Understanding and Mitigating LLM Hallucinations

Large language models (LLMs) have shown impressive capabilities but are known for generating non-factual or nonsensical statements, also known as "hallucinations." Detecting hallucinations is challenging and an active area of research. One approach to detect hallucinations is SelfCheckGPT, a zero-resource black-box hallucination detection method that checks the consistency of multiple samples generated by LLMs for the same prompt. The final score can be computed using semantic similarity measures like BERTScore or natural language inference methods such as NLI. Another approach involves leveraging an LLM to perform the consistency check, assigning scores based on whether the sentences contradict each other or not. Experimental results show that these approaches have varying performance levels, with the LLM-Prompt method being the best-performing version closely followed by the NLI version.