Referring understanding is a core task in vision-language research, where systems localize objects described in free-form language. In multi-object tracking, this extends to following those objects across frames. The abstract of the paper notes that existing works are constrained by limited language, which restricts how naturally people can refer to targets.

The paper, "Referring Multi-Object Tracking with Diverse Dynamic Expressions," addresses this constraint. Its focus is on scenarios where expressions are not only varied but also dynamic—changing as the context or scene evolves. This moves beyond static, simple descriptions toward more realistic and flexible interaction.

The significance is in pushing models to handle the complexity of real-world referring expressions. By pointing out the limitation, the work underscores the need for approaches that can cope with the full range of natural language used in tracking tasks.