Repurposing the scientific literature with vision-language models