NVIDIA's WorldTrace: Overcoming Memory Limits in Video World Models
nvidia · hf · 2026-08-10
NVIDIA's new research addresses visual persistence in interactive video world models. The study reveals that when rollouts extend beyond the training horizon, models struggle to retrieve visual information from the KV cache due to temporal Rotary Positional Embedding (RoPE) offsets, and naive compression corrupts memory.
To solve this, they propose WorldTrace, a training-free memory framework. It keeps compressed memory addressable by assigning distinct virtual positions to summary slots, featuring two approaches:
- WorldTrace-Field: Compresses history for temporal coherence.
- WorldTrace-Landmark: Stores verbatim scene traces at transitions for episodic recall.
They also introduced LoopBench, a benchmark evaluating a compressed cache's ability to reconstruct visited scenes after long detours. WorldTrace improved temporal consistency by 15.5% and episodic recall by 19.5%.
More from Research
- Jeff Dean's First Talk Post-Google: How to Pick Research Problems Worth 5 Years — FinanceYF5 · 2026-08-10
- Apple-π: First Benchmark for Law-Grounded Physical Video Reasoning — jiqizhixin · 2026-08-10
- Open Source Code for "Hands-On Large Language Models" Released with 300+ Diagrams — techNmak · 2026-08-10
- Mobile-GS: Real-time Gaussian Splatting for Mobile Devices (ICLR 2026) — tom_doerr · 2026-08-10
- ByteDance Introduces Modular TTT: Modularizing Test-Time Training with Systematic Ablation — ByteDance-Seed · 2026-08-10
- ByteDance Releases Douyin Multimodal Embedding Model, Deployed in Search — ByteDance · 2026-08-10