AIF-C01 Practice Question 48: Human Preference Alignment
AWS Certified AI Practitioner (AIF-C01) · Test 5 · question 8 of 10
An AI team's assistant gives answers that are technically accurate but often come across as unhelpful or poorly toned. To align the model with what people actually prefer, the team has reviewers rank several model responses to the same prompts from best to worst. Those rankings train a separate model that scores responses, and that score is used to further optimize the assistant. Which technique is the team using?
Show answer and explanation
Answer:
RLHF is a machine learning technique that uses human feedback to optimize a model. Humans rank multiple model responses for quality, and those rankings build a separate reward model that learns human preferences. The language model then uses the reward model's scores to favor responses people are most likely to prefer. RLHF helps with subjective qualities that are hard to capture through standard training, such as tone, helpfulness and appropriateness, which makes output more helpful and less harmful (option B).
- A is wrong because supervised fine-tuning trains on *labeled input-output pairs*, one correct answer per prompt. There's no ranking of alternative responses and no reward model. - C is wrong because RAG retrieves external documents to ground responses at inference time. It doesn't train the model on human preferences. - D is wrong because clustering groups unlabeled data by similarity. The reviewers' rankings are explicit feedback used to build a reward signal, not data to be clustered.
Reference: What is RLHF?