Ludwig Schmidt

This research focuses on optimizing language model training and predicting their real-world performance.

This study focuses on enhancing spatial accuracy in text-to-image generation.

A study highlights CLIP's reliance on spurious features in image recognition.

Including non-English data improves vision-language model performance and cultural understanding.

VLMs struggle with image classification, but better data integration can enhance their capabilities.

Leveraging language models improves predictions for tabular data across various fields.

MINT-1T is the largest open-source dataset for training multimodal models.

A guide to improving language model training with limited resources.

A new method enhances synthetic data quality for better language model alignment.

xGen-MM enhances multimodal models for better image and text learning.

KALE combines images with rich captions for better understanding.