AI hallucination is the single biggest threat to trusting machines with truth. But DeepRoot just proved that the problem isn't the model—it's the method.
DeepRoot is a multi-agent LLM system that builds a verified knowledge graph to coordinate multiple AI agents. Instead of letting a single LLM guess its way through ancient medical texts, DeepRoot separates grounding (what's real) from reasoning (what follows). The result? Hallucination rates drop from 87% to just 7–10%.
Tested on the Shen Nong Ben Cao Jing—a foundational text of traditional Chinese medicine—DeepRoot recovered 10 out of 21 held-out compound-disease pairs. That's 47.6% accuracy. Raw LLMs? A pathetic 4.8%. The difference isn't marginal; it's a return from noise to signal.
Why should you care? Because ancient medical knowledge isn't just a curiosity—it's a pharmacy of potential treatments buried in language that modern LLMs love to fabricate. Every hallucinated herb or disease risks creating misinformation that poisons downstream research. DeepRoot builds a verified knowledge graph as it reads, turning the text into an auditable, shareable database. The AI isn't just talking; it's taking notes.
The paper was accepted to the inaugural ACM CAIS workshop on “AI Agents for Discovery in the Wild” and posted on OpenReview and arXiv in late May/June 2026. The authors remain anonymous for now—adding a layer of mystery to a system that could reshape how we mine historical knowledge.
Who loses? Every single-agent LLM that claims to summarize texts without grounding. Who wins? Traditional medicine, historical research, and anyone tired of AI confidently lying about the past.
Trends suggest DeepRoot could be the blueprint for a new class of agents that verify before they reason. If the authors release the code and dataset, expect a gold rush in historical medicine mining. If not, this might remain a tantalizing hint of what AI could achieve when it learns to separate facts from fantasies.