WorldTrace beats sliding window recall by 32.6% at same cache size, tops all 12 LoopBench conditions

cindy_x_wu · x · 2026-11-21

WorldTrace, a long-horizon memory approach for world models built on Matrix-Game 2.0 and LingBot-World, improves average recall by 32.6% over a sliding window at equal cache size on LoopBench and ranks first in all 12 conditions without retraining. Related work: MemRoPE, ∞-RoPE, YaRN, MosaicMem.

Original post →

More from Research

Research channel →