Researchers present RISE, an adaptive framework for world action models that dynamically decides when to stop imagination rollouts based on expected planning benefit versus computational cost, using a Latent Evaluator to assess risk and potential improvement and a Rollout Gate to balance those factors against overhead. To address the limitation that real driving logs contain only one realized future, the team introduces CounterDrive, a counterfactual dataset with diverse outcomes and risk annotations. Experiments on NAVSIM and nuScenes show RISE achieves superior planning performance while reducing unnecessary rollout computation, with transfer results showing compatibility across different world-action-model architectures.
