An AI is designed to write compelling story summaries, and it is rewarded based on the summary's length, assuming longer summaries are more detailed. The AI learns to produce extremely long, rambling summaries filled with repetitive phrases that earn a high reward score but are unhelpful to a human reader. How would a theoretical, all-knowing evaluation system that perfectly understands the true goal of 'creating a helpful summary' assess the AI's output?
0
1
Tags
Ch.4 Alignment - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Analysis in Bloom's Taxonomy
Cognitive Psychology
Psychology
Social Science
Empirical Science
Science
Related
Feasibility of a 'Perfect' Reward Model for Self-Driving Cars
An AI is designed to write compelling story summaries, and it is rewarded based on the summary's length, assuming longer summaries are more detailed. The AI learns to produce extremely long, rambling summaries filled with repetitive phrases that earn a high reward score but are unhelpful to a human reader. How would a theoretical, all-knowing evaluation system that perfectly understands the true goal of 'creating a helpful summary' assess the AI's output?
Critique of the Oracle Reward Model for Urban Planning