Research map / Graph learning / Knowledge and temporal graphs
G-Retriever: Retrieval-Augmented Generation for Textual Graph Understanding and Question Answering
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces G-Retriever, a retrieval-augmented, soft-prompted LLM framework that answers questions over real-world textual graphs by retrieving a PCST-optimized subgraph, along with the GraphQA benchmark built from ExplaGraphs, SceneGraphs, and WebQSP.
Problem
Answering natural-language questions over large, real-world textual graphs is hard: existing graph-LLM work targets classification or toy graph queries, and textualizing whole graphs overflows LLM context windows and induces hallucinated nodes and edges.
Approach
Formulate subgraph retrieval for a query as a Prize-Collecting Steiner Tree problem over node and edge embeddings, then feed the retrieved subgraph to a frozen LLM via a GNN-encoded soft graph token plus a textualized subgraph, with optional LoRA tuning.
Key results
- G-Retriever with LoRA reaches 0.8705 on ExplaGraphs, 0.8683 on SceneGraphs, and 73.79 Hit@1 on WebQSP, the best among compared configurations.
- On WebQSP, PCST retrieval cuts average nodes by 99% and training time from 18.7 to 6.2 min/epoch (frozen LLM with prompt tuning setting).
- In a manual hallucination check on SceneGraphs, fully valid graphs rise from 8% (graph prompt tuning baseline) to 62% for G-Retriever.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add graph encoder (ablation) | G-Retriever · WebQSP | Hit@1: 54.62 ± 0.78 -> 70.49 ± 1.21 | Helped |
| Add projection layer (ablation) | G-Retriever · WebQSP | Hit@1: 69.70 ± 0.68 -> 70.49 ± 1.21 | Helped |
| Include the textualized graph alongside the question (ablation) | G-Retriever · WebQSP | Hit@1: 56.96 ± 1.83 -> 70.49 ± 1.21 | Helped |
| Add graph retrieval (ablation) | G-Retriever · WebQSP | Hit@1: 63.84 ± 0.41 -> 70.49 ± 1.21 | Helped |
| Increase retrieval k from 3 to 5 (ablation) | G-Retriever · WebQSP | Hit@1: 0.6977 -> 0.7063 | Helped |
| Increase retrieval k from 10 to 20 (ablation) | G-Retriever · WebQSP | Hit@1: 0.7248 -> 0.7039 | Hurt |
| Replace Graph Transformer with GCN (ablation) | G-Retriever · WebQSP | Hit@1: 70.49 -> 70.70 | No clear effect |
| Replace Graph Transformer with GAT (ablation) | G-Retriever · WebQSP | Hit@1: 70.49 -> 70.27 | No clear effect |
| Increase LLM size from Llama2-7b to Llama2-13b (ablation) | G-Retriever · WebQSP | Hit@1: 70.49 -> 75.58 | Helped |
Surprising findings
- Inference-only LLMs given the question alone outperform those given the textualized graph on WebQSP (61.16 vs 41.06 Hit@1), attributed to noise in the retrieved knowledge.
- Graph prompt tuning with a frozen LLM hallucinates nodes and edges frequently, with only 8% of graphs fully valid on SceneGraphs.
Limitations
- The retrieval component is static and not trainable.
- Hallucination evaluation relies on manual review of only 100 responses.
- ExplaGraphs and SceneGraphs are re-purposed from existing datasets rather than newly collected.
Experimental setup
Datasets: ExplaGraphs SceneGraphs WebQSP GQA
Benchmarks: GraphQA (ExplaGraphs, SceneGraphs, WebQSP)
Metrics: accuracy Hit@1 valid node/edge rate tokens and training time
Baselines: Zero-shot Zero-CoT CoT-BAG KAPING Prompt tuning GraphToken LoRA
Builds on: KAPING GraphToken MiniGPT-4 LoRA Sentence-BERT
Closest papers by meaning
- Simple is Effective: The Roles of Graphs and Large Language Models in Knowledge-Graph-Based Retrieval-Augmented Generation (ICLR 2025 · similarity 0.90)
- Think-on-Graph 2.0: Deep and Faithful Large Language Model Reasoning with Knowledge-guided Retrieval Augmented Generation (ICLR 2025 · similarity 0.90)
- RAGraph: A General Retrieval-Augmented Graph Learning Framework (NeurIPS 2024 · similarity 0.89)
- Talk like a Graph: Encoding Graphs for Large Language Models (ICLR 2024 · similarity 0.89)
- GFM-RAG: Graph Foundation Model for Retrieval Augmented Generation (NeurIPS 2025 · similarity 0.88)
- Fast Think-on-Graph: Wider, Deeper and Faster Reasoning of Large Language Model on Knowledge Graph (AAAI 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.