Metraipolis › archive › dc_ae07f89d03aa
Language-Conditional Dequantization: Recovering What Quantization Steals from Non-English Languages — arXiv:2608.11786
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2608.11786
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-19T05:27:37.659Z
- SHA-256 of the archived text
fe8ab7cd03499a85b41fb9d58f60967678205ac49a1b1653f02942be274ecfc6- Stable URL of this page
https://metraipolis.com/documento/dc_ae07f89d03aa
How to cite
Language-Conditional Dequantization: Recovering What Quantization Steals from Non-English Languages — arXiv:2608.11786 — arXiv (en), https://arxiv.org/abs/2608.11786 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_ae07f89d03aa (Metraipolis archive dc_ae07f89d03aa, sha256 fe8ab7cd03499a85b41fb9d58f60967678205ac49a1b1653f02942be274ecfc6)
BibTeX
@misc{dcae07f89d03aa,
title = {Language-Conditional Dequantization: Recovering What Quantization Steals from Non-English Languages — arXiv:2608.11786},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2608.11786},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_ae07f89d03aa; sha256 fe8ab7cd03499a85b41fb9d58f60967678205ac49a1b1653f02942be274ecfc6},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2608.11786. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Aggressive quantization disproportionately harms multilingual capability: in the sub-4B INT3 GPTQ regime, we measure 2-4x larger perplexity degradation on non-English languages than on English. We propose Language-Conditional Dequantization (LCD), a post-hoc method that attaches per-language rank-2 LoRA corrections to the linear layers of an already-quantized model, adding 0.12% parameters per language and training in under 20 minutes on a single GPU. Across Qwen2.5-3B and Llama-3.2-3B, LCD recovers 70-83% of the perplexity gap for non-Latin script languages and 17-28% of the GlobalMMLU accuracy gap, outperforming a language-agnostic correction of equal capacity by 3-9 points on typologically distant languages and a data-free low-rank baseline (LQER) by an order of magnitude. We further identify a perplexity-accuracy disconnect and trace it to where quantization concentrates damage: early-depth errors (Llama) propagate downstream and resist local correction, while late-depth errors (Qwen) do not. A layer-restricted variant of LCD validates this mechanism directly. Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-08-13 in cs.CL). Categorie arXiv: Computation and Language (cs.CL). Pagina del lavoro: https://arxiv.org/abs/2608.11786 — 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 2026-10-19T05:15:39.215Z dalla pagina https://arxiv.org/catchup/cs.CL/2026-08-13?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).