EvoOntology: A Self-Evolving Ontology Layer for Data Agents
The paper proposes an MCP-hosted ontology that data agents can query and revise while working.
EvoOntology is built around schema, content, and tool layers meant to sit between an agent and heterogeneous data sources. A builder agent constructs the ontology, then a self-evolution loop refines it with typed edits. Those edits are accepted only after paired evaluation conditioned on the LLM backbone. The authors report gains over strong baselines and semantic-layer methods across three data-agent benchmarks and four LLM backbones. HF Daily Papers' note
EvoOntology is built around schema, content, and tool layers meant to sit between an agent and heterogeneous data sources. A builder agent constructs the ontology, then a self-evolution loop refines it with typed edits. Those edits are accepted only after paired evaluation conditioned on the LLM backbone. The authors report gains over strong baselines and semantic-layer methods across three data-agent benchmarks and four LLM backbones. HF Daily Papers' note
score 4