Hallucination, Monofacts, and Miscalibration: An Empirical Investigation

Researchers conducted the first empirical investigation of hallucination in large language models, revealing that strategic repetition of just 5% of training examples can reduce AI hallucinations by up to 40%. The study introduces 'selective upweighting' as a technique that maintains model accuracy while significantly reducing false information generation.

• →Selective upweighting technique reduces AI hallucinations by up to 40% while maintaining accuracy levels.
• →Strategic repetition of only 5% of training examples can significantly improve model reliability.
• →The research challenges universal deduplication policies commonly used in AI training.
• →Study establishes empirical relationship between monofact rates and hallucination frequency in language models.
• →Findings reveal inherent tension between accuracy optimization and hallucination reduction in AI training.