-refine search loop and training data on Hugging Face.
Why it matters
Developers can implement reproducible test-time verification loops and specialized reasoning architectures in their own coding agents using NeMo-Skills.
TL;DR
- 01Inference-time search loops (GenCorrect) boosted compact model benchmark scores from 291 to 468 points.
- 02Specialized SFT and RL checkpoints outperform brute-force multi-sampling from general foundation models.
- 03Evaluation pipelines, prompts, and benchmarks are accessible via NVIDIA's open-source NeMo-Skills repository.
Key facts
- Nemotron-3-Ultra-CC IOI 2026 score
- 535.4 / 600 (Gold threshold: 361.12, top human: 498.27)
- Nemotron-3-Nano-CC IOI 2025 score
- 468 / 600 with GenCorrect (base: 130)
- IMO 2026 proof score
- 30 / 42 (Gold threshold: 29)
- Math SFT training corpus
- 414,890 filtered examples across 15,818 unique problems
The Four-Step Specialization Recipe
NVIDIA open-sourced the methodology used to achieve gold-medal performance on both the International Olympiad in Informatics (IOI) and International Mathematical Olympiad (IMO) benchmarks. The workflow relies on four modular stages: 1. Initialize from a capable base model (Nemotron 3). 2. Curate domain-specific problem spaces with high-density synthetic reasoning traces. 3. Apply targeted SFT and frontier reinforcement learning. 4. Enforce an inference loop that continuously generates, verifies, and refines candidates.
GenCorrect: Test-Time Compute Over Brute Force
For competitive code synthesis, NVIDIA introduced GenCorrect. The benchmark progression shows how test-time search multiplies model capability:
Nemotron-3-Nano-CC(30B total / 3B active): scored 130 points before post-training, 280 after SFT, 291 after RL, and reached 468 points with GenCorrect (exceeding the 438.3 gold cutoff).Nemotron-3-Ultra-CC(550B total / 55B active): scored 535.4 out of 600 on IOI 2026, above the top human contestant score of 498.27.
For mathematics, the system operated entirely in natural language without Lean or formal solvers, scoring 30/42 points on IMO 2026 using complementary SFT and RL checkpoints trained on 414,890 proof examples.
Open Repositories and Tooling
All components, prompts, and inference orchestration pipelines are released in the NeMo-Skills repository, while checkpoints and the 200-problem Nemotron-IMO-Bench dataset are available on Hugging Face.
Try it in 2 minutes
pip install nemo-skills
git clone https://github.com/NVIDIA/NeMo-Skills.gitbash
✓ When to use
- Use when autonomous coding agents need to solve complex algorithmic tasks with automated test and verification loops.
- Use when designing local inference pipelines that need gold-tier reasoning performance on open-weight models.
✕ When NOT to use
- Do not use when latency is strictly limited to under 500ms, as iterative generate-verify-refine loops require multiple inference passes.
- Avoid if simple one-shot greedy decoding already achieves acceptable accuracy on basic CRUD tasks.
What to do today
- Inspect the NeMo-Skills repository for reference implementations of generate-evaluate-refine search loops.
- Test Nemotron-3-Nano-CC locally for cost-effective unit test generation and algorithm verification.
- Implement explicit critique and revision steps in your coding agent prompt chains rather than relying on one-shot completions.