A benchmark for evaluating retrospective, present, and prospective reasoning in egocentric streaming video understanding.
EgoSAT is a benchmark for egocentric streaming interaction understanding. It evaluates whether vision-language models can reason about past, present, and future events under streaming observations, while also analyzing answerability and confidence calibration under temporal uncertainty.
For questions, please contact Yijia Lei at leiyj23@mails.tsinghua.edu.cn.
```