BenchmarkARI Bench
ARI Bench
A reference benchmark for recursive self-improvement of AI, examining the accuracy of models when iteratively improving their own outputs or code.
Timeline (last 90 days)
Nov 18
On the ARI Bench benchmark, Qwen 4 Max Preview reached 23% accuracy in Extra High mode, far behind GPT-6 Astra and Opus 5.5.
Unverified50%