Question

how does reinforcement learning improve the performance of AI models like chatGPT.
By exposing model only to positive feedback
by adjusting the model based on human ratings of multiple out put
simplifying the model's algorithms
by minimizing the training data

Answer 1

Correct choice: "by adjusting the model based on human ratings of multiple output(s)."

Why:
- Modern systems like ChatGPT use RL from human feedback (RLHF). Humans rank or rate different model outputs for the same prompt.
- Those ratings train a reward model that scores responses. The base language model is then fine-tuned with a reinforcement learning algorithm (e.g., PPO) to maximize that reward, so it produces answers more aligned with human preferences (helpfulness, safety, style).
Why the other options are incorrect:
- "By exposing model only to positive feedback": not true — RLHF uses relative preferences (positive and negative comparisons) and can learn from both good and bad examples.
- "Simplifying the model's algorithms": RLHF changes how the model is trained, not by simplifying its architecture or core algorithms.
- "By minimizing the training data": RLHF typically uses additional annotated data (comparisons/ratings), so it does not minimize data use.

Short summary: RLHF improves alignment and quality by using human judgments to shape a learned reward and then optimizing the model to produce outputs that humans prefer.