Hallucination Detection in LLMs: Fast and Memory-Efficient Finetuned Models

Hallucinations can broadly when implementing AI in high-risk settings, such be categorized into two types [2]: faithfulness hallucinations, as autonomous cars, medicine, or insurances. Large where the LLM deviates from provided Language Models (LLMs) have seen a surge in popularity instructions, and factual hallucinations, where there in recent years, but they are subject to hallucinations, is a disparity between the generated content and which may cause serious harm in high-risk verifiable facts. The risk arises when individuals settings. Despite their success, LLMs are expensive unaware of these limitations mistakenly treat such to train and run: they need a large amount outputs as ground-truth, leading to decisions based of computations and memory, preventing the use on erroneous information -- a concern particularly of ensembling methods in practice.