H-Net

Summary

H-Net is an end-to-end hierarchical sequence model that learns dynamic byte chunking jointly with language modeling.

Role In The Wiki

H-Net is the clearest example of replacing tokenizer preprocessing with learned hierarchical chunking.

ReinPatch makes H-Net relevant to time-series modeling as a transfer case study: it keeps the learned-boundary question but reports that H-Net’s original embedding-similarity chunking policy is not automatically the right inductive bias for continuous time-series forecasting.

Evidence

Relation To Foundation TSFM Agenda

Use the source-level agenda mappings rather than duplicating verdict rows here:

At the entity level, H-Net is the clearest example of replacing tokenizer preprocessing with learned hierarchical chunking. This page should stay as the object card; source pages carry slot-level verdicts, evidence, and missing pieces.