AIF-C01 Practice Question: A company uses an LLM to summarize medical…

Answer analysis

Option-by-option breakdown

For each option: why learners choose it and why it is or isn't the right answer here.

• ✗

  Increase temperature and top-p sampling parameters

  Why it's wrong here

  Raising temperature and top-p increases sampling randomness, producing more diverse and less grounded output, which worsens hallucination. It is tempting because these parameters tune creativity, and they would be correct when generating varied creative text rather than factual medical summaries.

• ✗

  Use a smaller model with less capacity

  Why it's wrong here

  A smaller model with less capacity has weaker factual recall and reasoning, typically increasing hallucination rates on complex medical content. It is tempting because smaller models reduce cost and latency, and one would be correct when throughput and expense matter more than factual fidelity.

• ✗

  Few-shot prompting and fine-tuning on more data

  Why it's wrong here

  Few-shot prompting and fine-tuning shape style and task format but do not ground output in source documents, so fabricated content persists. It is tempting because both are standard accuracy-improvement techniques, and they would be correct for adapting tone or domain vocabulary rather than verifying factual claims.

• ✓

  Retrieval-Augmented Generation (RAG) and Bedrock Guardrails

  Why this is correct

  Retrieval-Augmented Generation grounds each summary in retrieved source passages, so the model cites actual paper content rather than relying on parametric memory. Bedrock Guardrails then applies contextual grounding checks, filtering responses unsupported by those passages. Together they directly target the factual-accuracy constraint, reducing fabricated medical claims.

Go deeper

Related to this question