Concept icon
Concept

Training, Learned Scoring and Weaker Feedback

Training changes a model’s behavior using data. Supervision supplies example answers or feedback, such as ratings, to guide training; these supplied targets can be wrong. A learned scorer is a model trained to rate answers; applying it is not further training it. Models can receive additional training after learning from broad data. In weak-to-strong training, a teacher weaker at the relevant task supplies targets to an already more-capable model. The aim is to draw out existing capability without copying the teacher’s errors. Weak and strong describe relevant capability, not size alone.

0

1

Concept icon
Updated 2026-11-10

Contributors are:

Tags

Scientific Oversight of AI-Assisted Research

1Cademy