Metraipolis › archive › dc_6bcb8631fa0c

Quantizing Looped Transformers: Feedback Exposure and Calibration Blindness — arXiv:2609.30820

Source
arXiv (en)
Original
https://arxiv.org/abs/2609.30820
License (as archived)
CC0-1.0
Archived
2026-10-08T02:04:15.529Z
SHA-256 of the archived text
2a9239932c278d758f8d05f6a0f4790f9c1dc2b4fdde44f924bd721eabff2874
Stable URL of this page
https://metraipolis.com/documento/dc_6bcb8631fa0c

How to cite

Quantizing Looped Transformers: Feedback Exposure and Calibration Blindness — arXiv:2609.30820 — arXiv (en), https://arxiv.org/abs/2609.30820 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_6bcb8631fa0c (Metraipolis archive dc_6bcb8631fa0c, sha256 2a9239932c278d758f8d05f6a0f4790f9c1dc2b4fdde44f924bd721eabff2874)
BibTeX
@misc{dc6bcb8631fa0c,
  title = {Quantizing Looped Transformers: Feedback Exposure and Calibration Blindness — arXiv:2609.30820},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2609.30820},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_6bcb8631fa0c; sha256 2a9239932c278d758f8d05f6a0f4790f9c1dc2b4fdde44f924bd721eabff2874},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2609.30820. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Looped transformers reuse weights across recurrence steps, making low-bit quantization especially attractive. We identify two distinct failure modes of standard post-training quantization. On Huginn-3.5B, per-channel INT4 fails primarily at the non-residual loop-entry adapter, while quantizing the residual core is much less damaging. We call this feedback exposure: a quantized layer perturbs the recurrent state without an identity path, and the resulting error is fed back at later steps. Controlled experiments on linear filters and Mamba state-space models show that feedback exposure also occurs outside transformers. Grouped INT4 reveals a separate failure, calibration blindness: our one-step GPTQ baseline builds its Hessian from step-0 activations, leaving input directions used later in the recurrence nearly unweighted. Across nine checkpoints from seven looped architectures, one-step GPTQ is worse than round-to-nearest (RTN) on the primary task metric for five checkpoints. Accumulating the GPTQ Hessian across recurrence steps outperforms both one-step GPTQ and RTN on all nine checkpoints and recovers bf16-level accuracy on Huginn. These results separate two questions for PTQ on looped models: where quantization error enters the recurrence, and which states calibration sees.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-09-28 in cs.CL).
Categorie arXiv: Machine Learning (cs.LG) ; Computation and Language (cs.CL).
Pagina del lavoro: https://arxiv.org/abs/2609.30820 — 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 2026-10-08T01:38:49.484Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-09-28?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).