Safety-Preserving PTQ via Contrastive Alignment Loss
Fuente:
arXiv
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866915870447501312 |
|---|---|
| author | Wee, Sunghyun Kim, Suyoung Kim, Hyeonjin Hwang, Kyomin Kwak, Nojun |
| author_facet | Wee, Sunghyun Kim, Suyoung Kim, Hyeonjin Hwang, Kyomin Kwak, Nojun |
| contents | Post-Training Quantization (PTQ) has become the de-facto standard for efficient LLM deployment, yet its optimization objective remains fundamentally incomplete. Standard PTQ methods minimize reconstruction error (e.g., MSE or KL divergence) without accounting for behavioral alignment--a critical property instilled through safety fine-tuning. We demonstrate that this objective mismatch introduces a systematic vulnerability: models can maintain low perplexity yet exhibit significant degradation in safety alignment, revealing that perplexity alone is an insufficient and often misleading proxy for deployment readiness. To address this, we propose Contrastive Alignment Quantization (CAQ), which extends the PTQ objective design space by integrating a Contrastive Alignment Loss (CAL). CAL introduces a principled push-pull mechanism that jointly optimizes distributional fidelity and behavioral alignment: it steers the quantized model toward its safe, instruction-tuned counterpart while diverging from the unaligned, pre-trained reference. CAQ requires no specialized safety datasets, relying solely on standard calibration data, and introduces negligible computational overhead over existing transformation-based PTQ pipelines. We show that CAQ enables robust 4-bit (W4A4) quantization across diverse model families--including LLaMA, Qwen, and Mistral--achieving superior safety alignment where state-of-the-art PTQ methods fail, without sacrificing general capabilities. Anonymized code is available in the supplementary material. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2511_07842 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Safety-Preserving PTQ via Contrastive Alignment Loss Wee, Sunghyun Kim, Suyoung Kim, Hyeonjin Hwang, Kyomin Kwak, Nojun Artificial Intelligence Post-Training Quantization (PTQ) has become the de-facto standard for efficient LLM deployment, yet its optimization objective remains fundamentally incomplete. Standard PTQ methods minimize reconstruction error (e.g., MSE or KL divergence) without accounting for behavioral alignment--a critical property instilled through safety fine-tuning. We demonstrate that this objective mismatch introduces a systematic vulnerability: models can maintain low perplexity yet exhibit significant degradation in safety alignment, revealing that perplexity alone is an insufficient and often misleading proxy for deployment readiness. To address this, we propose Contrastive Alignment Quantization (CAQ), which extends the PTQ objective design space by integrating a Contrastive Alignment Loss (CAL). CAL introduces a principled push-pull mechanism that jointly optimizes distributional fidelity and behavioral alignment: it steers the quantized model toward its safe, instruction-tuned counterpart while diverging from the unaligned, pre-trained reference. CAQ requires no specialized safety datasets, relying solely on standard calibration data, and introduces negligible computational overhead over existing transformation-based PTQ pipelines. We show that CAQ enables robust 4-bit (W4A4) quantization across diverse model families--including LLaMA, Qwen, and Mistral--achieving superior safety alignment where state-of-the-art PTQ methods fail, without sacrificing general capabilities. Anonymized code is available in the supplementary material. |
| title | Safety-Preserving PTQ via Contrastive Alignment Loss |
| topic | Artificial Intelligence |
| url | https://arxiv.org/abs/2511.07842 |