VLMs are Biased
alive
HTTP 200
Last checked: 2026-03-02T21:38
Description
# Description
This webpage presents research demonstrating that Vision Language Models (VLMs) suffer from severe confirmation bias and fail at simple counting tasks when familiar objects are subtly modified. The site documents the "VLMBias" benchmark framework, which tests whether VLMs actually analyze images visually or rely on memorized knowledge—finding that models achieve 100% accuracy counting normal objects but only ~17% accuracy when those same objects are slightly altered (e.g., a 5-legged dog or 4-striped Adidas logo).