WorldKV: Efficient World Memory with World Retrieval and Compression

Researchers introduced WorldKV, a training-free memory system for video world models that preserves long-term scene consistency.

It stores evicted KV-cache chunks for selective retrieval and compresses redundant tokens to save memory without losing fidelity.

The approach roughly doubles inference throughput compared with full KV caching while keeping persistent worlds more coherent.

WorldKV targets the core tradeoff between fast autoregressive video generation and long-term consistency, without requiring fine-tuning.