Can Vision Language Models Learn from Visual Demonstrations of Ambiguous Spatial Reasoning?

TL;DR

This paper introduces SVAT, a benchmark to test if vision-language models can learn new visuospatial concepts from visual demonstrations, revealing current limitations and potential improvements with curriculum learning.

Contribution

The paper proposes a new benchmark, SVAT, to evaluate VLMs' ability to learn from visual demonstrations and investigates methods to enhance their in-context learning capabilities.

Findings

VLMs fail to learn new visuospatial tasks zero-shot.

Finetuning alone does not significantly improve learning.

Curriculum learning with simpler data enhances ICL performance.

Abstract

Large vision-language models (VLMs) have become state-of-the-art for many computer vision tasks, with in-context learning (ICL) as a popular adaptation strategy for new ones. But can VLMs learn novel concepts purely from visual demonstrations, or are they limited to adapting to the output format of ICL examples? We propose a new benchmark we call Spatial Visual Ambiguity Tasks (SVAT) that challenges state-of-the-art VLMs to learn new visuospatial tasks in-context. We find that VLMs fail to do this zero-shot, and sometimes continue to fail after finetuning. However, adding simpler data to the training by curriculum learning leads to improved ICL performance.