CoRe-MMRAG: Cross-Source Knowledge Reconciliation for Multimodal RAG
Reconcile what a model knows with what it retrieves, bringing visual and textual evidence together for knowledge-based question answering.
Research
Our work connects multimodal understanding, foundation models, and autonomous intelligence.
Multimodal understanding
Learning from visual, textual, and behavioral signals, with a focus on representations and evidence across knowledge sources.
Multimodal representation · Knowledge reconciliation · Recommendation
Attribute-driven Disentangled Representation Learning for Multimodal Recommendation
Learn disentangled representations guided by item attributes, connecting visual and textual information for recommendation.
Foundation models
Understanding and using the semantic capabilities of large models for retrieval, recommendation, and learning from sparse data.
Language models · Semantic learning · Information retrieval
A Non-parametric Multi-view Model for Short Text Clustering
Combine word-level evidence with language-model embeddings to cluster short texts without fixing the number of clusters in advance.
Understanding Before Recommendation: Semantic Aspect-Aware Review Exploitation via Large Language Models
Use large language models to extract semantic aspects from reviews and build more informative representations for recommendation.
Autonomous intelligence
Agents that plan, use tools, and interact with the world. Current work explores multimodal retrieval and evidence-grounded generation.
Tool use · Context management · Agentic retrieval
MM-ContextFold: Context Folding for Multimodal Agentic Retrieval
Keep the plan, load images when needed. A context-folding framework for agents working through extended multimodal retrieval tasks.
Open Multimodal Retrieval-Augmented Factual Image Generation
Retrieve real-world evidence before generating an image. An agentic approach to improving factual consistency in visual generation.