- https://arxiv.org/pdf/2308.07134.pdf
- Source: arxiv
TL;DR
InstructGLM serializes graph topology and node features as natural language instructions, unifying node classification, link prediction, and graph classification under a single LLM and matching or exceeding GNN baselines in zero-shot and few-shot settings.
Summary
Graph machine learning tasks are typically solved with graph neural networks requiring task-specific architectures. InstructGLM serializes graph topology and node features as natural language instructions, reformulating all graph tasks as text generation problems solvable by a single LLM. On Cora, CiteSeer, and ogbn-arxiv benchmarks, InstructGLM matches or exceeds GNN baselines in zero-shot and few-shot regimes, demonstrating that language models can serve as general-purpose graph reasoners without graph-specific inductive bias.
Key contributions
- Proposes InstructGLM, which serializes graph structure and node features as natural language for LLM-based graph task solving.
- Unifies node classification, link prediction, and graph classification under a single text-generation framework without task-specific architectures.
- Demonstrates zero-shot and few-shot performance that matches or exceeds GNN baselines on standard benchmarks including ogbn-arxiv.
- Shows that instruction tuning enables zero-shot transfer across graph task types without retraining.
When to cite
- When arguing that LLMs can replace graph-specific architectures when topology is serialized as natural language.
- When discussing the use of foundation models to reduce specialized graph ML engineering across downstream tasks.
- When citing evidence of zero-shot transfer from instruction-tuned LLMs to graph reasoning tasks.
- When exploring the limits of context-length as the primary bottleneck for LLM-based graph processing at scale.