Metraipolis › archive › dc_52cf1e13de26

UpLoop: Training Looped Transformers via Upcycling at No Extra Pretraining Compute — arXiv:2610.41720

Source
arXiv (en)
Original
https://arxiv.org/abs/2610.41720
License (as archived)
CC0-1.0
Archived
2026-11-03T07:55:26.030Z
SHA-256 of the archived text
6d4d2c72db1b96d7ceb048c4abde71cede792414915b74b255bb5ff5b01c8834
Stable URL of this page
https://metraipolis.com/documento/dc_52cf1e13de26

How to cite

UpLoop: Training Looped Transformers via Upcycling at No Extra Pretraining Compute — arXiv:2610.41720 — arXiv (en), https://arxiv.org/abs/2610.41720 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_52cf1e13de26 (Metraipolis archive dc_52cf1e13de26, sha256 6d4d2c72db1b96d7ceb048c4abde71cede792414915b74b255bb5ff5b01c8834)
BibTeX
@misc{dc52cf1e13de26,
  title = {UpLoop: Training Looped Transformers via Upcycling at No Extra Pretraining Compute — arXiv:2610.41720},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2610.41720},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_52cf1e13de26; sha256 6d4d2c72db1b96d7ceb048c4abde71cede792414915b74b255bb5ff5b01c8834},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2610.41720. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Looped language models can strengthen reasoning through repeated computation with shared weights, increasing execution depth without adding independent parameters. However, each additional pass increases training floating-point operations (FLOPs), making these gains more expensive to obtain. To address this cost, we propose UpLoop, a pretraining schedule that uses upcycling to train looped Transformers while exactly matching a standard dense baseline in final parameter count, total training tokens, and training FLOPs. Specifically, UpLoop begins with a shallow Transformer, progressively grows its depth by copying learned layers, and uses the computation saved early to support repeated execution of a shared middle block later in training. Experiments across model scales show improvements in language modeling and downstream performance, with the average score across 22 downstream tasks increasing from 22.99 to 27.52 at 1.3B parameters. These gains also extend to mathematical reasoning after supervised fine-tuning.

Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.CL, nuovi lavori del giorno).
Categorie arXiv: Machine Learning (cs.LG) ; Computation and Language (cs.CL).
Pagina del lavoro: https://arxiv.org/abs/2610.41720 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-11-03T07:49:46.030Z dalla pagina https://arxiv.org/list/cs.CL/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).