Model reports

ASI-Bench

First benchmark that jointly evaluates general intelligence, innovation, and autonomous execution, via 60 project-level scientific research tasks spanning 11 domains, progressively withdrawing human guidance. Very new (arXiv 2608.17271, code at github.com/apexin-ai/ASI-Bench) and intended to probe superintelligence-level behaviour, so high variance across seeds and scaffolds should be expected before any vendor adopts it. Only the seed-31415 instances are public; the seed-42 references stay private for official leaderboard scoring, so a locally computed score and a leaderboard score are not the same measurement. This registry's maintainer is one of the benchmark's authors.