Metraipolis › archive › dc_41f4e0aa2f86

Hybrid Latent Attention for Looped Language Models — arXiv:2610.07940

Source
arXiv (en)
Original
https://arxiv.org/abs/2610.07940
License (as archived)
CC0-1.0
Archived
2026-10-07T11:21:22.730Z
SHA-256 of the archived text
064ab7ff01a843f46f10b8648120ace31880717bde63da825eb61b7ac9d3aa74
Stable URL of this page
https://metraipolis.com/documento/dc_41f4e0aa2f86

How to cite

Hybrid Latent Attention for Looped Language Models — arXiv:2610.07940 — arXiv (en), https://arxiv.org/abs/2610.07940 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_41f4e0aa2f86 (Metraipolis archive dc_41f4e0aa2f86, sha256 064ab7ff01a843f46f10b8648120ace31880717bde63da825eb61b7ac9d3aa74)
BibTeX
@misc{dc41f4e0aa2f86,
  title = {Hybrid Latent Attention for Looped Language Models — arXiv:2610.07940},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2610.07940},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_41f4e0aa2f86; sha256 064ab7ff01a843f46f10b8648120ace31880717bde63da825eb61b7ac9d3aa74},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2610.07940. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Looped language models apply the same stack of layers T times to each token, which deepens the model without adding parameters but multiplies its key-value (KV) cache by T. The larger cache limits how many sequences a GPU can decode at once and slows each decoding step, which reads the whole cache. We propose Hybrid Latent Attention (HLA), which keeps exact keys and values within a sliding window of W recent tokens and stores each older token as a compact latent that the query of each loop reads directly, without reconstructing keys and values. We uptrain HLA on Ouro looped models (T=4) with 1.4B and 2.6B parameters, keeping the pretrained weights frozen and training only the added parameters to reproduce the original attention. The cache shrinks by 10.7x per token, fitting 4.0-8.8x as many concurrent sequences per GPU, and decoding throughput improves by 2.5x at 1K-token contexts and by up to 7.4x at 16K. HLA retains over 97% of the original accuracy on math, knowledge and reasoning benchmarks, and 96-100% on long-context retrieval up to 16K tokens. After supervised fine-tuning, it performs on par with the fine-tuned original model on competition-level math.

Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.AI, nuovi lavori del giorno).
Categorie arXiv: Computation and Language (cs.CL) ; Artificial Intelligence (cs.AI).
Pagina del lavoro: https://arxiv.org/abs/2610.07940 — 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 2026-10-07T11:15:24.786Z dalla pagina https://arxiv.org/list/cs.AI/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).