Activating Self-Correction via RLHF

A research team is developing a large language model to provide helpful and safe responses. They implement an iterative process where the model first generates a response, then critiques its own response against a set of principles (e.g., 'is the response factually accurate?', 'is it free of harmful bias?'), and finally, revises the response based on the critique. How does viewing this self-improvement process as an 'alignment problem' provide the most accurate analysis of the team's goal?

Analyzing Misaligned Self-Refinement

Connecting Self-Refinement and Alignment

Evaluating the 'Alignment' Framing of Self-Refinement