single mention · post-training
Group-relative length bonus
A group-relative length bonus augments the task reward during RL to reduce verbosity and inference costs while preserving task performance.1 For each prompt the policy samples a group of rollouts, and only correct rollouts receive a length-based bonus, computed relative to the shortest correct rollout in the group.2,3
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
To reduce verbosity and inference costs while preserving task performance, we augment the task reward with a group-relative length bonus for correct rollouts
— technical reportthe policy samples a group ofG rollouts{y1,...,y G}, each with correctness scoreci and lengthLi
— technical reportFor nonempty C, let Lmin = mini∈CLi
— technical report
Filed under post-training › reward modelling