How Should You Tune Prompt Model Settings to Reduce Unsupported Responses?
A company uses a prompt to generate responses based on internal support knowledge. The company requires the prompt to minimize generation of unsupported and incorrect information while ensuring responses remain useful. You need to tune the prompt model settings. How should you configure the model settings?
Community Insight
The question tests whether you can separate sampling parameters (temperature) from safety controls (content moderation); the trap is assuming a stricter moderation level reduces factual hallucination.
Prompt temperature controls how deterministic a model's output is when generating responses from internal support knowledge. Lowering the temperature is the correct configuration for minimizing unsupported or fabricated information while keeping answers useful.
Choosing to increase the content moderation level, assuming 'more moderation' equals 'more factual' — moderation filters harmful content, not unsupported claims grounded in the wrong source.
Community Discussion (1 comments)
Comments & Corrections
No comments yet — spotted an error or have a note? Share it below.
Expert Analysis
Why the Answer Is Correct
Temperature is the sampling parameter that governs randomness in token selection: higher values widen the distribution and invite creative, less grounded wording, while lower values concentrate probability on the most likely, source-consistent continuation. When a prompt must answer from internal support knowledge and must not invent unsupported details, decreasing the temperature is the direct lever. A low temperature keeps the model anchored to the phrasing and facts already present in the supplied knowledge rather than improvising plausible-sounding alternatives. It also satisfies the second half of the requirement — responses stay useful — because the model still answers, just more conservatively and consistently across runs.Why the Other Options Are Wrong
Content moderation settings (options A and B) govern whether harmful, violent, sexual or otherwise disallowed content is filtered or blocked; they do not evaluate factual grounding. Raising moderation does not stop a model from producing a confident but unsupported statement, and lowering it worsens safety without helping accuracy. Increasing the temperature (option D) is the opposite of what is needed: it deliberately introduces variability and creative divergence, which is desirable for brainstorming but counterproductive when the prompt must faithfully reflect internal support knowledge. Only option C changes the sampling behaviour that drives hallucination risk.Community Comment Notes
As aed747e noted, the intent is to "decrease the temperature setting" so the model becomes more deterministic and conservative, forcing factual responses tied to the supplied internal support knowledge. That reasoning matches the exam doctrine of treating temperature as the grounding-versus-creativity dial. No commenter argued that content moderation influences factual accuracy, which reinforces the single-answer conclusion.Exam Strategy
When a scenario mentions hallucination, unsupported claims, or staying faithful to provided knowledge, look for temperature (lower it) rather than moderation or top-p tuning. Reserve content moderation answers for questions about harmful or inappropriate output, and higher temperature for creativity or variety requirements.Official Reference
Exam Strategy
Map each setting to its purpose before reading the options: temperature = randomness/grounding, moderation = harmful content filtering. The phrase 'unsupported and incorrect information' is a direct signal to lower temperature, not to tighten moderation.
Frequently Asked Questions
Why doesn't increasing content moderation stop unsupported answers?
Moderation filters harmful or disallowed content categories; it does not check whether a claim is grounded in your internal support knowledge, so fabricated facts can still pass.
Will lowering the temperature make the prompt less useful?
No. It makes answers more consistent and conservative while still responding; only very low values risk terse or repetitive output, so test a moderate low setting.