Patronus AI builds training environments for scientific research agents

Patronus AI is developing reinforcement learning environments where science agents read prior work, propose tests, run code or simulations, and use results to decide what to investigate next. The approach emphasizes realistic, underspecified research problems rather than questions with a predetermined answer. Areas under exploration include engineering, scientific computing and biology, alongside skills such as experimental design.

Knowledge benchmarks highlight the gap between answering difficult questions and conducting research. Humanity’s Last Exam, which Patronus researchers contributed to, contains 2,500 expert-written questions at the edge of human knowledge. The best model scored 8% at launch. Choosing what to test and interpreting failed experiments remain difficult tasks for research agents.

Patronus also studies world models that let language models stand in for environments during agent training. GlobeBench measures whether models can predict an action’s effects: the best success rate was 49.6%, falling to 5.1% when a change needed to propagate to dependent objects. Its Generative Simulators work generates tasks, world dynamics and rewards together, allowing difficulty to increase as agents improve. These approaches aim to expand training situations while measuring limitations in predicting environmental responses.