Search papers, labs, and topics across Lattice.
This paper re-evaluates long-horizon failures in world models by distinguishing between kinematic and dynamic errors, proposing a novel diagnostic called imagined Kinematic-Consistency Error (iKCE). The authors demonstrate that world models, specifically a DreamerV3 checkpoint, exhibit significantly higher iKCE compared to real-physics rollouts, indicating a fundamental misalignment in how these models imagine future states. Notably, even when the model's trained policy experiences a reward collapse, its iKCE remains statistically unchanged, underscoring the limitations of current world models in capturing dynamic behaviors over extended horizons.
World models may be fundamentally misinterpreting future states, imagining them kinematically rather than dynamically, leading to significant long-horizon failures.
Long-horizon failure in world models is conventionally attributed to compounding error, a generic framing that does not distinguish what kind of error compounds. We propose a kinematic-vs-dynamic reframing: world models tend to imagine kinematically rather than dynamically. We operationalize this as the imagined Kinematic-Consistency Error, a per-step diagnostic that measures how far a rollout departs from a closed-form kinematic null, paired with a perturbation protocol that tests whether iKCE responds when physical conditions cross a regime boundary. We instantiate the diagnostic on a released DreamerV3 checkpoint trained on DMC walker-walk, where imagined iKCE runs roughly two orders of magnitude above that of matched real-physics rollouts. Across a friction sweep that crosses the gait-collapse boundary, the model's iKCE stays statistically flat even as the trained policy's reward collapses through the same range, providing the kinematic-not-dynamic signature. The diagnostic distinguishes kinematic from dynamic imagination at horizons longer than the embodiment's gait period.
