ClareNie/PersonaVLM
PersonaVLM by ClareNie is a 7 billion parameter personalized multimodal LLM framework accepted at CVPR 2026, built upon Qwen2.5-VL-7B. It specializes in long-term personalization by remembering, reasoning with a multi-type memory architecture, and aligning responses using a Momentum-based Personality Evolving Mechanism. This model is designed to transform general-purpose MLLMs into personalized assistants capable of multi-turn reasoning and evolving user personality inference.
Loading preview...
PersonaVLM: Long-Term Personalized Multimodal LLMs
PersonaVLM is an innovative 7 billion parameter personalized multimodal LLM framework developed by ClareNie, accepted at CVPR 2026. Built on Qwen2.5-VL-7B, its core innovation lies in enabling long-term personalization for multimodal agents. The framework integrates three key capabilities:
Key Capabilities
- Remembering: Proactively extracts and summarizes multimodal memories into a personalized database.
- Reasoning: Conducts multi-turn reasoning by retrieving relevant memories from a sophisticated multi-type memory architecture, including core, semantic, episodic, and procedural memory.
- Response Alignment: Infers and adapts to the user's evolving personality through a novel Momentum-based Personality Evolving Mechanism (PEM), ensuring outputs are consistently aligned with user preferences.
Benchmarking and Resources
To evaluate its personalization capabilities, PersonaVLM introduces Persona-MME, a comprehensive benchmark featuring over 2,000 curated interaction cases across 14 fine-grained tasks. The project provides access to model weights, training data (80k+ samples), the Persona-MME benchmark, and official code. This model is particularly suited for applications requiring an MLLM to maintain and adapt to a user's long-term persona and interaction history.