Multimodal Applications 关注模型如何在真实任务中使用图像、音频、视频和文本信息。知识组织从单模态感知开始,逐步进入 grounding、reasoning、工具调用和多轮环境交互。
Task Path
perception
-> recognition and extraction
-> grounding and spatial / temporal reasoning
-> multimodal answer or transformation
-> tool use and environment interaction任务范围
- visual question answering 和 image understanding;
- OCR、document understanding、chart / table reasoning;
- visual grounding、spatial reasoning 和 temporal reasoning;
- audio understanding、speech interaction 和 video understanding;
- multimodal RAG、tool use 和 GUI interaction。
评测关注
不能只看生成文本是否流畅,还需要区分 perception accuracy、grounding accuracy、reasoning correctness、hallucination、latency 和真实任务成功率。