{"id":"adaptive-token-granularity","title":"Changing the grain of a conceptual token","regions":["representation","computation","distributed-learning"],"question":"Can a sequence or paragraph be represented as one working unit and later refined without losing the relations needed by another model?","formulation":"Token merging is proposed beyond ordinary segmentation: local units could carry coarse learning and be decomposed when more resources become available.","representations":[{"name":"Coarse unit treated as an atomic handle","affords":"Reduces the number of visible units.","loses":"A shorter sequence does not by itself reduce all computation or retain every interaction.","epistemic_kind":"reconstruction"},{"name":"Decomposable unit with internal structure","affords":"Supports later refinement and inspection.","loses":"Needs a mapping between coarse and fine operations; an embedding alone may not be invertible.","epistemic_kind":"reconstruction"}],"tension":"Compression, semantic adequacy, and exact reversibility are separate objectives.","challenge":"Which downstream tasks tolerate coarse units, and which require distinctions erased during merging?","development":"active","epistemic_kind":"reconstruction","evidence":["basis-adaptive-token-granularity"],"next_move":{"text":"Hold tasks and compute accounting fixed; compare reconstruction and downstream errors at several granularities.","epistemic_kind":"model_proposed"},"open_questions":["Can a sequence or paragraph be represented as one working unit and later refined without losing the relations needed by another model?","Which downstream tasks tolerate coarse units, and which require distinctions erased during merging?"],"formal_status":"open_problem","version":1,"content_hash":"1018a0bdb1694df973af5c7544c6f94d173aaf8088860b5d17e763593a61d2fd"}