Hugging Face 推出 ConTextual 排行榜：評估多模態模型在富含文本場景中的圖文聯合推理能力

Hugging Face has announced the launch of a new multimodal benchmark and leaderboard called "ConTextual," aimed at addressing the shortcomings of existing multimodal evaluation tools. Traditional multimodal benchmarks tend to focus on object detection or general image description, but in real-world applications — such as reading financial charts, understanding internet memes, or recognizing signs in street-view imagery — models must simultaneously understand "text within images" and "the visual context surrounding that text." This is known as "joint text-image reasoning."