By name: kgpf-knowledge-graph-foundation-model description: > Knowledge Graph Foundation Model using Prior-data Fitted Network (PFN) for in-context learning. Unifies transferable relational regularities with inference-time context from structured neighborhoods. Use when: building KG reasoning systems, cross-graph transfer learning, in-context KG completion, multi-graph pretraining, or zero/few-shot adaptation to unseen knowledge graphs. Triggers: knowledge graph foundation model, in-context KG reasoning, PFN for graphs, cross-graph transfer, NBFNet neighborhood encoding.
KGPFN: Knowledge Graph Foundation Model via In-Context Learning
Overview
KGPFN (arXiv:2605.14907) combines two pillars of foundation models for knowledge graphs:
- Transferable relational regularities - learned via multi-graph pretraining
- In-context learning at inference time - conditioning on structured context
Architecture
Three-Level Context Encoding
| Level | Component | Purpose |
|---|---|---|
| Relation | Message passing on relation graphs | Cross-graph relational invariances |
| Local | Multi-layer NBFNet | Entity neighborhood structure |
| Global | PFN with feature + sample attention | Relation-specific global context from retrieved instances |
Key Design Decisions
Local Context (NBFNet):
- Encode query entity neighborhoods via multi-layer NBFNet
- Captures local graph topology around target entities
- Multiple layers = larger receptive field
Global Context (PFN):
- Retrieve large set of instances of the query relation + their local neighborhoods
- Aggregate within Prior-data Fitted Network framework
- Combines feature-level and sample-level attention
- Learns when to instantiate reusable patterns vs override with contextual evidence
Multi-Graph Pretraining:
- Train on diverse KGs simultaneously
- Model learns transferable relational regularities
- At inference: adapts to unseen graphs via ICL alone
Implementation Pattern
# 1. Relation Graph Message Passing
# Build relation graphs where nodes = relation types
# edges = co-occurrence patterns across entities
def encode_relations(kg):
rel_graph = build_relation_cooccurrence(kg)
rel_embeddings = message_passing(rel_graph)
return rel_embeddings
# 2. Local Neighborhood Encoding (NBFNet)
def encode_local_context(kg, query_entity, num_hops=3):
subgraph = extract_k_hop_neighborhood(kg, query_entity, k=num_hops)
return nbfnet_encode(subgraph, query_entity)
# 3. Global Context Construction (PFN)
def build_global_context(kg, query_relation, k=100):
instances = retrieve_relation_instances(kg, query_relation, top_k=k)
local_contexts = [encode_local_context(kg, inst.entity) for inst in instances]
return pfn_aggregate(local_contexts) # feature + sample attention
# 4. Inference: Combine all levels
def predict(kg, query):
rel_emb = encode_relations(kg)
local_ctx = encode_local_context(kg, query.entity)
global_ctx = build_global_context(kg, query.relation)
return combine_and_predict(rel_emb, local_ctx, global_ctx)
When to Use
- Zero-shot KG completion on unseen graphs - ICL alone achieves strong results
- Multi-domain KG reasoning - pretrain once, adapt via context
- Few-shot relation learning - retrieve similar instances as context
- Cross-graph transfer - no fine-tuning needed, just context retrieval
Benchmarks
Tested on 57 KG benchmarks, consistently outperforming competitive fine-tuned KG foundation models through in-context learning alone.