GILT: A New Graph Foundational Model Without LLMs

Researchers have introduced GILT (Graph In-context Learning Transformer), a graph foundational model that operates without Large Language Models (LLMs) and requires no per-graph tuning. Graph Neural Networks (GNNs) often fail to generalize to unseen graphs due to extreme heterogeneity in feature spaces, label sets, and topologies. Existing approaches either rely on LLMs, which struggle with numerical features, or require costly fine-tuning for each new graph. GILT addresses these limitations with an LLM-free, tuning-free architecture designed for in-context learning on graph data. The work is detailed in arXiv paper 2510.04567.