FrontierCS 2: A Benchmark for Computer Science R&D
The best agent passed 58% of its runs; 16 of the 48 research tasks saw no passing run.
FrontierCS 2 removes a paper’s contribution from its authors’ code, then asks agents to rebuild it inside the working repository. Agents can submit patches for development scores, while the evaluator, final tests and human reference remain private. The benchmark tests 12 agents across 13 domains, with Claude Opus 5.5 performing best. Beating the reference’s combined final score does not guarantee a pass: most runs that did so still fell short on individual metrics or workloads.
HF Daily Papers' note
FrontierCS 2 removes a paper’s contribution from its authors’ code, then asks agents to rebuild it inside the working repository. Agents can submit patches for development scores, while the evaluator, final tests and human reference remain private. The benchmark tests 12 agents across 13 domains, with Claude Opus 5.5 performing best. Beating the reference’s combined final score does not guarantee a pass: most runs that did so still fell short on individual metrics or workloads.
HF Daily Papers' note
score 5