Meta Wins Gold Medal in Kaggle Competition with Autonomous AI Research System AIRA₃

Meta's AI research organization, AI at Meta, unveiled its next-generation autonomous AI research system AIRA₃. AIRA₃ competed in the Kaggle live competition hosted by NVIDIA in June, ranking 8th out of approximately 4,000 teams and winning a Gold medal. The competition aimed to improve the reasoning capabilities of the 30B Nemotron model, with all competitors accessing the same information and being evaluated on a private test set.

AIRA₃'s architecture consists of multiple agents running for extended periods in isolated environments without a central controller. The agents coordinate asynchronously through a forum for sharing hypotheses and a shared file system for solution artifacts, compounding discoveries to enhance performance. During the live competition, it was configured as an ensemble of GPT 5.5 and Claude 4.8, but post-evaluation assessments demonstrated competitiveness across various model combinations, including the Muse Spark 1.2 model achieving Gold medal-level performance.

The system also possesses domain generalization capabilities. It can be applied simply by changing task specifications, reducing production GPU kernel latency by 27% in internal benchmarks. Additionally, it achieved Gold-level performance on a task translating Akkadian clay tablets from 4,000 years ago. Meta assessed that such self-knowledge compounding systems demonstrate the potential for accelerating AI research and enabling recursive self-improvement.