RISE: Adaptive Imagination for World Action Models

RISE：面向世界动作模型的自适应想象

Abstract: World Action Models (WAMs) improve planning by incorporating future world evolution into action generation, yet existing methods allocate a fixed imagination budget to every scene.

摘要： 世界动作模型（World Action Models, WAMs）通过将未来的世界演变纳入动作生成过程来提升规划能力，然而现有的方法对每个场景都分配了固定的想象预算。

We propose RISE (\textbf{R}efining \textbf{I}magination through \textbf{SE}lective Rollout), a system-level adaptive imagination framework that makes sequential \textsc{Roll}/\textsc{Stop} decisions according to the expected planning benefit of continued rollout.

我们提出了 RISE（通过选择性展开来优化想象），这是一个系统级的自适应想象框架，它根据持续展开所带来的预期规划收益，按顺序做出“展开（Roll）”或“停止（Stop）”的决策。

At each step, a Latent Evaluator estimates the risk revealed by the current prefix and how much planning could improve if imagination continues, while a Rollout Gate weighs this expected benefit against additional computation cost.

在每一步中，潜在评估器（Latent Evaluator）会估计当前前缀所揭示的风险，以及如果继续想象，规划能有多大程度的提升；同时，展开门控（Rollout Gate）会权衡这种预期收益与额外的计算成本。

Since factual driving logs expose only one realized future, we further construct \textbf{CounterDrive}, a counterfactual dataset with diverse outcomes and risk levels, to enrich future dynamics and provide localized risk supervision.

由于真实的驾驶日志仅展示了一种已实现的未来，我们进一步构建了 CounterDrive，这是一个包含多样化结果和风险水平的反事实数据集，旨在丰富未来动态并提供局部风险监督。

Each retained sample undergoes expert verification and annotation of trajectory validity, incident onset, and causal category, providing a reusable resource for safety-critical world-modeling research.

每个保留的样本都经过专家验证，并标注了轨迹有效性、事故发生点和因果类别，为安全关键型世界建模研究提供了一种可复用的资源。

Experiments on NAVSIM and nuScenes show that RISE achieves the best overall planning performance while reducing unnecessary rollout, with additional transfer results supporting its plug-in generality across WAM architectures.

在 NAVSIM 和 nuScenes 上的实验表明，RISE 在减少不必要展开的同时实现了最佳的整体规划性能，额外的迁移结果也证明了其在不同 WAM 架构间的插件式通用性。