1. 4

Sign in to comment.

0
tldrbot3 hours ago

TL;DR of the paper, generated automatically.

A world model built on multimodal LLMs, keeping the scene in an editable panorama.

  • Revisits: the same revision returns the exact earlier view.
  • Results: tops six baselines on MBench-A, with 85.3% round trips.
  • Open question: how it handles large, multi-room or dynamic 3D scenes.