Natural Language as a Universal Interface for Graph Foundation Models

This direction is split into other directions: natural language feature unification. Its last version stays readable and citable here.

Version 1 of 158 papers82 conceptsfrozen 2026-10-04

In the history of deep learning, fields like Natural Language Processing and Computer Vision have successfully transitioned from specialized, task-specific models to "foundation models"—single, massive architectures like GPT-4 or CLIP that generalize across diverse datasets. Graph Machine Learning, however, has long remained stuck in the era of "one-graph wonders."

A Graph Neural Network (GNN) trained to predict protein interactions cannot be easily applied to a social network or a citation graph. This is because graphs are fundamentally heterogeneous: their node attributes have different dimensions, their edge types have different meanings, and their underlying "geometries" vary wildly. Recent research suggests that the key to breaking this barrier is not a more complex graph algorithm, but a shift in how we represent graph data: using natural language as a universal interface.

The Feature Alignment Problem

The primary obstacle to a Graph Foundation Model (GFM) is feature misalignment. In a molecular graph, a node might be represented by a 100-dimensional vector of chemical properties; in a citation network, a node might be a 768-dimensional embedding of a paper abstract. Traditional GNNs require a fixed input dimension, making it mathematically impossible for a single model to process both datasets simultaneously.

As Haitao Mao and Yao Ma argue in Graph Foundation Models Are Already Here, the "Holy Grail" of the field is a universal graph vocabulary. Instead of treating node attributes as raw numbers, researchers are now treating them as natural language descriptions. By converting domain-specific attributes into text—for example, describing a molecule's properties or a user's interests in plain English—we can use a Large Language Model (LLM) as a universal encoder. This projects disparate data types into a shared semantic embedding space, effectively "leveling the playing field" for the GNN that follows.

Unifying Domains through Textualization

The core mechanism of this new paradigm is textualization. In the One for All (OFA) framework, Hao Liu and colleagues demonstrate that by describing graph data with natural language, a single set of weights can handle node, link, and graph-level classification across entirely different domains.

This works because language acts as a "semantic bridge." If two different datasets both contain information about "age" or "category," an LLM can recognize these shared concepts even if the original data formats were incompatible. This approach is echoed in UniGraph by Bryan Hooi and Xiaoxin He, which uses natural language to bridge structural and semantic gaps between molecular structures and social networks.

Topology-Aware Synthesis

A significant challenge arises when graphs lack textual metadata entirely. To solve this, Zehong Wang and Chuxu Zhang introduced TANS (Topology-Aware Node description Synthesis) in Can LLMs Convert Graphs to Text-Attributed Graphs?.

TANS uses an LLM to "read" the structural properties of a graph—such as its degree distribution or local clustering—and "write" a textual description for every node. By synthesizing these descriptions, even "text-free" graphs can be mapped into the same universal feature space as text-attributed graphs. This allows a model trained on a rich citation network to generalize its structural reasoning to a raw social network where no text was originally available.

Zero-Shot Reasoning and Transfer

The ultimate goal of this research is zero-shot transfer: the ability of a model to perform a task on a graph it has never seen before. Yuhan Li and Jia Li explore this in ZeroG, where they map node attributes and class labels into a shared semantic space. By converting classification into a similarity matching task between node embeddings and label embeddings, the model can generalize across datasets with entirely different feature schemes.

This logic extends beyond traditional graphs to tabular data. Models like MediTab by Zifeng Wang and Jimeng Sun use LLMs to translate heterogeneous medical tables into a shared natural language format. Similarly, LaTable by Rob Davis leverages LLM-encoded metadata to transfer statistical knowledge between datasets with different column headers.

Why It Works: Semantic Grounding

The reason this approach succeeds where previous attempts at "universal GNNs" failed is semantic grounding. Traditional GNNs are purely structural; they see nodes as anonymous points in space. By introducing natural language, we provide the model with a "prior" on what those nodes represent.

As Yang Li and Junier Oliva suggest in Towards Universal Neural Inference, semantic grounding allows models to recognize that conceptually related variables in different datasets—like "income" in one table and "salary" in another—represent the same underlying information. This linguistic anchor allows the model to focus on learning the relationships between concepts, which are often more transferable across domains than the raw data itself.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_natural_language_as_a_universal_interfac_92ecf11b,
  title = {Natural Language as a Universal Interface for Graph Foundation Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/natural-language-as-a-universal-interface-for-graph-foundation-models/rdr_92ecf11b463a46619758669c530881c4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_natural_language_as_a_universal_interfac_92ecf11b_v1,
  title = {Natural Language as a Universal Interface for Graph Foundation Models},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/natural-language-as-a-universal-interface-for-graph-foundation-models/rdr_92ecf11b463a46619758669c530881c4/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}