Video world models that generate footage frame by frame start forgetting what they already made once a rollout runs long enough. NVIDIA Research says it has isolated the specific reason that happens, and built a workaround that needs no retraining of the underlying model. The finding gives teams building interactive video generation, from game-world simulators to robotics training environments, a way to fix long-horizon drift without rebuilding their models from scratch.
The researchers, working across NVIDIA, Princeton, the University of Toronto and the Vector Institute, studied autoregressive video world models: systems that produce each new frame by attending back over a cache of previously generated frames. To keep that cache from growing without bound, the model compresses older frames into a small set of summary slots. Each slot carries a positional signal, similar to a timestamp, telling the model how far in the past that memory sits.
The team found that once generation runs past the length the model was trained on, those timestamps drift outside the range it ever learned to read. The memory is still physically stored in the cache. The model just can’t address it anymore, because the positional signal no longer resembles anything seen during training. NVIDIA calls this an addressability problem, and it flagged a second, compounding issue: the common trick of averaging old memories together to make room for new ones scrambles their rotated positional signals, a failure the team calls phase cancellation that destroys the content the summary was meant to preserve.
WorldTrace, the framework NVIDIA built to fix both problems, assigns each summary slot a position based on its rank in the cache rather than its distance from the current frame. That keeps every slot’s positional signal inside the range the model was trained on, no matter how long the video runs. The fix only changes how positions are assigned and how memories get written into the cache, so it works on an existing world model without retraining. That distinction is the commercial hook: retraining a video world model on longer sequences is slow and expensive, while a training-free patch can ship as an inference-time change to a model teams have already deployed.
Two writers sit on top of that addressable cache. WorldTrace-Field averages memories in a shared, aligned coordinate space before re-rotating them into position, avoiding phase cancellation, and targets visual coherence across long rollouts. WorldTrace-Landmark instead freezes verbatim traces of frames where the model entered a new scene, aimed at letting the system recognize a location it has generated before.
NVIDIA’s own benchmarks, run internally rather than independently verified, show WorldTrace-Field improving TempSSIM, a measure of frame-to-frame coherence, by 5.9 percent over a baseline called Block-Rel at eight times the training horizon and by 15.5 percent over a sliding-window baseline at 24 times the horizon. On a recall benchmark the team built called LoopMem, WorldTrace-Landmark scored 0.825 against 0.627 for the sliding-window baseline on a long path that revisits an earlier scene, and 0.941 against 0.892 on a shorter repeated-revisit path. The smallest gain came on wide 360-degree camera pans, where WorldTrace-Landmark scored 0.577 against a baseline of 0.559, a result the paper reports plainly rather than smooths into an aggregate.
The paper won best paper at an ICML 2026 workshop on turning frames into stories, a workshop track rather than the main conference program, and none of the figures above have been replicated outside NVIDIA’s own lab. Teams building or evaluating long-horizon video world models for interactive simulation should check whether their current key-value cache implementation ties memory position to absolute frame distance, since that is exactly the failure mode WorldTrace targets, and testing an inference-time patch is far cheaper than committing to a retraining run.
NVIDIA Research published the WorldTrace findings on its project page and in an accompanying paper carrying arXiv eprint 2608.07408, placing the work in August 2026.
