KAIST group pushes knowledge-intelligence separation to fix LLM trustworthiness

coallaoh · x · 2026-08-18

The author argues that the innate capabilities and fundamental limitations of parametric models need far more research, along with better human-in-the-loop interfaces — noting that the human/AI division of labor should shift by domain, since a coding agent and a clinical assistant have very different costs of a wrong answer.

Their Scalable Trustworthy AI group (moved to KAIST AI in Feb 2026) proposes "Knowledge-Intelligence Separation": analogous to the code-data separation of the 1960s, store knowledge in interpretable, editable external modules while keeping reasoning in the model. This enables faster customization, training data attribution by design, and knowledge editing/unlearning — addressing LLM hallucination, continual-learning failures, impractical GDPR compliance, and parameter-level privacy leaks. Research spans RAG and vector databases, memory-augmented architectures, multimodality, human-AI interaction, expert-in-the-loop systems, agentic AI, and privacy/security.

Related event: Human-AI division of labor should vary by domain, researchers argue(2 posts)→

Original post →

More from Research

Research channel →