Researchers have developed a method called Doubt-Preserving Quantization (DPQ) to preserve uncertainty in language models when they are quantized for deployment. The goal is to ensure that the model's confidence and margins are preserved after quantization. DPQ uses full-precision predictions to construct calibration mixtures of high-doubt examples and generic anchors, tailored to specific deployment targets. The method has been tested on 8 language models and 9 NLP benchmarks, showing improved preservation of uncertainty in various scenarios. --- Why it matters: This matters because quantization can have unintended effects on the uncertainty behavior of large language models, which is crucial for applications such as confidence-based decision-making and risk assessment. By preserving uncertainty, DPQ enables more reliable deployment of these models in real-world settings. Source: https://arxiv.org/abs/2608.21019