arxivcs.LGcs.AI2026-06-28
The Joint Effect of Quantization and Sampling Temperature on LLM Safety Alignment: A Factorial Analysis
Modern LLM deployments often combine quantization with higher sampling temperatures to reduce cost, latency, or repetition, yet safety evaluations usually treat these as fixed implementation details. We test whether models that are safe at FP16 with greedy decoding remain safe af…