NVIDIA's WorldTrace Fixes Scene Amnesia in Video World Models Without Retraining
NVIDIA Research has released WorldTrace, a training-free method that fixes "scene amnesia" in interactive video world models, with the paper submitted to NeurIPS 2026. The researchers found that worlds generated by such models degrade once the camera leaves and returns—a garage sprouts an extra door, a garden and roof vanish. The root cause: these models rely on the KV cache as an ever-growing visual memory, but once rollouts exceed training duration, temporal RoPE shifts move beyond the trained range and the cached content can no longer be addressed by attention.
Confirmed
- Core mechanism: WorldTrace applies without training, fixing the KV cache disconnection caused by out-of-range temporal RoPE shifts
- It ships with the LoopBench benchmark: the camera leaves a scene and returns to the original position, comparing return frames against starting frames while systematically varying route, rollout length, camera rotation, and revisit counts to test scene memory
- Results: revisit recall improved 32.6% (about 32%) at the same cache size
Why it matters
- Visual persistence is a key bottleneck for making interactive world models usable (games, simulation, agent training environments); WorldTrace offers a low-cost fix without retraining
- LoopBench is the first benchmark to systematically quantify world models' "spatial memory" deficits, providing a reproducible evaluation standard for future work
2026-11-21 ~ 2026-11-21 · 6 related posts
Primary sources
- [source] NVIDIA's WorldTrace boosts video world model scene recall by 32% without retraining — NVIDIAAI · 2026-11-21
- [source] NVIDIA's WorldTrace fixes scene forgetting in video world models, NeurIPS 2026 paper — cindy_x_wu · 2026-11-21
- LoopBench exposes world-model amnesia as scenes revisit; training-free WorldTrace fixes it — cindy_x_wu · 2026-11-21
- [source] LoopBench tests world-model memory; WorldTrace lifts recall 32.6% over sliding window — cindy_x_wu · 2026-11-21
- WorldTrace beats sliding window recall by 32.6% at same cache size, tops all 12 LoopBench conditions — cindy_x_wu · 2026-11-21
1 near-duplicate retellings: cindy_x_wu