Data Party THU

Sep 14, 2025 · Artificial Intelligence

Why Do Large Language Models Hallucinate? Uncovering the Root Causes and Practical Fixes

The article analyzes why large language models frequently generate confidently wrong answers, attributing hallucinations to statistical inevitability, data scarcity, and limited model expressiveness, and shows how RLHF exacerbates the problem by rewarding guesses, then proposes confidence‑threshold and "I don't know" strategies to mitigate it.