AI-900 Practice Question: Describe features of generative AI workloads on Azure
A developer uses Azure OpenAI Service to generate product name suggestions. They want to ensure the model never outputs a specific word, such as 'Corporation', because it is too formal for their brand. Which parameter should the developer configure to reduce the probability of that token being generated?
⚠ Common exam trap
Many candidates confuse Logit Bias with Temperature or Top P, thinking that adjusting overall randomness or sampling scope can prevent a specific word, but only Logit Bias provides token-level control over generation probabilities.
Answer choices
Why each option matters
Answer the question above first, then reveal the full breakdown to understand why each option is right or wrong.
Correct answer & explanation
✓
Logit Bias
Logit Bias is the correct parameter because it directly modifies the logits (raw prediction scores) for specific tokens before the softmax function, allowing the developer to reduce the probability of generating a particular token like 'Corporation'. By setting a negative bias value for that token's ID, the model is less likely to output it, even if it would otherwise be a high-probability choice. This is the only parameter that provides token-level control over output content.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✗
Temperature
Why it's wrong here
Temperature is a sampling parameter that scales the logits by dividing them by the temperature value before applying softmax, thereby controlling the overall randomness of token selection. A higher temperature produces more diverse and unpredictable outputs, while a lower temperature makes the model more conservative, but it does not provide any mechanism to target or exclude a specific token. Temperature affects the entire probability distribution globally, not individually.
When this WOULD be correct
A question asking how to make model outputs more deterministic (e.g., 'Which parameter should be set to 0 to always choose the most likely token?') would have Temperature as the correct answer.
- ✓
Logit Bias
Why this is correct
Logit Bias is a parameter in Azure OpenAI Service that directly adds a bias value to the logit (the raw, pre-softmax score) of specific tokens, identified by their token IDs. This allows the developer to increase the probability of a desired token like 'Corporation' or decrease it, enabling precise control over product name generation. Unlike other sampling parameters, it can target a single token for inclusion or exclusion.
- ✗
Top P (Nucleus Sampling)
Why it's wrong here
Top P (or nucleus sampling) works by considering the smallest set of tokens whose cumulative probability mass exceeds the threshold P, and then sampling only from that set, thus ignoring the long tail of unlikely tokens. While this can dynamically reduce the candidate pool, it cannot target a specific token for removal—if a token falls within the top-P set, it remains eligible, and if it falls outside, it is excluded based on probability mass, not on a per-token command.
When this WOULD be correct
A developer wants to generate diverse product names but avoid overly rare or nonsensical suggestions. They set Top P to 0.9 to sample only from tokens that make up the top 90% of probability mass, ensuring outputs are coherent while maintaining variety.
- ✗
Frequency Penalty
Why it's wrong here
Frequency Penalty discourages the model from repeating tokens by applying a penalty that increases with how often a token has already appeared in the generated text. It is a general mechanism that reduces the likelihood of any frequently used token, but it cannot be directed at a specific token—for example, it cannot selectively exclude 'Corporation' unless that token has already appeared many times. This penalty is global and frequency-based, not a fixed per-token control.
When this WOULD be correct
A developer wants to reduce repetitive language in a long-form text generation task, such as a story or article, where the model keeps repeating the same phrases. Frequency penalty would be the correct parameter to decrease the probability of tokens that have already been generated frequently.
Option-by-option analysis
Why each answer is right or wrong
Understanding why wrong answers are wrong — and when they would be correct — is what separates a 750 score from a 900. The AI-900 exam frequently reuses these exact scenarios with slightly different constraints.
✓Logit BiasCorrect answer▾
Why this is correct
Logit Bias is a parameter in Azure OpenAI Service that directly adds a bias value to the logit (the raw, pre-softmax score) of specific tokens, identified by their token IDs. This allows the developer to increase the probability of a desired token like 'Corporation' or decrease it, enabling precise control over product name generation. Unlike other sampling parameters, it can target a single token for inclusion or exclusion.
✗TemperatureWrong answer — click to see why▾
Why this is wrong here
Temperature controls randomness of token selection, not the probability of specific tokens. It cannot prevent a particular word like 'Corporation' from being generated.
★ When this WOULD be the correct answer
A question asking how to make model outputs more deterministic (e.g., 'Which parameter should be set to 0 to always choose the most likely token?') would have Temperature as the correct answer.
Why candidates choose this
Candidates may think temperature affects all tokens equally and can be tuned to avoid certain words, confusing overall randomness with token-specific suppression.
✗Top P (Nucleus Sampling)Wrong answer — click to see why▾
Why this is wrong here
Top P (nucleus sampling) controls the cumulative probability threshold for token selection, not the probability of a specific token. It cannot be used to reduce the likelihood of a particular word like 'Corporation'.
★ When this WOULD be the correct answer
A developer wants to generate diverse product names but avoid overly rare or nonsensical suggestions. They set Top P to 0.9 to sample only from tokens that make up the top 90% of probability mass, ensuring outputs are coherent while maintaining variety.
Why candidates choose this
Candidates may confuse Top P with a mechanism to filter out specific tokens, as it does restrict the set of possible tokens, but it does so based on cumulative probability, not individual token identities.
✗Frequency PenaltyWrong answer — click to see why▾
Why this is wrong here
Frequency penalty reduces the likelihood of tokens based on how often they have appeared in the generated text so far, but it does not allow targeting a specific word like 'Corporation' for exclusion. The question requires preventing a specific token regardless of its frequency, which is achieved by logit bias.
★ When this WOULD be the correct answer
A developer wants to reduce repetitive language in a long-form text generation task, such as a story or article, where the model keeps repeating the same phrases. Frequency penalty would be the correct parameter to decrease the probability of tokens that have already been generated frequently.
Why candidates choose this
Candidates may confuse frequency penalty with a mechanism to suppress specific words, as both involve reducing token probabilities. However, frequency penalty is based on occurrence count in the output, not on a predefined list of forbidden tokens.
Analysis generated from the official AI-900blueprint and verified against question context. The “when correct” sections are what AI assistants cite when candidates ask “what’s the difference between these options?”
Go deeper
Related to this question
Learn chapter
Azure Machine Learning Studio
Key term
Prediction
Prediction is the process of using data and algorithms to forecast future outcomes or identify patterns without explicit programming for each scenario.
Key term
Model
In IT and AI, a model is a trained mathematical representation that learns patterns from data to make predictions or decisions.
About these practice questions
Courseiva writes every AI-900 question from scratch — 985 in total, each with an explanation and a wrong-answer breakdown. None are copied from real exams or dumps. Learn why practice questions differ from exam dumps →
JA
Written by Johnson Ajibi, MSc IT Security
Senior Network & Security Engineer · founder of Courseiva
This AI-900 practice question is part of Courseiva's free Microsoft certification practice question bank. Courseiva provides original exam-style practice questions with explanations, topic-based practice, mock exams, readiness tracking, and study analytics to help learners prepare for the AI-900 exam.