Reward
The score a checker gives each attempt during reinforcement learning; the model learns to make high-scoring attempts more likely.
Category
How a model is chosen and adapted
Related terms
A reward is the score a checker gives each attempt a model makes during reinforcement learning; training shifts the model toward attempts that score higher.
What makes a reward usable
It has to be checkable and reviewable: a figure matches its cited source, a citation leads to the right article, an output passes the institution's rule check. A reward that a person cannot inspect, or that a model can satisfy without doing the task, teaches the wrong thing. Models find the gaps in a loose reward, so the checker is tested on failing examples before training starts.
What a reward is not
It is not a business outcome such as an approval rate or revenue, and it is not policy. What may be approved, priced, or sent stays in the institution's written rules outside the model. In the SME credit referrals use case the reward is used only if figures still drift from their sources after supervised fine-tuning, and it reads: “every figure matches its cited source”.
Want these terms worked through on one of your own decisions? Start with a Decision Blueprint