Scaling vision models does not consistently improve localisation-based explanation quality
Mateusz Cedro, Marcin Chlebus · 2026
Read Scaling vision models does not consistently improve localisation-based explanation quality by Mateusz Cedro, Marcin Chlebus (2026) 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/pmc-13575214.txt.
Description
Mateusz Cedro and Marcin Chlebus test whether bigger vision models give better explanations of their choices, in a 2026 study. Heat maps and like tools show which image parts drove a call. Users may trust that bigger models point more truly at the right parts. The team checked that hope across model families and data sets.
Eleven models from three families were trained or pre-trained and tested on medical and plain images with known masks. Five explanation tools were scored by how well their hot spots matched the masks. Deeper, larger models did not point better in most tests. Small models often matched or beat them. Some strong scorers pointed near nowhere near the right parts.
Catalog record
| CREATOR | Mateusz Cedro, Marcin Chlebus |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 62 KB |
| VERSION | 1.0 |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |