LQA: A Lightweight Quantized-Adaptive Framework for Vision-Language Models on the Edge
Xin Wang, Hong Jia, Hualin Zhou, Sheng Guang Wang, Yu Zhang, Ting Dang, Tao Gu · 2026
Read LQA: A Lightweight Quantized-Adaptive Framework for Vision-Language Models on the Edge by Xin Wang, Hong Jia, Hualin Zhou, Sheng Guang Wang, Yu Zhang, Ting Dang, Tao Gu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-07849.txt.
Description
Vision-language models are too heavy for edge devices, and test-time adaptation methods that could handle distribution shifts are themselves too costly. Xin Wang and six co-authors propose LQA, a lightweight quantized-adaptive method combining modality-aware quantization with gradient-free adaptation.
Selective hybrid quantization shrinks the model while a quantized, gradient-free mechanism adapts it on device without backpropagation. Across made and real distribution shifts, LQA improves adaptation results by 4.5 percent, uses less memory than full-fine models, and beats gradient-based adaptation with up to 19.9 times lower memory use across seven datasets. Edge AI builders gain a usable path to robust, private, on-device vision-language models.
Catalog record
| CREATOR | Xin Wang, Hong Jia, Hualin Zhou, Sheng Guang Wang, Yu Zhang, Ting Dang, Tao Gu |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 53 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-07849