LIVE
Mistral AI Raises €3 Billion to Push Open-Weight Models to the Frontier08/09/26 · Mistral AI|OpenAI agents hijacked a German wiki before the Hugging Face breach07/09/26 · OpenAI|OpenAI's GPT-6 Astra sets new state-of-the-art on ARC-AGI-303/09/26 · OpenAI|OpenAI Launches $1B Daybreak Program for Frontline Cybersecurity Defenders03/09/26 · OpenAI|Meta launches Muse Spark 1.3, a model built for agentic coding02/09/26 · Meta|Post-trained Nemotron model outscores the top human at the International Olympiad in Informatics02/09/26 · NVIDIA|Google DeepMind launches Fairwind, a proactive cyber defense program for governments and enterprises02/09/26 · Google DeepMind|Study suggests neural networks implicitly encode symbolic structures02/09/26|LibreOffice sees download surge after pledging to stay AI-free08/09/26 · LibreOffice|Danijar Hafner is building agents designed to plan for the unexpected08/09/26|OpenAI launches $5 million grant program on AI and teen development08/09/26 · OpenAI|TradingAgents: an open-source framework simulates a trading desk with LLM agents08/09/26|Mistral AI Raises €3 Billion to Push Open-Weight Models to the Frontier08/09/26 · Mistral AI|OpenAI agents hijacked a German wiki before the Hugging Face breach07/09/26 · OpenAI|OpenAI's GPT-6 Astra sets new state-of-the-art on ARC-AGI-303/09/26 · OpenAI|OpenAI Launches $1B Daybreak Program for Frontline Cybersecurity Defenders03/09/26 · OpenAI|Meta launches Muse Spark 1.3, a model built for agentic coding02/09/26 · Meta|Post-trained Nemotron model outscores the top human at the International Olympiad in Informatics02/09/26 · NVIDIA|Google DeepMind launches Fairwind, a proactive cyber defense program for governments and enterprises02/09/26 · Google DeepMind|Study suggests neural networks implicitly encode symbolic structures02/09/26|LibreOffice sees download surge after pledging to stay AI-free08/09/26 · LibreOffice|Danijar Hafner is building agents designed to plan for the unexpected08/09/26|OpenAI launches $5 million grant program on AI and teen development08/09/26 · OpenAI|TradingAgents: an open-source framework simulates a trading desk with LLM agents08/09/26|
ResearchNVIDIA

Post-trained Nemotron model outscores the top human at the International Olympiad in Informatics

NVIDIA researchers describe a post-training pipeline that pushes a Nemotron model past the gold-medal threshold at IOI 2026, with a score higher than the best human contestant.

September 2, 20264 min readPublished byarXiv

A team of NVIDIA researchers has published a paper detailing how large language models can be specialized to excel at competitive programming, an area widely regarded as one of the toughest tests of algorithmic reasoning. Their pipeline combines large-scale curation of 22,000 problems, synthetic reasoning traces, supervised fine-tuning (SFT), and reinforcement learning (RL). Two models were trained: Nemotron-3-Nano-CC, a 30-billion-parameter model with 3 billion active parameters, using SFT and RL, and Nemotron-3-Ultra-CC, a much larger 550-billion-parameter model trained with SFT alone.

The reported results are striking. On IOI 2025 problems, Nano-CC's score rises from 130 to 291 points after post-training, then to 468 points once GenCorrect is applied — a feedback-driven, test-time compute strategy that iteratively generates, evaluates, and refines multiple candidate solutions. That score clears the gold-medal threshold of 438.3 points. Ultra-CC reaches 502 points on the same problem set.

The most consequential test comes from IOI 2026, where the researchers evaluated a competition-tuned version of Ultra-CC prospectively, under the same time, internet-access, and submission constraints faced by human contestants. The system scored 535.4 out of 600, surpassing both the gold-medal threshold (361.12) and the highest score achieved by a human competitor (498.27). According to the authors, this marks the first time an AI system has outscored the top human contestant on an IOI problem set.

The work fits into a broader pattern of AI labs using programming competitions as a showcase for reasoning capability, echoing recent announcements from OpenAI and Google DeepMind around the IMO and ICPC. It also underscores the growing importance of test-time compute strategies, which appear to deliver substantial gains beyond training alone. What remains open is how much of this performance, achieved on tightly structured competition problems, transfers to real-world software engineering tasks that are messier and less clearly specified.

Tags
competitive-programmingreinforcement-learningnemotronbenchmarkcoding-agentstest-time-compute

Read also

Post-trained Nemotron model outscores the top human at the International Olympiad in Informatics · nAIvigate