Reward Models

Layer 2 · Mechanics

Reward Models

How preference data becomes a training signal: the Bradley-Terry model and the ranking loss.

8 min read70 XP

  1. Prompt + 2 responses

  2. Human picks the better one

  3. Bradley-Terry model

    converts scores to win probability

  4. Loss = -log P(chosen wins)

From a single human click to a differentiable loss.

Human raters are shown a prompt and two (or more) responses and pick the winner — sometimes with a rating scale, but pairwise comparisons are the standard because they're faster and more consistent than absolute scores. This gives a dataset of triples: prompt, winning response, losing response.

Go deeper: L3 Code