JANUS anticipates results before launching research agent experiments

A world model helps the agent choose which leads to execute. The actual results are then used to improve prediction and scientific exploration.

JANUS adds a prediction step to the classic loop that offers an idea, executes and returns. In an evolutionary graph, the world model estimates the value of a track and the information it could bring before the agent engages experience.

Actual returns feed into research skills and exploration strategies. The authors evaluate the device on MLS-Bench-Lite and research tasks, with announced improvements in efficiency and final performance. The predictive model guides the choice of tests; it does not replace their experimental validation.

Pre-publication filed on arXiv on 9 October 2026; The results are those reported by the authors and remain to be confirmed independently.