LLM+Graph@VLDB'2025 Workshop Summary

Abstract: The integration of LLMs with graph-structured data has become a pivotal and fast evolving research frontier, drawing strong interest from both academia and industry. The 2nd LLM+Graph Workshop, co-located with the 51st International Conference on Very Large Data Bases (VLDB 2025) in London, focused on advancing algorithms and systems that bridge LLMs, graph data management, and graph machine learning for practical applications. This report highlights the key research directions, challenges, and innovative solutions presented by the workshop's speakers.

• The paper presents how LLMs improve graph querying and repair while graph approaches boost LLM reliability and indexing.
• It details methodologies like Riemannian geometry for advanced graph learning and LLM reasoning frameworks for property graph consistency.
• Practical case studies from ByteDance, Neo4j, and Ant Group demonstrate scalable, unified models and highlight ongoing research challenges.

Workshop Overview: Integration of LLMs and Graph Technologies

The LLM+Graph@VLDB'2025 Workshop reported extensive progress and open challenges in synergizing LLMs with graph-structured data. Addressing both academic and industrial perspectives, the workshop highlighted recent advancements in algorithms, systems integration, and practical deployments across diverse domains, including data management, graph learning, and agentic AI. The proceedings underscored the dual capability of LLMs to enhance graph-centric tasks and the reciprocal value of graphs in augmenting LLM reliability and expressiveness.

Keynote Highlights

Three keynotes encapsulated the field's theoretical and practical trajectories.

Duality Between LLMs and Databases:

Prof. M. Tamer Özsu examined the intersection of LLMs and database systems, advocating for a symbiotic relationship. In LLM4DB, LLMs facilitate data management tasks such as Natural Language Querying and semantic evaluation (e.g., SQLyzr for NL2SQL), while embedded LLM functions yield efficient inference over database content. The DB4LLM direction stressed the foundational role of databases in LLM empowerment, especially in vector retrieval and indexing. MIRAGE, a graph-based ANNS solution, was introduced, achieving high indexing and search performance. The keynote emphasized multimodal RAG and the imperative for complex reasoning, concluding that the data management community must confront hallucinations and operational costs to realize the technology's transformative potential.

Graph Foundation Models via Riemannian Geometry:

Prof. Philip S. Yu argued the inadequacy of Euclidean GNNs and advanced Riemannian Geometric Deep Graph Learning for more expressive modeling. Heterogeneous curvature spaces and product manifolds were proposed to capture diverse graph geometries, enabling structural abstraction and self-supervised learning. Demonstrations included Ricci curvature for clustering (Congregate, RicciNet), physics-informed modeling (Pioneer), and transferable localization (Trace). The RiemannGFM concept proposes a universal structural vocabulary mapped to Riemannian geometry, facilitating generalizable models that overcome oversmoothing and oversquashing. This paradigm shift opens pathways to structurally aware, scalable graph foundation models.

LLM Reasoning for Property Graph Consistency:

Prof. Angela Bonifati analyzed automated and user-centric property graph repairs. The cQAR framework ensures collaborative, convergent repair, but empirical studies revealed LLMs are unreliable for unsupervised graph inconsistency correction, with low correctness rates despite valid formats. Common failure modes included hallucination and indecision. In rule generation, LLMs produced redundant and sometimes meaningless rules; a RAG pipeline was presented to filter and generate actionable Cypher queries. Theoretical analysis posited the necessity of explainability and human-in-the-loop systems until LLMs satisfy data consistency requirements.

Industrial Applications

ByteDance:

GraphRAG and graph foundation models were operationalized at massive scale. Dual-Granularity Prompting enables efficient fraud detection: target node detail is preserved, and metapath-based neighbors are summarized, improving prediction in complex contexts. Over 50 GraphRAG scenarios are deployed, with long-term vision focused on unified foundation models.

Neo4j:

Innovations include GraphRAG pipelines combining vector retrieval and subgraph pruning for GNN+LLM tasks, yielding superior retrieval accuracy but raising challenges such as neighborhood explosion and multi-hop answer identification. Text-to-Cypher projects contributed datasets and models, optimizing inference with schema filters. The GDS Agent autonomously applies graph algorithms for analytical reasoning, but scaling, benchmarking, and global planning remain open issues.

Ant Group:

Chat2Graph was presented as a graph-native, agentic reasoning system. With $10^{13}$ edges, the architecture incorporates Planning, Toolkit, and Memory graphs for strategic planning, tool recommendation, and layered memory management. Alipay voice assistant uses GraphRAG, but Chat2Graph demonstrates broader applicability for complex analytic tasks. Open sourcing furthers ecosystem development.

Peer-Reviewed Research Themes

LLMs in Graph Construction and Reasoning:

• Colombo et al. leveraged fine-tuned LLMs for knowledge graph construction from unstructured legislative documents.
• Ji et al. evaluated twelve LLMs for taxonomic reasoning, offering systematic test cases.
• Publio et al. delivered explainable SHACL validation with multi-language contextual RAG explanations.
• Mihindukulasooriya et al. demonstrated prompt optimization techniques (DSPy, APE) for triple extraction, statistically improving KG construction quality.

Graphs for LLM Enhancement and Retrieval:

• Zhang et al. deployed Locality-Sensitive Hashing to scale tree-based Graph-RAG retrieval.
• Schneider et al. identified gaps in LLM spatial reasoning and proposed graph integration for search engines targeting complex spatial queries.
• Zhou et al. conceptualized agentic systems composed of planning, execution, knowledge, and tools, mapping future research in agent-centric architectures.

Panel Discussion and Future Directions

The panel synthesized viewpoints on collaboration versus competition between LLMs and graph technologies. Panelists rejected the notion of LLMs superseding graph databases, instead promoting natural language interfaces and hybrid feedback mechanisms. The enduring necessity for expertise in schema design was underscored, given LLMs' current limitations in knowledge graph construction.

Forward-looking areas were identified:

• Foundational Models: Universal structural vocabularies and deep geometric learning for transferrable reasoning.
• Scalable Data Systems: Integration of LLMs with ever-larger, dynamic graphs and hybrid queries involving vector and graph indices.
• Trustworthy AI: Automated data repair, explainable outputs, and interactive frameworks promoting transparency.

Conclusion

The LLM+Graph@VLDB'2025 Workshop documented major strides in bridging LLMs and graph data management, learning, and reasoning. While LLMs have demonstrated utility across querying, prompt optimization, and rule generation, they are constrained in unsupervised graph repair and structural comprehension. Graph-enhanced retrieval, explanatory pipelines, and agentic architectures are establishing new standards, but demand further scalability, reliability, and explainability. The symbiosis of LLMs and graphs is poised to redefine AI systems, with unresolved theoretical and practical questions driving ongoing research in scalable infrastructures, trust, and foundational modeling (2604.02861).