Addressable Memory for Video World Models
WorldTrace keeps long video rollouts visually persistent by making compressed memory slots addressable inside the model’s trained position range.
The paper says video world models lose reliable recall when rollouts exceed their training horizon because RoPE time offsets move out of distribution. It also says naive KV-cache compression can corrupt memory by mixing incompatible positional phases. WorldTrace is training-free, with one variant for temporal coherence and another that stores scene traces at detected transitions for episodic recall. The authors report gains of +15.5% and +19.5% on LoopBench for those two settings. Source: HF Daily Papers' note.
The paper says video world models lose reliable recall when rollouts exceed their training horizon because RoPE time offsets move out of distribution. It also says naive KV-cache compression can corrupt memory by mixing incompatible positional phases. WorldTrace is training-free, with one variant for temporal coherence and another that stores scene traces at detected transitions for episodic recall. The authors report gains of +15.5% and +19.5% on LoopBench for those two settings. Source: HF Daily Papers' note.
score 5