Ontology-based checks catch LLM agent errors barely better than random, arXiv study finds
JeremyCMorgan · x · 2026-11-06
An arXiv paper (2610.37809) makes the checker itself the object of study, testing whether knowledge-graph ontologies (curated, LLM-generated, or data-derived) reliably validate LLM agent answers.
Key findings:
- Curated schemas flag errors on real agent answers barely more precisely than random checks, and published ontologies can contradict their own underlying data
- On UniProt and a private graph with 10^9 triples, the curated ontology wrongly flags more correct answers than data-derived or LLM-derived alternatives, pushing agents to "fix" answers that were right
- Each ontology source trades caught errors against false alarms; combining sources or feeding checker feedback back to the agent doesn't clearly pay off
Takeaway: structural schema validation is not factual verification, and wrapping agent outputs in ontology-derived checks deserves reevaluation.
More from coding & agent
- PSA: AI agents love slapping broad :has() rules on CSS, causing huge style recalcs — jh3yy · 2026-11-07
- YC-backed Trepo debuts kitchen hardware that lets AI agents shop for groceries in the background — ycombinator · 2026-11-07
- Cognition ships Devin Review 2.0: 5× cheaper, 1.5× faster PR review — smtabatabaie · 2026-11-07
- Air moves its agentic development experience directly into JetBrains IDEs, free tiers included — therealdanvega · 2026-11-07
- PSA: keep CSS :has() rules tight — agents' broad selectors trigger huge style recalcs — jh3yy · 2026-11-07
- Solo founder credits AI chief-of-staff bots for record-breaking revenue month — joshgonsalves_ · 2026-11-07