EvoOntology

EvoOntology is a self-evolving ontology layer for data agents, introduced by Meiduo Chong, Shaolei Zhang, Ju Fan, and Xiaoyong Du in "EvoOntology: A Self-Evolving Ontology Layer for Data Agents" (arXiv:2609.15779, September 2026). It addresses what the authors call the agent-data gap. An agent working over heterogeneous tables, files, and databases can normally only see that data through generic tools, such as listing column names or file paths one call at a time. Hand-written semantic layers pasted into a prompt do not scale to many data sources or adapt as the agent's behavior changes.

EvoOntology packages the ontology itself as an MCP server with three layers: a schema layer, a content layer, and a tool layer. A data agent queries it directly at runtime instead of reading a static description. A dedicated builder agent constructs the initial ontology autonomously, and a self-evolution loop then continuously revises it through small, typed edits. Each candidate edit is attributed to a specific weakness in the ontology and is kept only if a paired evaluation on the same underlying model shows it actually improves results, rather than being applied unconditionally.

Across three data-agent benchmarks and four backbone language models, EvoOntology outperforms both raw data exploration and hand-built semantic-layer baselines. On the DDR-Bench benchmark, average accuracy across six backbones rises by 17.8 points, from a 4.8-point gain on Qwen3.5-Flash to a 26.7-point gain on GPT-5.5. Execution accuracy on the BIRD benchmark rises by 7.4 points. The authors report that edits to the tool layer, rather than the schema or content layers, account for 57% of the total gain from evolution.