JANUS is a framework that unifies future outcome forecasting with past experience reflection, operating through a Predict‑Verify‑Evolve loop powered by a co‑evolving world model. Within an evolving graph the research agent proposes candidate hypotheses, while the world model scores them by forecasting value improvement, information gain, and exploration utility, then decides whether to keep, refine, or prune each branch before execution. Ground‑truth feedback from execution is distilled into evolving skills: hierarchical research skills, exploration‑strategy skills for the world model, and coordination protocols between them. This mechanism enables continual co‑evolution—the agent sharpens candidate generation while the world model improves long‑horizon predictive judgment. Extensive evaluations on MLS‑Bench‑Lite and real research tasks show that JANUS markedly outperforms state‑of‑the‑art baselines in both efficiency and performance ceiling.
Review