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.
More from Research
- Anthropic researcher claims Riemann Hypothesis proof with Claude, backed by 17.4M-line Lean formalization — Kyrannio · 2026-11-21
- Anthropic's Levent Alpöge claims a Riemann Hypothesis proof discovered working with Claude — Kyrannio · 2026-11-21
- Stanford AI100 releases 2026 report on the scaling era and mainstream generative AI — stanfordnlp · 2026-11-21
- New Multi-Agent Coordination Bench finds frontier agents coordinate poorly; only Opus gains — AlexGDimakis · 2026-11-21
- Riemann Hypothesis Result's Dramatically Written Intro Draws Attention — lishali88 · 2026-11-21
- Stanford CS329Z Lecture 9 Released: Optimizing AI Agents via Prompts vs. Weights — stanfordnlp · 2026-11-21