MMD-Flagger: Leveraging Maximum Mean Discrepancy to Detect Hallucinations
Researchers have proposed a method to detect hallucinations in large language models. Hallucinations are factual errors generated by these models when they lack ground-truth labels. MMD-Flagger uses the Maximum Mean Discrepancy (MMD) metric to track the stability of model outputs across different decoding temperatures. It identifies hallucinations based on characteristic shapes in the MMD trajectory. The method was tested on multi-lingual claim verification benchmarks using m
Researchers have proposed a method to detect hallucinations in large language models. Hallucinations are factual errors generated by these models when they lack ground-truth labels. MMD-Flagger uses the Maximum Mean Discrepancy (MMD) metric to track the stability of model outputs across different decoding temperatures. It identifies hallucinations based on characteristic shapes in the MMD trajectory. The method was tested on multi-lingual claim verification benchmarks using modern language models like Llama-3 and Gemma-3.
---
Why it matters: This matters because large language models are increasingly used in autonomous agents, but their tendency to generate factual errors is a significant safety concern. Detecting hallucinations at test-time can help build trustworthy AI systems.
Source: https://arxiv.org/abs/2506.01367
This article was originally published at: https://arxiv.org/abs/2506.01367