Hallucination scores for language models change little across prompting techniques such as Zero-shot, Few-shot, CoT, and Instruction formats because the prompts are semantically equivalent and decoding is low-entropy, causing outputs to be dominated by the models' learned alignment policies.

Authors

• Person: Not available Organization: Frontiers
  Survey and analysis of hallucinations in large language models

Sources

• Survey and analysis of hallucinations in large language models www.frontiersin.org Frontiers via serper

Referenced by nodes (2)

• Zero-Shot concept
• CoT concept