Formula

Segment Score as Difference of Sequence Scores

When automating the rating of segments using a strong language model, it can be challenging to obtain human preferences directly at the segment level. One strategy is to use the language model to assign a cumulative rating score to the sequence up to the current segment, and then subtract the score of the sequence up to the previous segment. The score for a specific segment yˉk\bar{\mathbf{y}}_k is thus calculated as the difference between the LLM's score for the sequence ending at segment kk and the score for the sequence ending at segment k1k-1. The relationship is formulated as: s(yˉk)=s(yˉ1...yˉk)s(yˉ1...yˉk1)s(\bar{\mathbf{y}}_k) = s(\bar{\mathbf{y}}_1...\bar{\mathbf{y}}_k) - s(\bar{\mathbf{y}}_1...\bar{\mathbf{y}}_{k-1}) where s()s(\cdot) represents the rating score assigned by the language model.

Image 0

0

1

Updated 2026-05-03

Contributors are:

Who are from:

Tags

Ch.4 Alignment - Foundations of Large Language Models

Foundations of Large Language Models

Computing Sciences

Foundations of Large Language Models Course