Escaping the SpuriVerse: Can Large Vision-Language Models Generalize Beyond Seen Spurious Correlations?
Sep 1, 2025·,,,
,,,·
1 min read
Yifei Yang
Changping Lee
Sheng Shen Feng
Dongxu Zhao
Bingbing Wen
Andrew Z. Liu
Yulia Tsvetkov
Bill Howe
Abstract
We study whether large vision-language models can generalize beyond spurious correlations observed during training, introducing a benchmark to probe robustness to spurious signals.
Type
Publication
NeurIPS 2025 Datasets and Benchmarks
We introduce a benchmark to evaluate whether large vision-language models can move beyond spurious correlations and generalize robustly to distribution shifts that break shortcut cues.
