HalluGuard: Demystifying Data-Driven and Reasoning-Driven Hallucinations in LLMs

Researchers introduce HalluGuard, a new framework that identifies and addresses both data-driven and reasoning-driven hallucinations in Large Language Models. The system achieved state-of-the-art performance across 10 benchmarks and 9 LLM backbones, offering a unified approach to improve AI reliability in critical domains like healthcare and law.

• →HalluGuard introduces the first unified theoretical framework to address both data-driven and reasoning-driven hallucinations in LLMs.
• →The system achieved state-of-the-art performance across 10 diverse benchmarks and 9 popular LLM backbones.
• →The framework formally decomposes hallucination risk into training-time mismatches and inference-time instabilities.
• →HalluGuard uses NTK-based scoring to leverage geometric representations for improved hallucination detection.
• →The research addresses critical reliability issues for LLMs in high-stakes domains like healthcare, law, and scientific discovery.