Implement generative AI and agentic solutions 2.3 Optimize and operationalize generative AI systems
A team uses a non-reasoning chat model to answer multistep factual questions.
Study guide skill: Implement model reflection, chain-of-thought evaluations, and self-critique loops
Question B6-025
A team uses a non-reasoning chat model to answer multistep factual questions. Answers are sometimes wrong, and reviewers find them hard to check. Which two prompt techniques should the team apply? (Choose two.)
Select 2 answers.
Answer: D, E
Chain-of-thought prompting has the model work step by step and show its steps, which reduces inaccurate outcomes and makes the response easier to assess; it applies only to non-reasoning models. Breaking the task down, such as extracting facts first and then generating queries to verify them, also helps LLMs perform better. Extracting a reasoning model's reasoning by any method other than the reasoning summary parameter isn't supported, a higher temperature makes output more random, and an output cue only shapes the format.
Deep dive
How it works
Both correct answers come from Microsoft's prompt engineering techniques page. That page notes that these techniques aren't recommended for reasoning models like gpt-5 and the o-series, which fits this team's non-reasoning model.
Chain-of-thought prompting (only applicable to non-reasoning models): instead of splitting the task into smaller steps, you instruct the model to proceed step by step and present all the steps involved. Microsoft says this reduces the possibility of inaccurate outcomes and makes assessing the model response easier. Both problems in the question, wrong answers and hard-to-check answers, are addressed. The sample prompt asks for a step-by-step approach, reasoning before the final answer, and a fixed answer format.
Break the task down: LLMs often perform better when a task is broken into smaller steps. Microsoft's example restructures a prompt so the model first extracts relevant facts and then generates search queries that can verify those facts. With many factual claims, that can make a significant difference.
The other options, and when you'd use them
- Output cue for a bulleted list: A cue jumpstarts the output and suggests a format, such as bullet points. It shapes formatting, not accuracy.
- Raise temperature to 0.7: A higher temperature makes the output more random and divergent. That's the wrong direction for factual accuracy.
- Extract an o-series model's internal reasoning: Extracting model reasoning by any method other than the reasoning summary parameter isn't supported, may violate the Acceptable Use Policy, and can lead to throttling or suspension.
Exam tip
Non-reasoning model, multistep facts = chain of thought + break the task down (extract facts, then verify). Don't try to pull reasoning out of reasoning models.
References
Last reviewed: