Research map / Graph learning / Knowledge and temporal graphs
A Prompt-Based Knowledge Graph Foundation Model for Universal In-Context Reasoning
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 KG-ICL, a prompt-based knowledge graph foundation model that performs in-context reasoning over unseen KGs using query-related prompt graphs, a unified tokenizer, and two message passing networks, pre-trained on three KGs and evaluated on 43 KGs.
Problem
Existing KG reasoning models are trained separately per KG and cannot generalize or transfer to unseen entities, relations, and KGs across transductive, inductive, and fully-inductive settings.
Approach
Build a query-specific prompt graph around an example fact (neighbors of subject/object plus k-hop paths), map its entities and relations to predefined tokens via a unified tokenizer (distance-based entity labels, query-identity relation labels), encode it with a prompt MPNN, and use the resulting relation representations to initialize a conditional MPNN for KG reasoning without parameter updates.
Key results
- KG-ICL pre-train reaches average MRR 0.442 and H@10 0.606 across 43 KGs, versus ULTRA pre-train at 0.396 and 0.557.
- KG-ICL finetune achieves the best average MRR of 0.473 across all groups, versus supervised SOTA at 0.351.
- Ablation: removing the prompt graph drops average MRR from 0.442 to 0.132.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| finetune KG-ICL pre-train for 5 epochs | KG-ICL · Average (43 KGs) | MRR: 0.442 -> 0.473 | Helped |
| add prompt graph generation and encoding (ablation) | KG-ICL · Average (43 KGs) | MRR: 0.132 -> 0.442 | Helped |
| add the unified tokenizer (ablation) | KG-ICL · Average (43 KGs) | MRR: 0.403 -> 0.442 | Helped |
| use the proposed token representation instead of GraIL one-hot labeling (ablation) | KG-ICL · Average (43 KGs) | MRR: 0.431 -> 0.442 | Helped |
| combine neighbors with 3-hop paths instead of neighbors only (ablation) | KG-ICL prompt graph · Average (43 KGs) | MRR: 0.425 -> 0.442 | Helped |
| add neighbors to 1-hop paths (ablation) | KG-ICL prompt graph · Average (43 KGs) | MRR: 0.153 -> 0.442 | Helped |
| add neighbors to 2-hop paths (ablation) | KG-ICL prompt graph · Average (43 KGs) | MRR: 0.171 -> 0.442 | Helped |
| add neighbors to 3-hop paths (ablation) | KG-ICL prompt graph · Average (43 KGs) | MRR: 0.410 -> 0.442 | Helped |
| replace KG reasoning message passing with NBFNet (ablation) | KG-ICL · Average (43 KGs) | MRR: 0.442 -> 0.416 | Hurt |
| increase the number of prompt examples from 1-shot to 20-shot (ablation) | KG-ICL · 43 KGs | MRR | No clear effect |
| use the same pre-training mixture as ULTRA (ablation) | KG-ICL · Inductive (14 KGs) | MRR: 0.554 -> 0.547 | No clear effect |
| use the same pre-training mixture as ULTRA (ablation) | KG-ICL · Transductive (16 KGs) | MRR: 0.346 -> 0.357 | No clear effect |
Surprising findings
- Removing the prompt graph causes a very large performance drop (average MRR 0.132), and the pre-trained model is trained on only three small KGs yet beats ULTRA pre-trained with a larger mixture.
- KG-ICL pre-train underperforms the supervised SOTA on transductive datasets (0.346 vs 0.365 MRR).
Limitations
- Transductive performance of the pre-trained model lags the supervised SOTA (0.346 vs 0.365 MRR).
- Fully-inductive and inductive evaluation relies on subsets of datasets; some results use the 1-vs-50 protocol.
- Preprocessing cost grows for large KGs with many relations (AristoV4 took 995 seconds).
Experimental setup
Datasets: FB V1 NELL V1 CoDEx-small FB15k-237 WN18RR NELL-995 YAGO3-10 CoDEx AristoV4 DBpedia100k ConceptNet100k Hetionet
Benchmarks: 43 KG datasets across inductive, fully-inductive, and transductive settings FB15k-237 WN18RR NELL-995 YAGO3-10
Metrics: MRR Hits@10
Baselines: Supervised SOTA ULTRA InGram RED-GNN NBFNet
Builds on: ULTRA GraIL NBFNet RED-GNN PRODIGY
Closest papers by meaning
- LightPROF: A Lightweight Reasoning Framework for Large Language Model on Knowledge Graph (AAAI 2025 · similarity 0.89)
- Graph-constrained Reasoning: Faithful Reasoning on Knowledge Graphs with Large Language Models (ICML 2025 · similarity 0.88)
- Towards Graph Foundation Models: Training on Knowledge Graphs Enables Transferability to General Graphs (NeurIPS 2025 · similarity 0.88)
- Towards Foundation Models for Knowledge Graph Reasoning (ICLR 2024 · similarity 0.87)
- Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning (ICLR 2024 · similarity 0.87)
- Think-on-Graph: Deep and Responsible Reasoning of Large Language Model on Knowledge Graph (ICLR 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.