FrontierCS 2 launches 48 paper-based computer science tasks for AI agents

FrontierCS 2 gives AI agents 48 research-code tasks across 13 computer science domains, according to its creators. Agents can test patches during development, but their final code passes only if it beats the paper’s original implementation on every metric and hidden workload. The researchers report that no run passed on 16 tasks; a private set of more than 100 tasks is planned for a future announcement.