MMD-Flagger: Leveraging Maximum Mean Discrepancy to Detect Hallucinations
The episode analyzes the 'MMD-Flagger' paper, which introduces Maximum Mean Discrepancy (MMD) to detect AI hallucinations. Hosts explain how MMD measures statistical divergence by comparing outputs generated at varying temperatures. This process reveals a U-shaped trajectory that signals when an AI's probabilistic behavior fails to align with factual data.