Research map / Graph learning / Graph self-supervised learning

UniGTE: Unified Graph–Text Encoding for Zero-Shot Generalization across Graph Tasks and Domains

NeurIPS 2025 · Graph learning › Graph self-supervised learning · approach: LLMs and graph foundation models · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes UniGTE, an instruction-tuned LLM encoder with learnable alignment tokens and structure-aware graph-text attention feeding a frozen LLM decoder that predicts task answers and reconstructs graph prompts, achieving zero-shot graph learning across node, link, graph-level tasks.

Problem

Zero-shot generalization to unseen graph tasks and domains without task-specific supervision; GNNs are tied to fixed label spaces and LLMs struggle to capture graph structure and are sensitive to node ordering.

Approach

An LLM-based encoder with learnable alignment tokens and a structure-aware graph-text attention (shared graph position index, shortest-path distance bias, natural-language edge-description bias, masking) produces compact task-aware graph representations; a frozen LLM decoder predicts the answer and reconstructs the graph prompt as auxiliary supervision.

Key results

• Zero-shot Pubmed node classification accuracy 0.870 vs 0.781 for TEA-GLM (second best) in the in-domain table.
• Zero-shot graph regression on Esol/Lipo/Freesolv MAE 2.54/1.03/9.18 vs GOFA 4.93/1.36/14.98.
• Inference time per sample reduced vs Vicuna-7B and GOFA on several datasets (e.g., Cora 0.6s vs 2.6s for Vicuna-7B).

Measured effects reported in the paper

Surprising findings

• LLaGA and GraphGPT often fail to outperform their base Vicuna-7B in zero-shot settings, which the authors attribute to lack of permutation invariance.
• GOFA, when instruction-tuned under the authors' pipeline, fails to match even an untuned LLM on many tasks.

Limitations

• Gains on link prediction are less pronounced than on node classification and graph regression.
• Zero-shot Cora result is weaker relative to other settings; Cora node classification accuracy only 0.215.

Experimental setup

Datasets: Arxiv Children Computer FB15K237 ChEMBL Pubmed Cora Photo Sports BACE HIV PCBA

Benchmarks: Pubmed Cora Photo Sports BACE HIV PCBA WikiCS Reddit Instagram Esol Lipo

Metrics: accuracy AUC MAE legality rate

Baselines: Vicuna-7B OFA GraphGPT LLaGA TEA-GLM GOFA

Builds on: GOFA LLaGA TEA-GLM GraphGPT RoPE

Closest papers by meaning

• A Unified Graph Language Model for Multi-Domain Multi-Task Graph Alignment Instruction Tuning (NeurIPS 2026 · similarity 0.91)
• Towards Graph Foundation Models: Training on Knowledge Graphs Enables Transferability to General Graphs (NeurIPS 2025 · similarity 0.90)
• SSTAG: Structure-Aware Self-Supervised Learning Method for Text-Attributed Graphs (NeurIPS 2025 · similarity 0.90)
• Talk like a Graph: Encoding Graphs for Large Language Models (ICLR 2024 · similarity 0.90)
• Generalization Principles for Inference over Text-Attributed Graphs with Large Language Models (ICML 2025 · similarity 0.90)
• Towards Graph Foundation Models: Learning Generalities Across Graphs via Task-Trees (ICML 2025 · similarity 0.90)