Addressable Memory for Video World Models
WorldTrace keeps long video rollouts visually addressable without retraining the model.
The paper says video world models lose reliable access to old frames when rollouts go past their training horizon, partly because RoPE time offsets move out of distribution. Its proposed framework assigns compressed memory slots virtual positions that stay in range, avoiding corruption from naive cache compression. The authors test two variants: one for temporal coherence and one for episodic recall after scene transitions. On LoopBench, they report +15.5% temporal consistency and +19.5% episodic recall. ArXiv · AI/CL/LG's note
The paper says video world models lose reliable access to old frames when rollouts go past their training horizon, partly because RoPE time offsets move out of distribution. Its proposed framework assigns compressed memory slots virtual positions that stay in range, avoiding corruption from naive cache compression. The authors test two variants: one for temporal coherence and one for episodic recall after scene transitions. On LoopBench, they report +15.5% temporal consistency and +19.5% episodic recall. ArXiv · AI/CL/LG's note
score 5