Researchers from UC Berkeley, including PhD student Qiuyang Mang and collaborators from Princeton, have introduced FrontierCS 2. This benchmark evaluates the ability of AI agents to improve upon existing computer science research without access to the original papers or the internet.
The test set comprises 48 open-ended tasks spanning 13 domains, such as robotics, databases, and machine learning. Each agent is provided with a code repository and must rebuild and enhance the implementation within a five-hour timeframe. Performance is measured against both known metrics and hidden workloads.
During initial testing, 12 agent-model pairs were evaluated. Claude Opus 5.5 achieved the highest success rate at 58%. Subsequent updates in November added Claude Sonnet 5.5 to the leaderboard, which recorded a 50% pass rate. Notably, 16 of the 48 tasks remained unsolved by all tested models.
FrontierCS 2 succeeds the original FrontierCS benchmark presented at the ICML 2026 conference. The new iteration focuses on real-world environments by utilizing functional codebases from academic studies. To prevent data staleness, the team plans to release over 100 additional private tasks in the future.






