Meta's autonomous research system AIRA3 places 8th of about 4,000 teams in NVIDIA's Kaggle reasoning challenge

Meta says its autonomous AI research system AIRA3 entered NVIDIA's Nemotron reasoning competition on Kaggle in June and finished 8th out of roughly 4,000 teams, a gold-medal result. The system had the same tools and information as human competitors and was graded externally on the same private test

In a four-post thread on its official X account on 5 September, Meta said its autonomous AI research system AIRA3 entered a live Kaggle competition run by NVIDIA in June to teach a 30-billion-parameter Nemotron model to reason better. Meta says AIRA3 placed 8th of roughly 4,000 teams, gold-medal level, with the same tools and information as human competitors and external grading on a private test set.

Meta calls this “a reliable signal that AIRA3 can improve a targeted capability of an AI model at a level similar to human experts.” The thread does not name the team on the leaderboard; the results are Meta's statement, unverified independently.

Competition and models

Per the Kaggle page, the NVIDIA Nemotron Model Reasoning Challenge ran from mid-March to mid-June. Entrants all started from Nemotron-3-Nano-30B and submitted a LoRA adapter with rank at most 32, scored on accuracy on a new NVIDIA Research reasoning benchmark. 4,182 teams entered; first prize: $25,000 and 5 DGX Spark.

By Meta's account, the 8th-place live entry combined GPT 5.5 (with OpenCode) and Claude 4.8 (with ClaudeCode). In post-hoc runs on the same private test set, Meta's own Muse Spark 1.2 (with MuseCode) also reached gold level; Muse Spark 1.1 and GLM 5.2 (both with OpenCode), silver.

How it works

Meta says AIRA3 has no central controller: many long-running agents, each a model paired with a coding harness, work in isolated environments and coordinate asynchronously through a shared forum for hypotheses and findings and a shared filesystem for solution artifacts, building on each other's discoveries; search strategies emerge dynamically.

Generalization claims

Meta says only the task specification changes between domains, citing a 27% latency reduction on production GPU kernels in an internal benchmark and gold level in another Kaggle competition translating 4,000-year-old Akkadian clay tablets into English. It calls itself “early” and names “recursive self-improvement” as the goal; its predecessor, AIRA2, appeared on arXiv in March.

For the corporate reader

One data point that part of model fine-tuning can be handed to autonomous systems; also a single competition, the company's own statement, not a production environment.