Interactive video world models rely on Key-Value (KV) caches as a growing visual memory to carry forward previously generated frames. However, researchers at Hugging Face have identified a critical limitation: as rollouts extend beyond the training horizon, these models can no longer reliably address stored content. The cause is that temporal Rotary Positional Embeddings (RoPE) offsets fall outside the range seen during training, making it difficult for the model to retrieve relevant visual information through attention.
Additionally, naively compressing the cache in the RoPE-rotated space corrupts memory by averaging together incompatible positional phases. To overcome these issues, the team proposes WorldTrace, a training-free memory framework designed for long-horizon visual persistence. WorldTrace keeps compressed memory addressable by assigning each summary slot a distinct, in-distribution virtual position.
Within this addressable cache, the researchers explore two compression approaches: WorldTrace-Field, which compresses history for temporal coherence, and WorldTrace-Landmark, which stores verbatim scene traces at detected transitions for episodic recall. To evaluate these methods, they introduce LoopBench, a benchmark that tests whether a compressed cache can reconstruct a previously visited scene after a long detour.
Results on LoopBench show that WorldTrace-Field improves temporal consistency by +15.5%, while WorldTrace-Landmark improves episodic recall by +19.5%, extending visually persistent generation without retraining.