LoopBench exposes world-model amnesia as scenes revisit; training-free WorldTrace fixes it
cindy_x_wu · x · 2026-11-21
cindyxwu presents work on spatial memory in video world models:
- LoopBench: the camera leaves a scene and later returns to the same pose; the return frame is compared against the start while varying route, rollout length, camera rotation, and number of revisits.
- Why it fails: old frames often remain in the KV cache, but attention can't retrieve them reliably — past the training horizon, RoPE offsets between queries and cached keys leave the trained range, and averaging keys for compression lets differently rotated keys cancel. With a sliding-window cache, a garage grows a third door and the garden and roof vanish.
- WorldTrace (training-free): keeps each memory slot addressable at a virtual position inside the trained range; frames whose merge barely changes attention outputs are averaged (WorldTrace-Field), while distinct frames like scene entries stay verbatim (WorldTrace-Landmark). The scene comes back intact at the same cache size.
More from Multimodal
- ElevenMusic Adds Chaos Parameter to Control How Far Generated Music Strays — charis_ai · 2026-11-21
- Full workflow fixes Fable 5.2 video audio entirely inside Claude — AIwithGhotai · 2026-11-21
- NVIDIA ships Alpamayo 2 Edge, a 2B-parameter in-vehicle reasoning VLA model — NVIDIAAI · 2026-11-21
- Creator shows HeyGen-powered video edit: 'editing to another level' — aziz4ai · 2026-11-21
- AI-built oil painting engine: 2,300 commits in 8 days as Claude Code runs a 'studio' of sessions — FotoRe_store · 2026-11-21
- "AGI Before GTA6": AI-Generated Music Video Riffs on 13-Year GTA VI Wait — Yamapama · 2026-11-21