RUC open-sources EvoOntology: self-evolving ontology layer lifts data agents by up to 26.7 points
udmrzn · x · 2026-10-17
Renmin University of China open-sourced EvoOntology, a self-evolving ontology layer for data agents that topped Hugging Face Daily Papers (855 GitHub stars).
The problem: agents querying enterprise data guess schema mappings and metric definitions, often producing plausible-but-wrong numbers. The paper shows stuffing a data dictionary into the prompt actually makes Claude-Sonnet-5 score 15 points lower.
How it works:
- Build an ontology from the questions you ask — metric definitions, table relations, business constraints — each verified against real data before admission
- Serve concepts on demand via MCP instead of stuffing context
- After each task, analyze where the agent erred and patch only that part; a new version ships only if it beats the old one on the same model and data
Results: same Claude-Sonnet-5 gains 8.8 points (nearly 24-point gap vs prompt-stuffing); GPT-5.5 gains 26.7 on DDR-Bench. Zero model weight changes. Ready-made plugins for Claude Code and Codex.
More from coding & agent
- 'Tip-of-the-tongue' queries: a blind spot for modern agentic retrieval — bclavie · 2026-10-17
- vLLM gets autonomously optimized by AI agents with clear per-step performance gains, a first — vllm_project · 2026-10-17
- Repo just 2.5 days old already has 8,639 closed PRs merged minutes apart — ravisparikh · 2026-10-17
- Context rot: why Every's head of evals starts a fresh Claude thread for every task — danshipper · 2026-10-17
- "Check the environment before the prompt" — until the prompt actually was the bug — ClickOk5811 · 2026-10-17
- LEASH: weak-to-strong escalation matches Claude Opus 5 at 19.2% lower cost on OSWorld — xwang_lk · 2026-10-17