AI
AI system scores higher than top human on International Olympiad in Informatics
Researchers trained specialized language models with synthetic reasoning and reinforcement learning to achieve record performance on competitive programming's most prestigious competition.
Researchers have developed an AI system that surpassed the highest-scoring human contestant at the International Olympiad in Informatics (IOI) 2026, marking a significant milestone in machine reasoning applied to competitive programming.
Training approach#
The work combined a multi-stage pipeline designed specifically for competitive coding challenges. The team curated a collection of 22,000 programming problems to serve as training material. They then enhanced the models' reasoning capabilities through synthetic reasoning traces and applied supervised fine-tuning followed by reinforcement learning.
Two model variants were created:
- Nemotron-3-Nano-CC (30B parameters): trained with both supervised fine-tuning and reinforcement learning
- Nemotron-3-Ultra-CC (550B parameters): trained with supervised fine-tuning alone
Test-time optimization strategy#
The researchers introduced GenCorrect, a novel strategy that operates at test time by generating multiple solution attempts, evaluating them against test cases, and iteratively refining the most promising approaches. This technique substantially boosted performance beyond the base models alone.
Performance benchmarks#
On IOI 2025, the improvements were dramatic. Before post-training, Nano-CC scored just 130 points. After the specialized training process, it reached 291 points, and with GenCorrect applied, it climbed to 468—exceeding the gold medal threshold of 438.3 points. The larger Ultra-CC model achieved 502 points.
When evaluated on IOI 2026 under real competition conditions—matching time limits, no internet access, and standard submission rules—the system scored 535.4 out of 600 total points. This exceeded both the gold medal threshold of 361.12 and the highest human score of 498.27.
Significance#
According to the researchers, this represents the first instance of an AI system outscoring the top human competitor on a complete IOI problem set, demonstrating that large language models can be effectively specialized for complex algorithmic reasoning through targeted training and inference-time strategies.
