While Multimodal Retrieval-Augmented Generation (MM-RAG) has shown promising results, it still struggles with complex multi-hop reasoning tasks. Existing methods primarily focus on independent instance-level matching, which often fails to capture explicit relationships across modalities and documents. Although Graph-enhanced methods introduce structural modeling, they face a fundamental challenge in multimodal scenarios: incorporating fine-grained visual features leads to rapid graph expansion and retrieval noise, whereas coarse-grained representations cause the discarding of critical local evidence. To address this dilemma, we propose DualG-MRAG, a Dual-tier framework that introduces a decoupled architecture comprising Macro-reasoning and Micro-matching Graphs for Multimodal RAG. Specifically, to suppress retrieval noise by isolating global structural reasoning from fine-grained evidence matching, we construct a Macro Graph for global topological routing and a Micro Graph for precise local verification. Subsequently, to enable dynamic relevance propagation across heterogeneous evidence sources, we formulate retrieval as a query-driven message passing process via a GNN Retriever. Furthermore, to provide the generative model with coherent structural guidance, we introduce a dynamic programming decoding mechanism that extracts explicit reasoning paths directly from the GNN's forward pass, replacing the standard input of isolated document chunks. Extensive experiments demonstrate that DualG-MRAG outperforms baselines in both evidence recall and complex QA accuracy.
Jiachen Tao, Qingyun Sun, Haonan Yuan et al.· 0 citations
Graph Structure Learning (GSL) is fundamental to effective graph-based modeling, particularly for life omics analysis. Traditional heuristics-based GSL methods impose overly rigid inductive biases, while optimization centric GSL suffers from non-convexity and inherent interpretability challenges. Evolutionary algorithms (EAs) have emerged as a promising alternative for their robust global exploration and ability to navigate discrete search spaces without gradient information. However, such stochastic mutations essentially bottleneck evolutionary efficiency. Furthermore, both optimization-based and existing evolutionary paradigms lack semantic guidance, resulting in inexplainable graph structures. Specifically, the topological modifications tend to be arbitrary, failing to reveal the underlying algorithmic logic required for biological insights. To address these problems, we propose CEAgent-GSL, a Code-level Evolutionary LLM Agent framework. CEAgent-GSL treats graph generation as an evolving population of executable scripts, leveraging an LLM for semantic mutation guided by task feedback. Recognizing that omics analysis demands not only high performance but also transparent and verifiable construction principles, we introduce Evolutionary Rule Abstraction to distill generalized logic from optimization trajectories, overcoming the opacity of traditional evolution. Furthermore, to ensure the scientific rigorousness of these discovered principles, we design a Counterfactual Validation Module to strictly verify the causal reliability of abstracted rules. Through this transparent evolutionary paradigm, CEAgent-GSL overcomes the limitations of conventional parametric tuning, and facilitates highly efficient and interpretable GSL, rigorously validated across 11 foundational bioinformatics algorithms on 10 diverse omics datasets. Besides individual task optimization, it can seamlessly integrate domain expertise with task-specific requirements by retrieving and adapting proven logic, thus generating customized rules for novel scenarios.
Zhongyu Xing, Hanwen Luo, Maozu Guo et al.· Proceedings of the 32nd ACM...· 0 citations