SugarCREPE

'SugarCrepe++' is a dataset used to evaluate the sensitivity of generative vision-language models (VLMs) to lexical and semantic alterations in prompts.

Papers using SugarCREPE (8)

PolyGen: Fully Synthetic Vision-Language Training via Multi-Generator Ensembles2026Contrastive vision-language learning with paraphrasing and negation2025COCO-Tree: Compositional Hierarchical Concept Trees for Enhanced Reasoning in Vision Language Models2025A Good CREPE Needs More Than Just Sugar: Investigating Biases In Compositional Vision-language Benchmarks2025Advancing Compositional Awareness In CLIP With Efficient Fine-tuning2025Contrastive Region Guidance: Improving Grounding in Vision-Language Models without Training2024 · 13 citesSensitivity of Generative VLMs to Semantically and Lexically Altered Prompts2024A New Method to Capturing Compositional Knowledge in Linguistic Space2024