Tongyi Lab Launches VimRAG: A Multimodal RAG Framework with Memory Graph

ME News reports that on April 10 (UTC+8), Alibaba’s Tongyi Lab officially launched its next-generation multimodal RAG framework, VimRAG, designed to address the longstanding “state blind spot” issue in existing systems. VimRAG upgrades traditional linear historical records to a Multimodal Memory Graph, organizing the reasoning process with a dynamic directed acyclic graph (DAG) structure to effectively eliminate redundant retrieval and track exploration paths in real time. It introduces Graph-Modulated Visual Memory Encoding, enabling adaptive token allocation for high-load visual data such as images, combined with the GGPO mechanism to achieve fine-grained credit assignment and enhance reasoning attribution accuracy. According to published evaluation data, VimRAG demonstrates outstanding performance across multiple multimodal benchmarks including SlideVQA, MMLongBench, and LVBench, with the Qwen3-VL-8B-Instruct version achieving the highest overall score among comparable solutions. VimRAG aims to advance multimodal RAG from “simple retrieval” to “structured, reliable reasoning,” providing a stronger system-level solution for handling complex long-form documents and multimodal hybrid scenarios. (Source: BlockBeats)