Inventing nonexistent medical test results is a form of hallucination commonly associated with generative AI models. Reducing top-p sampling (A) restricts the model to choosing from the most probable next tokens, reducing randomness and preventing the generation of fabricated or implausible content. AAIA emphasizes configuration tuning (temperature, top-k, top-p) as critical controls for mitigating hallucinations.
Increasing context (B) helps provide more information but does not directly reduce hallucinations. Increasing temperature (C) would make hallucinations worse by adding randomness. Frequency penalties (D) discourage repetition, not hallucinations. The strongest mitigation is reducing sampling randomness via reduced top-p.
References:
ISACA, AAIA Exam Content Outline - Domain 2: AI Operations (generative AI safeguards, hallucination mitigation).