safety
fact
bearish
Quantization of large language models increases bias in open-ended responses while passing standard safety evaluations, with roughly one in four open-ended answers containing stereotypes across eight languages
Computation and Language28 Jul 2026