RISE: Adaptive Imagination for World Action Models
RISE spends more rollout compute only when its evaluator expects more imagined future to improve planning.
The paper frames that as a fix for World Action Models that use the same imagination budget on every scene. Its Latent Evaluator estimates revealed risk and possible planning gain, while a Rollout Gate compares that gain with added compute. The authors also introduce CounterDrive, a counterfactual driving dataset with expert checks for trajectory validity, incident onset, and causal category. Experiments on NAVSIM and nuScenes report stronger planning performance with less unnecessary rollout. HF Daily Papers' note
The paper frames that as a fix for World Action Models that use the same imagination budget on every scene. Its Latent Evaluator estimates revealed risk and possible planning gain, while a Rollout Gate compares that gain with added compute. The authors also introduce CounterDrive, a counterfactual driving dataset with expert checks for trajectory validity, incident onset, and causal category. Experiments on NAVSIM and nuScenes report stronger planning performance with less unnecessary rollout. HF Daily Papers' note
score 5