Leveraging Large Language Models for Graph Learning: Opportunities, Current Progress, and Future Directions

Why apply large language models (LLMs) to graph learning? LLMs excel at text understanding and reasoning, and many graph tasks involve rich textual information or can be expressed as text, making LLMs a promising tool for graph inference.

Graph data characteristics are divided into three categories: Pure Graphs (no text, e.g., traffic networks), Text‑Paired Graphs (graphs with accompanying text, e.g., molecular structures), and Text‑Attributed Graphs (nodes contain abundant text, e.g., academic citation networks). These traits determine how LLMs can be leveraged.

Current state of LLM‑based graph learning includes:

For Pure Graphs, LLMs can answer queries about topology (connectivity, shortest paths) by converting graph information into textual prompts.

For Text‑Paired Graphs, multimodal contrastive learning combines graph encodings with textual embeddings to improve tasks such as molecular property prediction.

For Text‑Attributed Graphs, LLMs encode node text into embeddings that are fed to GNNs, achieving better performance than traditional GNNs alone.

Roles of LLMs in graph tasks are threefold:

Enhancer/Encoder : LLMs generate explanations or embeddings that augment GNN inputs (explanation‑based or embedding‑based).

Predictor : LLMs directly perform graph inference either by flattening graph structures into text (Flatten‑based) or by combining GNN‑derived embeddings with textual prompts (GNN‑based).

Aligner : LLMs and GNNs are jointly trained via contrastive learning or knowledge distillation to align textual and structural representations.

Potential research directions include investigating whether LLMs truly learn graph topology versus relying on textual context, discovering universal structural features beneficial across domains, and designing methods that enable LLMs to focus on complex topological patterns.

Q&A highlights emphasize that better textual descriptions can improve Pure Graph reasoning, unified graph learning reduces domain‑specific engineering effort, and practitioners should choose the LLM role that best fits their data and tasks.