PrefMoE Prevents Preference Collapse in Personalized MLLMs.

Personalized Multimodal Large Language Models (MLLMs) aim to generate responses tailored to individual users. However, current methods often rely on broad profile information and fail to capture the nuances of diverse user preferences. This leads to a phenomenon called "group preference collapse," where MLLMs become insensitive to individual preferences, drifting towards dominant population-level choices due to suppressed preference signals. Researchers have identified that this collapse stems from unreliable preference use during generation. To counter this, they propose PrefMoE, a preference-centric framework designed to maintain individual preference sensitivity. PrefMoE separates stable profile information from dynamic preference-related representations. The framework decomposes preferences into shared prototypes and personalized residuals, preserving these individualized residuals through techniques like imbalance-aware learning, counterfactual pseudo-user augmentation, and residual decorrelation. It then routes these distinct profile and preference factors through separate LoRA adaptation paths. Experiments across various MLLM backbones demonstrate that PrefMoE significantly improves preference-sensitive personalization and effectively reduces group preference collapse.