Researchers have proposed a new method to improve the robustness of large language models when they are quantized, or converted into lower-precision formats. This is done by injecting noise into the model's pre-attention logits, with the amount of noise determined by the model's sensitivity to discretization errors. The authors claim that their approach outperforms other methods in certain scenarios, achieving up to 37% relative gains in accuracy on ImageNet-1K and improving perplexity by up to 40% on WikiText. --- Why it matters: This matters because large language models are often deployed in low-precision formats to reduce memory and computational costs. However, this can lead to decreased performance due to quantization errors. The proposed method provides a way to identify the optimal amount of noise to inject into the model, which could improve its robustness and accuracy in these scenarios. Source: https://arxiv.org/abs/2608.20988