Researchers have proposed a method to personalize vision-language models without retraining them. The approach involves providing user profiles or visual references at inference time and using a calibrated residual decoding framework to estimate the contribution of personalization. This method improves personalized multimodal understanding on several benchmarks, including identity-sensitive tasks. --- Why it matters: This matters because it enables the creation of more effective and efficient vision-language models that can adapt to individual users without requiring extensive retraining or updates. Source: https://arxiv.org/abs/2608.22263