Surprisingly Simple and Effective Multi-Domain Graph Foundation Model Through Graph-to-Table Alignment

Jul 16, 2026

Upload by Aurelia

Graph Foundation Models (GFMs) aim to learn transferable representations across diverse graph domains, enabling adaptation to varied downstream tasks. Existing approaches largely split between GNN- and LLM-based paradigms, often constrained by limited pre-training scale or heavy dependence on text attributes. Tabular foundation models (TFMs) are a promising text-free alternative, yet their ability to capture graph structure is underexplored. GTAlign proposes a graph-to-table alignment mechanism using a unified latent graph encoder and community-guided continual pre-training with few-shot episodes, followed by in-context inference.

LanguageEnglish

CategoryTechnology

FormatPDF

Pages18

Reading Time45 min

Abstract

Introduction

Background on Graph Foundation Models

Limitations of Existing Paradigms

Motivation for Text-Free Tabular Approaches

Proposed Method: GTAlign

Unified Latent Graph Encoder

Community-Guided Continual Pre-Training

In-Context Inference

arXiv :2607 . 1 1374v 1 [ cs .LG] 13 Jul 2026 SURPRISINGLY SIMPLE AND EFFECTIVE MULTIDOMAIN GRAPH FOUNDATION MODEL THROUGH GRAPH-TO-TABLE ALIGNMENT Chunyu Hu 1 , Tianyin Liao 1 , Ge Lan 1 , Xingxuan Zhang3 , Jianxin Li2 , Peng Cui3 , Ziwei Zhang2 ∗ 1Nankai University, 2Beihang University, 3Tsinghua University [huchunyu@mail.nankai.edu.cn](huchunyu@mail.nankai.edu.cn) , [1120230329@mail.nankai.edu.cn](1120230329@mail.nankai.edu.cn) [lange@nankai.edu.cn](lange@nankai.edu.cn) , [xingxuanzhang@hotmail.com](xingxuanzhang@hotmail.com) , [lijx@buaa.edu.cn](lijx@buaa.edu.cn) [cuip@tsinghua.edu.cn](cuip@tsinghua.edu.cn) , [zwzhang@buaa.edu.cn](zwzhang@buaa.edu.cn) ABSTRACT Graph Foundation Models (GFMs) have emerged as a promising paradigm for learning transferable representations across diverse graph domains. Recent ad vancements in GFMs have been largely dominated by two paradigms: Graph Neu ral Network and Large Language Model (LLM) based methods. However, these methods often face a fundamental dilemma between training with limited data and a heavy reliance on textual attributes. Tabular foundation models (TFMs) offer a potential alternative, as node features and representations can be natu rally organized in a tabular form. However, how to enable TFMs to effectively capture structural information of graphs remains largely unexplored. The key challenge is to learn a graph-to-table alignment mechanism that enables graph structural understanding for TFMs. To address this, we propose GTAlign, a sur prisingly simple yet effective Graph-to-Table Alignment framework for text-free Graph Foundation Model. Specifically, we first pretrain a graph encoder that maps diverse graphs into a unified latent space to capture domain-agnostic graph rep resentations. To further bridge the gap between graph topology and the tabu lar representation space, we propose community-guided continual pre-training, where pseudo-labels derived from graph community are used to construct few shot prediction episodes. Lastly, we adapt the graph encoder for an unseen target domain and perform in-context inference. Extensive experiments on five bench mark datasets demonstrate that GTAlign significantly outperforms state-of-the-art baselines on both node and graph classification, offering a simple, effective, and text-free GFM model. Code will be released upon acceptance. 1 INTRODUCTION Graph Neural Networks (GNNs) have been the cornerstone of graph representation learning Kipf & Welling (2017); Velikovi et al. (2018); Xu et al. (2019), boasting a wide range of applications. However, traditional GNNs are typically task-specific and domain-dependent, requiring training from scratch on specific domains and failing to generalize across unseen domains Hu et al. (2020) . Inspired by the success of Large Language Models (LLMs) as foundation models in natural language processing Zhou et al. (2025), recent research attention has been shifting towards Graph Foundation Models (GFMs) Mao et al. (2024); Zi et al. (2024); Yu et al. (2025a), which aim to learn generalpurpose graph representations through pre-training on multi-domain graph datasets, which can be adapted for diverse downstream tasks Wang et al. (2025) . Existing GFMs can be categorized into three types based on their backbone architectures Liu et al. (2025): GNN-based Liu et al. (2023), LLM-based Wang et al. (2024a), and GNN+LLM-based models Liu et al. (2024c) . However, despite the initial success, these architectures still face notable limitations. Particularly, GNN-based methods are often constrained by the scale and diversity of available graph pre-training corpora, which limits their ability to generalize across different domains and limits their performance in downstream ∗ Corresponding author. tasks Liu et al. (2025) . Meanwhile, LLM-based and GNN+LLM-based models often rely heavily on textual attributes, which restricts the model’s usability and effec

FAQs

What problem does GTAlign address in graph foundation modeling?

GTAlign targets the difficulty of enabling tabular foundation models to effectively capture graph structural information, especially in multi-domain settings. It focuses on learning a graph-to-table alignment mechanism for text-free GFMs.

How does GTAlign connect graph topology with the tabular foundation model input space?

GTAlign first pretrains a graph encoder that maps diverse graphs into a unified latent space to produce topology-aware representations. Those unified representations are treated as tabular tokens and fed into the TFM.

What role does community-guided continual pre-training play?

GTAlign introduces community-guided continual pre-training using pseudo-labels derived from graph community structure. These pseudo-labels construct few-shot prediction episodes to bridge graph topology and the tabular representation space.

Jul 16, 2026

Upload by Aurelia