Detection and mitigation of hallucination in large language models

As more sophisticated large language models (LLMs) and artificial intelligence (AI) solutions continue to emerge, an eminent challenge called Hallucination; exist with these models. Hallucination describes the situation where LLMs and AI models generate very coherent but factually incorrect output or response to a query. The presence of hallucination in LLMs discourages their use in many real-world scenarios in the health, legal, and other related sectors, to avoid the implications of hallucinated outputs. Consequently, NLP researchers have continued to seek sophisticated methods, beyond simple fact-checking, to detect and mitigate LLM hallucinations. Diverse methods including retrieval-augmented generation (RAG), advanced decoding, and prompting techniques have been proposed, but no single solution seem to sufficiently eliminate the challenge compared to validation by human experts. In this comparison therefore, some recent solutions for the detection and mitigation of hallucination in different downstream NLP tasks are highlighted.