About object detection

I think that you push below token in llm
```
['<cls>', '<x1>', '<y1>', '<x2>', '<y2>', '<cls>', '<x1>', '<y1>', '<x2>', '<y2>', '<cls>', '<x1>', '<y1>', '<x2>', '<y2>', ...]
```

about object detection loss,
did you use hungarian matching like detr?


Or if you use just next token prediction by cross entropy loss, how to sort the ground-truth box?