A new benchmark, Into the Unknown, asks whether AI agents can carry out research on open scientific problems. The benchmark gives an agent a science paper along with a research direction to pursue, then tasks it with writing a follow-up. This setup is designed to test more than retrieval or summarisation—it requires the agent to extend prior work into unknown territory.

The benchmark's focus on open problems distinguishes it from typical AI evaluations, which often rely on closed tasks with known answers. By starting from a real paper and a direction, the benchmark aims to approximate the early stages of scientific inquiry. The abstract does not report results, so the feasibility of such agents remains an open question.

As a single-source announcement, this is an early description of the benchmark rather than a comparative study. Its significance lies in defining a concrete evaluation for AI-driven research, a step toward understanding whether agents can contribute to genuine scientific discovery.