Describe Artificial Intelligence workloads and considerations →mediumMultiple ChoiceObjective-mapped
AI-900 Practice Question: Describe Artificial Intelligence workloads and considerations
What is 'reinforcement learning from human feedback' (RLHF) and how is it used in training AI models?
⚠ Common exam trap
Many exam-takers confuse RLHF with simple supervised learning (Option C) or real-time feedback loops (Option D), missing the key distinction that RLHF uses a learned reward model from human preferences to guide reinforcement learning, not direct human rewriting or live retraining.
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
✓
Using human preference ratings to train a reward model that guides language model optimisation
Reinforcement learning from human feedback (RLHF) is a technique where human evaluators rank or rate model outputs (e.g., text completions) to create a reward model. This reward model then guides the fine-tuning of a language model using reinforcement learning, typically with Proximal Policy Optimization (PPO), to align outputs with human preferences.
Answer analysis
Option-by-option breakdown
For each option: why learners choose it and why it is or isn't the right answer here.
- ✗
A technique where humans physically assist robots in learning physical tasks
Why it's wrong here
Humans physically assisting robots to learn motor tasks is an example of imitation learning, kinesthetic teaching, or learning from demonstration, where trajectories or action sequences are observed and cloned. RLHF is specifically a technique for aligning language model outputs, not physical robot control, and it relies on a reward model trained from human preference judgments on text responses rather than on haptic guidance or physical manipulation. There is no embodiment or physical interaction involved in RLHF.
- ✓
Using human preference ratings to train a reward model that guides language model optimisation
Why this is correct
RLHF trains a reward model from human preference ratings: raters compare multiple model outputs, and a scalar reward function is learned to capture which responses are more helpful, harmless, and aligned. This reward model then serves as the optimization signal during reinforcement learning fine-tuning (typically with proximal policy optimization) to guide the language model's policy toward higher-scoring, more human-aligned outputs. The human preference ratings are the core supervision data that make the reward model possible.
- ✗
Having human reviewers manually rewrite model outputs to improve training data
Why it's wrong here
Having human reviewers manually rewrite model outputs is associated with supervised fine-tuning (SFT), where corrected text serves as target tokens for standard next-token prediction. RLHF does not directly edit outputs; instead it collects human preference ratings over multiple candidate responses, uses those ratings to fit a scalar reward model, and then optimizes the language model policy against that reward model via reinforcement learning. The key distinction is preference-based reward modeling versus label-based imitation learning.
- ✗
Allowing end users to flag incorrect answers to automatically retrain the model in real time
Why it's wrong here
End-user flagging of incorrect answers describes an online or continuous learning loop where the model is updated from live, aggregated user signals. RLHF, by contrast, uses curated, pairwise (or ranked) preference comparisons from human raters to train a separate reward model, which is then used as a fixed reward signal during a separate optimization phase. Real-time automatic retraining from user flags is neither a reward-model-based process nor does it involve the structured human preference elicitation central to RLHF.
Go deeper
Related to this question
Learn chapter
Machine Learning Core Concepts
Key term
Reinforcement learning
Reinforcement learning is a machine learning approach where an agent learns to make decisions by taking actions in an environment to maximize cumulative reward.
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
This AI-900 question is part of Courseiva's 985-question bank — original exam-style content with full explanations and wrong-answer analysis, never real exam questions or exam 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.