World models simulate an environment and predict its evolution under actions, and they are increasingly used in real‑world applications such as robotics. Complex systems need the same tool because a single action often has a negligible immediate effect; the crucial factor is the outcome that unfolds over subsequent steps. For example, seeding nodes for a marketing campaign or immunizing nodes against an epidemic changes little on its own, while the eventual diffusion determines success. Designing an algorithm therefore requires iteratively scoring candidate actions by their expected payoff, which traditionally relies on extensive simulation and becomes a computational bottleneck.
We propose an action‑conditioned Network World Model that learns a network’s diffusion dynamics under interventions over time. Given the current network state and a candidate action, the model predicts the resulting diffusion outcome. This model serves as a fast evaluator inside an algorithm‑design loop where a coding agent creates and refines executable algorithms using feedback from full rollouts, action‑level credit, and counterfactual probes of alternative interventions.
Across eight network tasks and five diffusion models, the algorithms designed with this model match or exceed the strongest reported baselines in 138 of 141 settings, while achieving up to 14.5× faster rollouts compared to Monte Carlo simulation. The code will be released upon acceptance.
Review