Ran Afek (Tel Aviv U) has posted “Accuracy and Fluency: Understanding Hallucinations in Large Language Models” on SSRN. Here is the abstract:
Large Language Models (LLMs) have demonstrated remarkable capabilities in generating fluent, contextually appropriate text, yet they are inherently prone to hallucinations, producing outputs that appear plausible but are factually inaccurate. Drawing on the findings of Kalai et al. (2025), this paper examines how hallucinations arise, not merely from corrupted training data but from the intrinsic architecture and probabilistic nature of LLMs. This study highlights how statistical patterns learned from massive datasets underpin both the strengths and limitations of language generation models. Ultimately, LLMs favor conversational fluency over strict factual precision, reflecting the probabilistic and context-driven nature of human language itself.