AIAny
Icon for item

EvoOntology: A Self-Evolving Ontology Layer for Data Agents

Provides a self-evolving ontology layer that enables LLM-based data agents to query and interact with heterogeneous data via an MCP server; it auto-builds and iteratively refines schema, content, and tool layers based on agent interactions.

Introduction

Data agents often falter because heterogeneous data lives outside the agent and is only reachable through generic tools; static semantic descriptions either don’t scale or fail to adapt to different agent behaviors. EvoOntology’s core insight is to turn the semantic layer into an interactive, self‑evolving service (an MCP server) that agents actively query and that adapts based on observed interactions.

Key Findings
  • Interactive ontology as a runtime service: the ontology is packaged as an MCP server with three components—schema (object types and reference rules), content (grounded domain knowledge and data mappings), and tool interfaces—so agents can call functions rather than passively ingesting context. This reduces brittle prompt-heavy exploration of raw sources.
  • Autonomous construction + continuous refinement: a builder agent issues probe queries to construct an initial ontology grounded in observed data; a self-evolution loop then uses attribution analysis to find weaknesses, proposes typed edits (schema/content/tool), and admits edits only after a paired, backbone-conditional evaluation on held-out validation cases.
  • Practical tool API: the tool layer exposes compact interfaces (e.g., fbrowse for top semantic matches and fresolve for retrieving records and linked objects) plus a session manifest to bootstrap agent usage.
  • Empirical result: across multiple data-agent benchmarks and several LLM backbones, the method consistently outperforms direct-exploration and static semantic-layer baselines, indicating better bridging of the agent–data gap.
Who it helps and tradeoffs

Great fit if you run LLM-driven data agents over many heterogeneous, evolving sources and need a maintainable semantic layer that adapts to agent behavior and task patterns. It is less appropriate for tiny static datasets or scenarios where building/validating an ontology adds unacceptable latency: the self-evolution loop requires interaction logs, validation cases, and evaluation overhead (paired validation and backbone-conditional checks). Expect additional system complexity (MCP server, tool integration) and compute/annotation cost for iterative validation, but gain more robust, queryable semantics and fewer brittle multi-step agent failures.

Information

  • Websitearxiv.org
  • AuthorsMeiduo Chong, Shaolei Zhang, Ju Fan, Xiaoyong Du
  • Published date2026/09/14

Categories

More Items

Turns implemented functionality in open-source codebases into executable RL training environments by automatically generating behavioral specifications, execution-grounded tests, and post-rollout filtering; yields 5,545 verifiable coding tasks across 23 languages to train coding agents.

Provides reproducible, multi‑platform environments and execution‑grounded verification for hybrid computer‑use agents that interleave GUI exploration and code/CLI. Key features: a five‑platform harness, RecreationBench (250 tasks), reference‑grounded programmatic and visual assertions for automatic scoring.

Transforms source code into verifiable, reusable agent skills by extracting atomic operations, workflows, and recurring patterns and validating them via source-body-blind reconstruction. Produces CodeSkillBank (1,006,822 accepted records from 19,769 GitHub repos) and yields ~11.7% average downstream improvement.