A self-building semantic layer for data agents, with no scores published

single source· 1 articles · confidence: low · first seen 2026-09-13 20:00 UTC

What this means for you

Nothing to act on yet. The code is on GitHub, but the paper reports no scores, names no benchmarks and gives no evaluation dates, so its central claim cannot be checked. The pattern is worth borrowing — MCP is a sensible place to put a semantic layer — but not yet worth adopting.

A preprint posted to arXiv on 13 September describes EvoOntology, a semantic layer — a structured description of what a dataset holds — placed between a data agent and the tables, files and databases it must query. It runs as an MCP server (the protocol for handing an agent tools to call), holding schema, content and tool layers. A builder agent constructs the layer; a loop then refines it, and each edit is accepted only after a paired evaluation conditioned on the backbone model (the LLM driving the agent). The authors report it beats existing semantic-layer approaches on three data-agent benchmarks across four backbones, without scores, benchmark names or evaluation dates.

Key facts

  • ·The EvoOntology preprint was posted to arXiv on 13 September 2026 (arXiv:2609.15779). source
  • ·The ontology is packaged as an MCP server with three layers: schema, content and tool. source
  • ·A builder agent constructs the ontology; a self-evolution loop refines it with edits accepted only after a paired, backbone-conditional evaluation. source
  • ·The authors report evaluations on three data-agent benchmarks with four LLM backbones, but the abstract gives no scores and names no benchmarks or evaluation dates. source
  • ·Code is released at github.com/ruc-datalab/EvoOntology. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire