Double Gold: Specializing NVIDIA Nemotron for IOI and IMO 2026
雙金入袋！NVIDIA 透過微調 Nemotron 征服 2026 國際奧賽數理與資訊殿堂
NVIDIA demonstrated how a four-step specialization recipe can transform Nemotron-3 base models into domain experts. At IOI 2026, the 550B-parameter Nemotron-3-Ultra-CC scored 535.4/600, surpassing both the gold threshold and the top human score. At IMO 2026, the specialized system scored 30/42, beating the gold threshold using pure natural language reasoning without external provers. By co-designing model tuning, high-quality data, and test-time search loops (GenCorrect), they achieved frontier-level problem solving, now fully open-sourced.
Key points
The Specialization Recipe
Rather than building new foundations, the team combined strong base models with curated data, post-training (SFT/RL), and a 'generate-evaluate-refine' inference loop.
Tuning Dynamics at Scale
For the 30B Nano model, combining SFT and RL yielded steady gains. For the 550B Ultra model, just one epoch of SFT outperformed a fully-tuned Nano model.
Natural Language Reasoning
The IMO system used no formal provers, external tools, or web access, relying entirely on natural language proof generation, critiquing, and refining.
Test-Time Co-Design
Gold results relied on combining capable specialized models with structured test-time search and verification loops (such as GenCorrect) during inference.
How it works
| IOI 2026 (程式/資訊) | IMO 2026 (數學) | |
|---|---|---|
| Specialized Models | Nemotron-3-Ultra-CC (550B SFT) | Nemotron 3 Ultra (SFT 與 RL 組合) |
| Score achieved | 535.4 / 600 (金牌門檻 361.12) | 30 / 42 (金牌門檻 29) |
| Key Strategy | GenCorrect 多輪迭代修正迴圈 | 生成-驗證-重構 多模型協同審查 |
| External Tools | 有（程式編譯與回饋） | 無（完全基於自然語言推理） |
Why it matters
This achievement proves that open-weight base models can reach human-olympiad-level performance without retraining from scratch. By using a structured, reproducible specialization recipe and test-time search, enterprises and researchers can replicate this success in other complex, high-reasoning domains like software engineering and scientific research.
Who it affects
- AI Developer
- AI Researcher
- Product Manager
- Enterprise Leader
How to use it
- 1Advanced Coding & Debugging: Employing multi-round 'generate-evaluate-refine' (GenCorrect) workflows for automated software engineering and complex algorithm synthesis.
- 2Complex Logic & Academic Proofing: Generating and self-verifying rigorous arguments in high-precision fields like math, legal analysis, or finance using SFT and meta-verification.
Limitations & caveats
- Unofficial & Unsupervised Runs: The IOI benchmark was a simulated prospective run under competition constraints, not an official supervised entry.
- High Test-Time Compute Overhead: Achieving gold-level performance relies heavily on multi-round search and refinement, increasing latency and operational compute costs during inference.
Related
Building Persistent 3D Object Memory: How Ledger Tracks Objects from Egocentric Videos
打造過目不忘的 3D 空間記憶：Ledger 如何透過第一人稱影片追蹤隱形物體
Researchers introduce Ledger, a framework that builds a persistent 3D object memory from egocentric videos, significantly improving spatial question-answering accuracy for embodied agents.
Decoupling Exploration from Optimization: How ExpDis Boosts LLM Reasoning and Solution Diversity
探索與優化解耦：全新強化學習框架 ExpDis 提升大語言模型的推理多元性
The ExpDis framework decouples exploration from optimization in RLVR. By training explorers with novelty bonuses and distilling filtered trajectories into a student model, it prevents model degradation while fostering diverse reasoning.
Clipped Decentralized SGD: Achieving Optimal Convergence and Linear Speed-Up Under Heavy-Tailed Noise
去中心化 SGD 克服重尾雜訊：梯度裁剪如何實現最佳收斂與線性加速
This study proves that clipped decentralized SGD (DSGD) achieves order-optimal convergence rates and linear speed-up under heavy-tailed noise for non-convex optimization.