Hallucination

Hallucination is a well-known failure mode of large language models where they produce content that is not grounded in their training data or provided context. The model “hallucinates” by generating information that sounds coherent and confident but is incorrect.

Causes of hallucinations:

• The model prioritizing plausible-sounding text over factual accuracy
• Insufficient or contradictory training data
• The model’s inherent lack of grounding in real-world truth
• Prompt ambiguity leading to speculative completions

Techniques to reduce hallucinations include RAG (retrieving factual context), prompt engineering (explicitly asking for citations), fine-tuning on factual datasets, and using confidence calibration methods. Despite progress, hallucination remains an open challenge in LLM deployment.