Imagined Rollouts are Kinematic, Not Dynamic: A Diagnosis of Long-Horizon World-Model Failure

Modern generative world models render increasingly realistic action-controllable futures, yet they frequently hallucinate: rollouts remain visually fluent while drifting from the ground-truth dynamics. We hypothesize that hallucination concentrates in low-coverage regions of the state-action space, where lightweight data-centric signals can both detect it and guide mitigation. To test this, we introduce MMBench2, a 427-hour, 210-task dataset for visual world modeling with ground-truth actions, rewards, and live simulators, and train a 350M-parameter world model on it. We identify three distinct hallucination modes: perceptual, action-marginalized, and scene-diverging -- each anchored to a different stage of the pipeline, and develop three signals that accurately predict where the model will fail. To close coverage gaps at training time, we develop a coverage-aware sampling technique; to close them online, our hallucination predictors serve as curiosity rewards for targeted data collection, yielding a data-efficient finetuning recipe that adapts the pretrained world model to entirely unseen environments with as few as 50 real environment trajectories. Overall, our findings reveal that hallucination in world models is inherently a data coverage issue, and that the same signals used to detect it can also be used for mitigation. An interactive web version of our paper is available at https://www.nicklashansen.com/mmbench2

World models are usually evaluated as components of model-based reinforcement learning (MBRL) systems, while the world models themselves are rarely studied in isolation. We examine five representative visual world-model agents in Atari Pong: DreamerV3, DIAMOND, TWISTER, Simulus, and STORM. After reproducing their training pipelines and matching the reported agent performance, we freeze the learned world models and evaluate them with a closed-loop rollout diagnostic: a policy trained separately from the corresponding MBRL agent interacts with each frozen model, and the generated video trajectories are inspected for visual and dynamical errors. Across all five models, the rollouts contain clear failures, including ball disappearance, incorrect ball motion, and invalid ball-paddle interactions. Beyond visual trajectories, we further evaluate them with pixel-space zero-shot MBRL, where a new policy is trained entirely inside a frozen world model and then evaluated in the real environment. Across all five models, the resulting policies substantially underperform those produced by the corresponding original MBRL training pipelines. The gap is particularly large for DreamerV3, whose mean return drops from -5.5 to -20.9, near the minimum Pong return of -21. We hypothesize that insufficient modeling of task-critical concepts, such as the ball in Pong, may contribute to these failures. We therefore propose Concept-Guided Spatial Regularization (CGSReg), an auxiliary pixel reconstruction loss applied to segmented concept regions. Experiments show that CGSReg improves both closed-loop rollouts and pixel-space zero-shot MBRL in DreamerV3, DIAMOND, and TWISTER. Its effects vary across the remaining models and evaluation metrics, indicating that CGSReg alone does not address all world-model bottlenecks.

World Action Models (WAMs) enable decision-making through imagined rollouts by predicting future observations and actions. However, the reliability of these imagined futures remains under-examined: is a generated future merely visually plausible, or is it dynamically compatible with the action sequence it claims to model? In this work, we identify action-state consistency, the alignment between predicted actions and induced state transitions, as a missing reliability axis for WAMs. Through a systematic study across representative joint-prediction and inverse-dynamics models, we find that action-state consistency systematically separates successful and failed rollouts across many tasks and follows similar success-failure trends as learned value estimates. These results suggest that consistency captures decision-relevant structure beyond visual realism. We further identify background collapse as an important boundary condition, where low-dynamics failed trajectories can become deceptively consistent because static futures are easier to predict. Building on these findings, we introduce a value-free consensus strategy for test-time selection, which ranks candidate rollouts by agreement among predicted futures. This strategy improves success rates on RoboCasa and RoboTwin 2.0 without additional training or reward modeling. Taken together, our findings establish action-state consistency as both a diagnostic tool for evaluating WAM reliability and a practical signal for value-free planning.