AWS Certified Cloud Practitioner

Explanation:

Setting temperature to 0 and specifying strict output format is the most effective technique for reducing hallucinations in large language models because:

1. Temperature = 0: This setting makes the model deterministic and always selects the highest probability token, reducing randomness and variability that can lead to hallucinations.

2. Strict output format: By specifying exactly how the output should be structured (e.g., JSON format, specific templates, or constrained formats), you guide the model to produce more reliable and structured responses.

3. Other techniques comparison:

   • Self-consistency decoding (A): Generates multiple reasoning paths and takes the majority vote, which improves reasoning but doesn't specifically target hallucination reduction.
   • Few-shot prompting (B): Provides examples to guide the model, which helps with task understanding but doesn't directly control hallucination.
   • Chain-of-thought prompting (D): Encourages step-by-step reasoning, which improves reasoning quality but doesn't inherently prevent hallucinations.

4. Hallucination control: The combination of deterministic output (temperature=0) and format constraints provides the strongest guardrails against the model generating fabricated or incorrect information.