The authors constructed multiple-choice video question-answering tasks from walking tours covering distances of up to 1 km.
Current vision-language models fail to build city-scale mental maps
Across real-world walking videos spanning up to one kilometre, models relied mainly on visual recognition and text matching rather than integrating routes into geometric maps.
Big Tech
Aravindh Mahendran · Michael King · Matthew Koichi Grimes · Antoine Yang · Tyler Zhu · Joseph Heyward · +9 more
Google DeepMind · Princeton University · Toyota Technological Institute at Chicago · Google
Research Digest··2 min read
Mahendran et al.
Why this paper
From Google DeepMind and 3 others
In one line
VLMs cannot infer city-scale spatial layouts from video; they bypass reasoning via 2D recognition and text-matching.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§