Autonomous Knowledge Graph Exploration with Adaptive Breadth-Depth Retrieval
Joaquín Polonuer, Marinka Zitnik, Lucas Vittor, Iñaki Arango, Ayush Noori, David A. Clifton, Luciano Del Corro
arXiv (Cornell University)
Problems Identified (5)
Breadth-depth retrieval tradeoff: Retrieving evidence from knowledge graphs for language model queries requires balancing broad graph coverage with multi-hop relational traversal.
Shallow similarity retrieval: Similarity-based knowledge graph retrievers can provide broad coverage but remain shallow.
Fragile seed-based traversal: Traversal-based retrieval methods depend on seed node selection that can fail for queries spanning multiple entities and relations.
Breadth-depth retrieval tradeoff: Retrieving evidence from knowledge graphs for language model queries requires balancing broad graph coverage with multi-hop relational traversal.
Shallow similarity retrieval: Similarity-based knowledge graph retrievers can provide broad coverage but remain shallow.
Proposed Solutions (5)
ARK adaptive KG retriever: ARK is a tool-using knowledge graph retriever that lets a language model adaptively control the breadth-depth retrieval tradeoff.
Lexical-search and neighborhood-exploration tools: ARK uses global lexical search over node descriptors and one-hop neighborhood exploration that can compose into multi-hop traversal.
Training-free adaptive traversal: ARK alternates between breadth-oriented discovery and depth-oriented expansion without relying on fragile seeds, preset hop depth, or retrieval training.
Label-free tool-use distillation: ARK's tool-use trajectories are distilled from a large teacher model into an 8B model via label-free imitation.
ARK adaptive KG retriever: ARK is a tool-using knowledge graph retriever that lets a language model adaptively control the breadth-depth retrieval tradeoff.
Results (3)
STaRK retrieval performance:
Training-free baseline improvement:
Distilled 8B improvement:
Research Domain
knowledge graph retrieval for language model question answering