AIRA2 (arXiv) is Meta’s follow-on research agent, built around fixes for three bottlenecks the field had identified: synchronous single-GPU execution, a generalization gap in which validation-based selection overfits over long search horizons, and the ceiling imposed by fixed single-turn operators. It reports an 81.5% mean percentile rank on MLE-bench-30 at 24 hours (83.1% at 72) against a 72.7% baseline, exceeds human state of the art on 6 of 20 AIRS-Bench tasks, and its ablations attribute earlier reports of overfitting to evaluation noise. In this corpus it shares the March 31 development with Bilevel Autoresearch and follows Meta’s hyperagents of the previous week.
AISystem
AIRA2
Meta's second-generation AI research agent, which attacks three structural bottlenecks in research agents with an asynchronous multi-GPU worker pool, a hidden consistent evaluation protocol and interactively debugging ReAct agents.
Sources
Attribution and reporting linked by this record; links do not imply independent verification.
- Welcome to March 31, 2026 (theinnermostloop.substack.com)
Development timeline
Newest reporting dates first. Reporting dates may differ from event dates.
Knowledge graphopen in browser →
Developed by
Cite or correct this record
Suggest a correction — opens a GitHub draft for you to review and submit.