Probing Visual Concepts in Lightweight Vision-Language Models for Automated Driving

Researchers analyzed Vision-Language Models (VLMs) used in automated driving to understand why they fail on simple visual tasks. They identified two failure modes: perceptual failure where visual information isn't encoded, and cognitive failure where information is present but not properly aligned with language semantics.

• →Vision-Language Models commonly fail on simple visual questions crucial for automated driving applications.
• →Object presence is explicitly encoded in VLMs while spatial concepts like orientation are only implicitly encoded.
• →Two distinct failure modes exist: perceptual failure (missing visual encoding) and cognitive failure (misaligned language semantics).
• →Linear separability of visual concepts degrades significantly as object distance increases.
• →Even when visual concepts are properly encoded in model activations, the model may still produce incorrect answers.