Changing the grain of a conceptual token

Can a sequence or paragraph be represented as one working unit and later refined without losing the relations needed by another model?

Formülasyon

Token merging is proposed beyond ordinary segmentation: local units could carry coarse learning and be decomposed when more resources become available.

Gerilim

Compression, semantic adequacy, and exact reversibility are separate objectives.

Açık soru

Which downstream tasks tolerate coarse units, and which require distinctions erased during merging?