PersonaVLM
A research framework that turns general multimodal LLMs into personalized assistants able to remember, reason over, and align with a user's traits across long-term interaction.
PersonaVLM is a personalized multimodal agent framework for long-term personalization, built on top of general multimodal LLMs such as Qwen2.5-VL. It works through three capabilities: remembering, which extracts and summarizes multimodal conversation history into structured core, semantic, episodic, and procedural memories; reasoning, which retrieves and integrates relevant long-term memories during multi-turn conversation; and response alignment, which infers a user's evolving traits through a momentum-based personality mechanism.
The project also introduces Persona-MME, a benchmark of over 2,000 curated interaction cases across 14 personalization tasks used to evaluate long-term personalization in multimodal models.
It ships as a research implementation with model weights, training data, and the benchmark, alongside an interactive Gradio demo.