AAAI 2026 Oral

Jan 25, 2026·
Enjun Du
Enjun Du
· 2 min read
Oral Presentation at AAAI 2026
Abstract
Foundation models have demonstrated remarkable capabilities for general representation and reasoning in computer vision and natural language processing. However, developing analogous foundation models for knowledge graphs presents unique challenges due to their dynamic nature, with entities and relations requiring continuous updates. Furthermore, a significant gap exists in current models ability to perform cross-domain reasoning, particularly for critical applications in biomedical knowledge graphs. We present \textbf{GraphOracle}, a novel foundation model for universal knowledge graph reasoning that addresses these limitations. GraphOracle achieves domain-independent representations by modeling causal graphs of relationship interactions, enabling extension to arbitrary graph structures and rapid adaptation through fine-tuning. By integrating multiple unimodal foundation models, GraphOracle demonstrates robust cross-domain reasoning capabilities that extend seamlessly to biomedical knowledge graphs. Extensive experiments across Transductive, Inductive, and BioKG benchmarks demonstrate that GraphOracle consistently outperforms existing approaches and can surpass state-of-the-art models with minimal fine-tuning. Our work represents a significant advance in universal knowledge graph reasoning with important implications for biomedical discovery and cross-domain applications.
Date
Jan 25, 2026 10:00 AM — 10:30 AM
Event
Location

Singapore

Singapore EXPO, Singapore,

This page highlights my Oral Presentation at AAAI 2026 held at Singapore EXPO, Singapore on January 25, 2026.
Click on the Slides button above to view the presentation materials, or watch the talk video for the full oral presentation.

🧠 GraphOracle: Efficient Fully-Inductive Knowledge Graph Reasoning

Our AAAI 2026 oral presentation introduces GraphOracle, a foundation model designed for universal and fully-inductive knowledge graph reasoning, with strong cross-domain generalization and biomedical applicability.

🚀 Key Contributions

  • Foundation Model for Knowledge Graph Reasoning
    We propose GraphOracle, a novel foundation model that enables domain-independent and fully-inductive KG reasoning, addressing the limitations of static entity- and relation-dependent models.

  • Relation-Dependency Graph (RDG)
    GraphOracle models causal interaction structures among relations via Relation-Dependency Graphs, allowing reasoning to generalize to unseen entities, relations, and domains.

  • Cross-Domain & Biomedical Reasoning
    By integrating multiple unimodal foundation models, GraphOracle demonstrates strong cross-domain transferability, extending naturally to biomedical knowledge graphs (BioKGs).

  • Efficient Adaptation with Minimal Fine-Tuning
    Extensive experiments on Transductive, Inductive, and BioKG benchmarks show that GraphOracle consistently outperforms existing methods and can surpass state-of-the-art models with very limited fine-tuning.

📊 Experimental Results

GraphOracle achieves:

  • Superior reasoning accuracy across multiple KG benchmarks
  • Robust performance under inductive and cross-domain settings
  • Strong scalability and efficiency suitable for large, evolving knowledge graphs

🔗 Resources

For more information about the conference, please visit the AAAI 2026 website.

AAAI 2026 Oral Foundation Model Knowledge Graph Reasoning Inductive Reasoning BioKG
Enjun Du
Authors
Enjun Du
Research Assistant