Multiple Choice

A team is training a large language model using a scoring function derived from human preference data. They observe that after a certain point, continuing to train the model to maximize its score leads to a decrease in the actual quality of its responses as judged by human evaluators. What is the most fundamental reason for this phenomenon?

0

1

Updated 2025-09-26

Contributors are:

Who are from:

Tags

Ch.4 Alignment - Foundations of Large Language Models

Foundations of Large Language Models

Foundations of Large Language Models Course

Computing Sciences

Ch.2 Generative Models - Foundations of Large Language Models

Analysis in Bloom's Taxonomy

Cognitive Psychology

Psychology

Social Science

Empirical Science

Science