EvidenceDemonstrated research proof-of-conceptv1.22.1
In plain English
This page shows what kind of support exists for each claim: real systems, experiments, early evidence, architectural reasoning, open questions, or speculative scenarios.
- Why this matters: AI risk can come from the whole arrangement, not one obvious model.
- What to look for: data, memory, routes, adapters, tools, evaluators, updates, and rollback paths.
- Technical version below: the expert terminology remains available and is linked through the glossary.
Alignment-Aware Quantization for LLM Safety
Evidence card
- Claim
- Conventional low-perplexity quantization objectives can miss safety degradation in studied models.
- Evidence level
- Emerging evidence
- Source
- https://arxiv.org/abs/2511.07842
- Publication date
- 2025-11-11
- Authors or institution
- Sunghyun Wee, Suyoung Kim, Hyeonjin Kim, Kyomin Hwang, Nojun Kwak
- System tested
- Post-training quantization with alignment-preserving contrastive loss across LLaMA, Qwen, and Mistral families as reported.
- Limitations
- Preprint; method scope and robustness require independent replication.
- What the evidence does show
- Conventional low-perplexity quantization objectives can miss safety degradation in studied models.
- What the evidence does not show
- That all low-bit deployments are unsafe or that AAQ is sufficient in every setting.
- Date last reviewed in UTC
- 2026-06-26T00:00:00Z
Site use
This source supports Cognivirus.com pages related to quantization, alignment preservation, safety metrics. Its role is bounded by the limitations listed above.