Meta's AIRA₃ Agent Wins Gold in Kaggle Competition with Model Ensemble

AI at Meta ·

Key Info

Meta's autonomous AI research system AIRA₃ competed live in a Kaggle competition, with an ensemble of models (GPT 5.5 with OpenCode + Claude 4.8 with ClaudeCode) placing 8th and earning gold-medal level performance. In additional post-hoc evaluations, Muse Spark 1.2 (with MuseCode) also reached gold-medal level, while Muse Spark 1.1 (with OpenCode) and GLM 5.2 (with OpenCode) achieved silver-medal level performance.

Highlights

  • The live competition entry, a combination of two leading models, ranked 8th overall with gold-medal performance.
  • Post-hoc assessments showed similar or slightly lower but still strong results for other Meta models and third-party models.
  • The results highlight the value of ensembling multiple AI models for tackling complex, real-world research challenges.
Loading...