Prepare and test your skills
Prepare and test your skills
Free sample question, answered and explained below. The Generative AI Leader practice bank has 2,191 more like it. Try 10 free: instant score, no account.
An AI engineering team has pre-trained a large language model on extensive text data. While the model is articulate, it occasionally generates biased responses, exhibits reasoning errors, and struggles to adhere to safety policies. The team wants to post-train the model so that it learns how to be safe, helpful, and aligned with human values through iterative exploration and feedback.
Which machine learning technique uses an agent-environment feedback loop and cumulative reward signals to achieve this alignment?
Supervised Fine-Tuning (SFT) is a model adaptation technique where a pre-trained base model is trained on curated, high-quality pairs of prompt inputs and target demonstrations. The model adjusts its weights using cross-entropy loss to maximize the exact token probability of the provided reference answers, learning through imitation rather than trial and error.
The team had a prepared dataset of curated question-and-answer pairs and wanted to train the model to mimic specific response formats without utilizing reward models or feedback optimization loops.
Reinforcement Learning from Human Feedback (RLHF) is an advanced machine learning method that optimizes foundation models by aligning their outputs with human preferences, safety standards, and intended goals. In this framework, the generative model operates as an agent taking actions (generating text or responses) within an environment (the user prompt and evaluation context). Instead of simply predicting subsequent words from a static dataset, the model explores diverse candidate answers and receives numeric scores via a feedback loop driven by a reward model calibrated on human judgements.
RLHF directly addresses the limitations of standard generative pre-training. Rather than treating all web data equally, it introduces human-guided feedback loops that turn a raw, unconstrained language generator into a safe, controllable enterprise assistant.
Unsupervised Pre-training (often called self-supervised pre-training) is the foundational phase of training large language models. During this stage, the model processes immense volumes of unstructured text from the internet, books, and code repositories to predict masked or subsequent tokens across sequences without human labeling.
The scenario described the initial stage of building a foundational model from scratch on large-scale text collections to acquire general linguistic knowledge and semantic understanding.
Reason and Act (ReAct) is an agentic prompting design pattern used during inference. It prompts a language model to alternate between generating natural language thoughts (reasoning steps), executing external actions (such as invoking search engines, calculators, or database APIs), and observing the resulting feedback from the tools to reach an answer.
The team needed an application architecture where a deployed model queries external APIs and databases sequentially to solve multi-step problems in real time without retraining or fine-tuning the model.
Change the inputs and watch how the outcome changes.

Change the technique and the objective to see how the training loop changes.
Answer them yourself and get your score the moment you finish. No account needed.
Full access includes
Keep the momentum going with these hand-picked practice scenarios
Last updated