Addressable memory lets video world models return to earlier scenes
NVIDIA and Princeton researchers show that interactive world models lose track of earlier scenes once a rollout runs longer than their training length, and propose a fix that needs no retraining. They report 15.5% better temporal consistency and 19.5% better recall on a new benchmark that tests returning to a place after a long detour.
Walking away and coming back is what a camera does all day. A world model that forgets the room when you turn around cannot hold continuity, so this is the specific problem to watch — and it is good to see it measured directly.
Continuity is the first thing a script supervisor protects and the first thing generative video loses. This paper is useful because it isolates one precise mechanism behind that loss in interactive world models and measures it, instead of filing it under the general heading of drift.
According to the abstract, the models in question use a key-value cache as a growing visual memory of what has been seen. Once a rollout runs past the length the model was trained on, the temporal position encodings (RoPE offsets) attached to older memories fall outside the range the model knows, and it struggles to retrieve them. Compressing that cache naively makes things worse, because it blends entries with incompatible positional phases. The authors’ answer, WorldTrace, is training-free: each compressed summary slot gets its own virtual position that stays inside the familiar range. They test two variants. WorldTrace-Field compresses history for temporal coherence and reports a 15.5% improvement in temporal consistency; WorldTrace-Landmark keeps verbatim traces of the scene at detected transitions and reports a 19.5% improvement in episodic recall. To measure the latter, they built LoopBench, which asks whether a model can reconstruct a place it visited after a long detour.
LoopBench is the part worth borrowing as a way of thinking. Most video evaluations score a clip in isolation. Production cares about the return: the reverse shot back to the kitchen after the scene has played out in the hallway, the second pass along the same storefront in a chase. A world model steered by a camera, like the systems announced since in Runway’s GWM Worlds 2 and Google’s Genie 4, is only useful for virtual scouting or previs if the set is still the set when the camera comes back. A benchmark that tests exactly that gives buyers a question to put to vendors that is more precise than how long a session can run.
Training-free matters in practice. A fix that lives in the memory layer can, in principle, be applied to existing models rather than waiting for the next generation of them, which is the kind of improvement that tends to reach tools relatively quickly. The landmark approach is also intuitive to anyone who has worked on a set: keep an exact record of each location at the moment you leave it, and compress what happens in between. It is close to what a continuity photo does.
The limits are those of any single paper. The gains are relative and reported on the authors’ own new benchmark, and the abstract does not say how close to a faithful return either variant gets, how it behaves over very long sessions, or what it costs in compute. Remembering what a scene looked like is also not the same as remembering what happened in it: a chair that was moved or a door that was opened is state, not just appearance. Our view is that memory and state will decide whether world models become a previs instrument or remain a demo, and this is one of the first papers to frame the problem in terms a production would recognise.