← All works
Full-stack recurrence

Block Recursive Transformers

Block Recursive Transformers for Structured Parameter Sharing

Josep Marques, Hyunyoung A. Cho, Youngeun Kim, Youjia Zhang, Aecheon Jung, Kyuhong Shim, Sungeun Hong

Argues that how shared layers are organized matters. BRT splits the stack into contiguous blocks and recurses each block before moving to the next, an intermediate topology between per-layer repetition and looping the entire stack.

Inside the method

[ Block 1 ] × r → [ Block 2 ] × r → … → [ Block n ] × r

Simplified conceptual schematic. Consult the paper for the complete architecture.

Recurrence family
Full-stack recurrence
Depth control
Static per-block recursion
KV / state strategy
See paper

Reading note

Improves validation NLL and downstream accuracy over static recursive sharing in decoder-only models up to 1.7B, with additional encoder-decoder and encoder-only studies; no routing modules or extra objectives.

Sources checked 2026-10-28. This catalog does not imply independent reproduction.

Related works

Cite this work

@misc{brt2026,
  title = {Block Recursive Transformers for Structured Parameter Sharing},
  author = {Josep Marques and Hyunyoung A. Cho and Youngeun Kim and Youjia Zhang and Aecheon Jung and Kyuhong Shim and Sungeun Hong},
  year = {2026},
  eprint = {2610.24998},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.24998}
}

Suggest a correction on GitHub ↗