LoopNav: Benchmarking Spatial Consistency in World Models

Kewei Lian, Shaofei Cai, Yitao Liang, Anji Liu · 2025

Read LoopNav: Benchmarking Spatial Consistency in World Models by Kewei Lian, Shaofei Cai, Yitao Liang, Anji Liu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-22976.txt.

Description

Kewei Lian and three co-authors present a study from twenty twenty-five on a test for world models that must keep space straight. A world model fakes how a scene shifts as an agent moves. If it forgets where things are, plans built on it fail. Past test sets stress looks, not long range space sense. Their set, LoopNav, is built on loops.

Agents walk loops in the game Minecraft and come back to the start. The set holds a vast span of video with the moves logged. A new score checks if the model's inner scene graph stays in line on the loop. It shrugs off pixel level noise. Data, tests, and code are open for later work.

Catalog record

CREATORKewei Lian, Shaofei Cai, Yitao Liang, Anji Liu
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE70 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-22976

Browse the full library · About the Tower of Babel