Benchmark
FrontierCS 2
Test whether AI agents can improve on published computer-science research code.
Published leaderboard
Explore scores and evaluation settings on the publisher’s leaderboard.
Open leaderboardCoverage
48 released research tasks across 13 computer-science domains, including machine-learning systems, databases, security, and robotics. A separate private set of more than 100 tasks is not included.
Methodology
Agents receive research repositories with the central contribution removed and have five hours to develop a replacement, using feedback on development workloads. Their chosen final patch must beat the original implementation on every required metric and held-out workload.