Knowledge Graphs Boost LLM Accuracy from 65% to 99% in Industrial Operations
· Madhulatha Mandarapu, Sandeep Kunkunuru
A new paper from the KDD 2026 team flips a big assumption about AI agents on its head.…
A new paper from the KDD 2026 team flips a big assumption about AI agents on its head. Instead of asking whether better orchestration or a smarter LLM boosts performance on industrial tasks, they found the real bottleneck is the data layer itself.
The researchers used AssetOpsBench, a benchmark of 139 industrial maintenance scenarios where GPT-4 agents hit 65% accuracy using flat document stores like CouchDB, YAML, and CSV. That became their baseline. Then they built a knowledge graph with 781 nodes, 955 edges, and 16 relationship types over the same data. The results were striking.
Deterministic graph handlers without any LLM hit 99% (137 out of 139). LLMs that generated Cypher queries against the graph landed at 82 to 83%. That’s a huge jump from the 65% baseline, using the same GPT-4 model. The key insight: instead of having the LLM reason directly over raw text and tables, they had it produce structured queries. The graph handled the execution deterministically.
They also added 40 graph native scenarios, like multi-hop dependency tracking and PageRank criticality, and tested on the expanded AssetOpsBench of 467 scenarios across six domains. The deterministic handlers scored a perfect 467 out of 467, with an average score of 0.848.
The takeaway is simple. For structured operational work, the data model matters more than the LLM or the orchestration pattern. Knowledge graphs act as a clean integration layer between messy industrial data and LLM reasoning. The next question is how far this inverted approach can scale.