[Submitted on 3 Apr 2026 (v1), last revised 5 Nov 2026 (this version, v4)]
Abstract:We show that robustness to post-training quantization (PTQ) is a transferable direction in weight space. We call this direction the quantization vector: extracted from a donor task by simple weight-space arithmetic, it can be used to patch a receiver model and improve PTQ Top-1 accuracy by up to 60% in a 3-bit setting, without receiver-side quantization-aware training (QAT). When using no receiver validation data, the method provides a zero-shot, low-cost alternative to QAT for extremely low-bit deployment. Across multiple vision and language models and more than 30 tasks, donor quantization vectors often yield substantial gains even when donor and receiver tasks differ markedly. We further prove rigorously that quantization vectors are well-defined and do not suffer from reparameterization symmetries, and provide a local geometric account of their effect. Together, these results suggest that quantization robustness can be partially isolated, reused, and transferred through simple weight-space algebra.
Submission history
From: Daniele Solombrino [view email]
[v1] Fri, 3 Apr 2026 19:33:25 UTC (834 KB)
[v2] Sat, 11 Apr 2026 19:49:30 UTC (844 KB)
[v3] Fri, 8 May 2026 06:46:47 UTC (2,627 KB)
[v4] Thu, 5 Nov 2026 14:09:30 UTC (1,913 KB)
