NVIDIA's WorldTrace boosts video world model scene recall by 32% without retraining

NVIDIAAI · x · 2026-11-21

NVIDIA Research's WorldTrace tackles scene inconsistency in interactive video world models: beyond the training horizon, temporal RoPE offsets exceed the trained range so KV-cache content becomes unaddressable, and naive compression in rotated space corrupts memory. WorldTrace is training-free — it assigns each summary slot a distinct in-distribution virtual position and lets the model's own attention decide what to keep verbatim vs merge. On the LoopBench scene-recall benchmark it improves average DINOv2 recall by 32.6% over sliding-window at the same attended-cache size. NeurIPS 2026 paper.

Original post →

More from Research

Research channel →