Winoground

Winoground is a benchmark that evaluates the compositional generalization capabilities of vision-language models by assessing their ability to align directional semantic relations in text with spatial relationships in images.

Papers using Winoground (6)

Inference-Time Structural Reasoning for Compositional Vision-Language Understanding2026Looking to Learn: Token-wise Dynamic Gating for Low-Resource Vision-Language Modelling2025COCO-Tree: Compositional Hierarchical Concept Trees for Enhanced Reasoning in Vision Language Models2025Winoground: Probing Vision and Language Models for Visio-Linguistic Compositionality2022 · 14 citesCross-modal Attention Congruence Regularization for Vision-Language Relation Alignment2022Prompting Large Vision-Language Models for Compositional Reasoning2024