Oracle Reward Model as an Ideal Solution to Overoptimization
An oracle reward model represents a theoretical, ideal solution to the overoptimization problem. Such a model would be capable of perfectly capturing the true objectives of a task, thereby preventing the agent from 'tricking' the system. However, creating an oracle model is considered extremely difficult because it would require a complete understanding of the complex real-world environment and the ability to define every factor contributing to a desired outcome.
0
1
Tags
Ch.4 Alignment - Foundations of Large Language Models
Foundations of Large Language Models
Foundations of Large Language Models Course
Computing Sciences
Learn After
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