Align-GRAG: Reasoning-Guided Dual Alignment for Graph Retrieval-Augmented Generation

TL;DR

Align-GRAG addresses two core challenges in graph-based retrieval-augmented generation: noisy retrieval introducing irrelevant graph content and a representation gap between graph structures and language models. It introduces a reasoning-guided dual alignment framework with a Graph Aligner that performs node-level pruning via KL divergence and representation alignment via contrastive learning, anchored by LLM-generated summarized reasoning chains. The approach enables efficient pruning and unified graph-language representations, leading to grounded, context-aware generation. Extensive experiments on GraphQA across commonsense reasoning, scene understanding, and knowledge graph reasoning demonstrate state-of-the-art performance and robust generalization, highlighting the value of integrating reasoning-guided graph alignment into GRAG pipelines.

Abstract

Large language models (LLMs) have demonstrated remarkable capabilities, but still struggle with issues like hallucinations and outdated information. Retrieval-augmented generation (RAG) addresses these issues by grounding LLM outputs in external knowledge with an Information Retrieval (IR) system. Building on this foundation, graph-based RAG systems go a step further by retrieving subgraphs, which preserve the relationships between knowledge entities and provide more comprehensive context. However, graph RAG faces two challenges: (1) Retrieving relevant information introduces irrelevant nodes (especially in dense graph databases, where retrieval usually extends to adjacent nodes), and leads to overly lengthy inputs that hinder efficiency; (2) The representation gap between graph and language during generation with LLMs limits the ability to fully leverage graph structures for enhanced understanding. To address these limitations, we propose Align-GRAG, a novel reasoning-guided dual alignment framework in post-retrieval phrase. It first formulates a subgraph by retrieving nodes and edges. Then an Aligner is proposed to jointly optimize a graph encoder with an LLM-summarized reasoning chain. It achieves dual alignment of graph node and representation by leveraging KL divergence loss and contrastive loss, facilitating efficient pruning of irrelevant knowledge and establishing a unified semantic space. The Generator integrates the aligned graph data with LLM to produce coherent and accurate answers. Experiments on the GraphQA benchmark across three tasks (including common sense reasoning, scene graph understanding, and knowledge graph reasoning) validate the effectiveness of our method. The codes are available in this repository\footnote{https://anonymous.4open.science/r/Align-GRAG-F3D8/}.

Figures (11)

• Figure 1: Comparison of Align-GRAG with traditional RAG (including post-retrieval methods such as reranker). Left: query and textual graph (entities and relations). Middle (a): Traditional RAG retrieves top‑k passages and may use a reranker, but still suffers from noisy/irrelevant evidence and loses the underlying graph structure, which hurts answer grounding. Right (b): Align‑GRAG introduces a graph aligner that jointly performs node alignment and representation alignment, preserving structural relations and filtering irrelevant content, showcasing its effectiveness in graph RAG scenarios.
• Figure 2: The Align-GRAG framework includes three components: ➀ Graph Retriever, ➁ Graph Aligner, and ➂ Graph Generator. It first retrieves subgraphs, then trains the aligner with an LLM-based summarized reasoning chain. The trained aligner prunes irrelevant nodes, generates graph embeddings, and feeds them into the LLM to produce the final answer.
• Figure 3: Ablation study with different alignment strategy.
• Figure 4: Representation Alignment Analysis: The cosine similarity score between graph embeddings and language embeddings (aligned using the aligner module vs. the unaligned setting).
• Figure 5: Hyperparameters Analysis of the Number of seed nodes.