Research map / Graph learning / Knowledge and temporal graphs
A Foundation Model for Zero-shot Logical Query 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 UltraQuery, a zero-shot foundation model for complex logical query answering on arbitrary knowledge graphs, built from an inductive link predictor (ULTRA) plus non-parametric fuzzy logic operators, trained once on FB15k237 queries.
Problem
Complex logical query answering over knowledge graphs has been limited to transductive or entity-only inductive settings; existing models learn graph-specific entity/relation vocabularies and cannot answer queries on new KGs with new entities and relations without retraining.
Approach
Derive both relation projections and logical operators as vocabulary-independent functions: the projection is an inductive GNN (initialized from pre-trained ULTRA) operating on relation meta-graphs, and logical operators are non-parametric product-logic fuzzy operations, enabling zero-shot query answering on any KG.
Key results
- Zero-shot UltraQuery averages 0.309 MRR on EPFO queries vs 0.196 for the best baseline across 23 datasets, and 0.178 vs 0.105 on negation queries.
- On the 11 new inductive (e,r) WikiTopics-CLQA datasets, UltraQuery reaches 0.280 EPFO MRR vs 0.014 for the edge-type heuristic baseline.
- UltraQuery (177k parameters) lags the 30M-parameter QTO on transductive benchmarks but beats it on FB15k in EPFO and negation queries.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune the Ultra projection operator on complex queries rather than use the edge-type heuristic baseline | UltraQuery · Inductive (e,r), EPFO avg across 11 datasets | MRR: 0.014 -> 0.280 | Helped |
| fine-tune the Ultra projection operator on complex queries rather than use the edge-type heuristic baseline | UltraQuery · Inductive (e,r), negation avg across 11 datasets | MRR: 0.004 -> 0.193 | Helped |
| fine-tune on complex queries instead of using QTO | UltraQuery · Transductive, EPFO avg across 3 datasets | MRR: 0.468 -> 0.411 | Hurt |
| fine-tune on complex queries instead of using QTO | UltraQuery · Transductive, negation avg across 3 datasets | MRR: 0.259 -> 0.240 | Hurt |
| fine-tune on complex queries instead of using GNN-QE | UltraQuery · Inductive (e), EPFO avg across 9 datasets | MRR: 0.328 -> 0.312 | Hurt |
| fine-tune on complex queries instead of using GNN-QE | UltraQuery · Inductive (e), negation avg across 9 datasets | MRR: 0.176 -> 0.139 | Hurt |
| fine-tune the projection operator on complex queries rather than freeze the pretrained operator with thresholding (ablation) | UltraQuery · 23 datasets, EPFO avg | MRR: 0.279 -> 0.309 | Helped |
| fine-tune the projection operator on complex queries rather than freeze the pretrained operator with thresholding (ablation) | UltraQuery · 23 datasets, negation avg | MRR: 0.107 -> 0.178 | Helped |
| add intermediate-score thresholding to frozen UltraQuery LP (ablation) | UltraQuery LP · 23 datasets, EPFO avg | MRR: 0.242 -> 0.279 | Helped |
| add intermediate-score thresholding to frozen UltraQuery LP (ablation) | UltraQuery LP · 23 datasets, negation avg | MRR: 0.088 -> 0.107 | Helped |
| add NELL995 query data to FB15k237 training (ablation) | UltraQuery 2G · 11 inductive (e,r) datasets, EPFO avg | MRR: 0.280 -> 0.310 | Helped |
| add NELL995 query data to FB15k237 training (ablation) | UltraQuery 2G · 20 inductive datasets, negation avg | MRR: 0.166 -> 0.158 | Hurt |
Surprising findings
- The frozen ULTRA link predictor with score thresholding (UltraQuery LP) reaches higher 1p performance than the fine-tuned model but performs much worse on negation and multi-hop queries, showing the multi-source propagation issue.
- Training on additional datasets (2G) improves EPFO on inductive (e,r) datasets but slightly reduces negation performance.
Limitations
- Projection operator is the only learnable component; multi-source propagation remains a challenge for pre-trained link predictors.
- Lags behind the larger trained QTO on transductive benchmarks and ROC AUC for easy vs hard answer ranking is behind GNN-QE.
- Only graph structure is used without node or edge features; scalability to very large KGs is untested.
Experimental setup
Datasets: FB15k237 NELL995 FB15k BetaE query datasets WikiTopics
Benchmarks: FB15k237 BetaE queries NELL995 BetaE queries FB15k BetaE queries GNN-QE inductive (e) datasets (FB 106%-550%) WikiTopics-CLQA (11 topics)
Metrics: MRR Hits@10 ROC AUC MAPE
Baselines: QTO GNN-QE edge-type heuristic BetaE Query2Box CQD FuzzQE NodePiece-QE
Builds on: ULTRA GNN-QE BetaE NBFNet CQD
Closest papers by meaning
- NQE: N-ary Query Embedding for Complex Query Answering over Hyper-Relational Knowledge Graphs (AAAI 2023 · similarity 0.90)
- Is Complex Query Answering Really Complex? (ICML 2025 · similarity 0.90)
- Context-aware Inductive Knowledge Graph Completion with Latent Type Constraints and Subgraph Reasoning (AAAI 2025 · similarity 0.87)
- Consistent Query Answering over Existential Rules with Open and Closed Predicates (AAAI 2025 · similarity 0.87)
- Improving Complex Reasoning over Knowledge Graph with Logic-Aware Curriculum Tuning (AAAI 2025 · similarity 0.86)
- Rethinking Complex Queries on Knowledge Graphs with Neural Link Predictors (ICLR 2024 · similarity 0.86)
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.