Alibaba Tongyi Laboratory Releases VimRAG: Reconstructing Multimodal Retrieval and Reasoning with Memory Graphs

CryptoWorld News reports, ME News reports, on April 10 (UTC+8), Alibaba Tongyi Laboratory (Tongyi Lab) officially launched the new-generation multimodal RAG framework VimRAG, focusing on tackling the long-standing “state blind spots” problem in existing systems. VimRAG upgrades traditional linear history records into a Multimodal Memory Graph, organizing the reasoning process with a dynamic directed acyclic graph (DAG) structure, effectively eliminating redundant retrieval and tracking exploration paths end to end. It introduces Graph-Modulated Visual Memory Encoding, which targets high-load visual data such as images, achieves adaptive Token allocation, and incorporates the GGPO mechanism to enable fine-grained credit assignment and improve the accuracy of reasoning attribution. Based on published evaluation data, VimRAG performs exceptionally well across multiple multimodal benchmark tests such as SlideVQA, MMLongBench, and LVBench, with the Qwen3-VL-8B-Instruct version achieving the leading overall score among comparable solutions. VimRAG aims to move multimodal RAG from “simple retrieval” to “structured reliable reasoning,” providing a stronger system-level solution for handling complex long documents and multimodal mixed scenarios.