OaK Builds Dynamic Ontologies for Reliable Agents
OaK dynamically constructs task-specific schemas, knowledge graphs, and typed reasoning functions from task requirements and training data. Its frozen ontology kernel improves evidence grounding and multi-step agent performance across TravelPlanner, CRMArenaPro, and ToolQA.
OaK’s strongest idea is treating the ontology as an executable interface, not merely a retrieval layer.
- –Compiles recurring multi-step reasoning into typed, callable functions
- –Uses OWL/HermiT validation and judge-guided repair to catch schema defects
- –Achieves a 55.9% final pass rate on TravelPlanner with DeepSeek-v4-flash, versus 15.3% for ReAct
- –Leads CRMArenaPro and ToolQA results under both tested model backbones
- –Construction adds meaningful LLM, validation, and graph-building overhead, so production value will depend on reliability and cost
DISCOVERED
1h ago
2026-08-26
PUBLISHED
1h ago
2026-08-26
RELEVANCE
AUTHOR
Discover AI