Meta's AIRA₃ Takes Kaggle Gold With Controller-Free Agent Swarm

Meta says AIRA3, its autonomous AI research system, placed 8th of about 4,000 teams and won a gold medal in a live Kaggle competition run by NVIDIA. Kaggle hosts public machine-learning contests in which teams submit solutions and are ranked. In this one, every team had to fine-tune a 30-billion-parameter Nemotron model, NVIDIA's open model family, so that it reasoned better. According to Meta's thread, all competitors had the same information and were graded externally on a private test set, which means the final ranking came from questions no team could see or tune against. Meta reads the result as "a reliable signal" that AIRA3 can improve a targeted capability of an AI model at a level similar to human experts.

The opening post does not say which models AIRA3 ran on or how it organizes its work. It calls AIRA3 the next generation of Meta's autonomous AI research system and says it outperformed human competitors "who had access to the same frontier tools." Meta described the system's makeup in later posts in the same thread, which were not available in the evidence for this note, so the architecture behind the entry is not covered here.

That makes the result one externally graded data point on one fine-tuning problem. The ranking was scored by an outside party on a private test set, but from the opening post alone it is not possible to say how much of the result reflects AIRA3's own methods rather than the frontier tools that, by Meta's account, the human teams also had.