Open Models Almanac
Methodspost-trainingreinforcement learning algorithm

single mention Β· post-training

Group-relative length penalty

A group-relative length penalty is a reinforcement-learning reward adjustment that deducts reward from successful rollouts longer than a reference length derived from the prompt's group of sampled rollouts.1 The deduction applies only to successful rollouts, and a minimum pass-rate gate skips difficult groups to preserve room for exploration while curbing excessive token growth.2

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. we compute the reference length β„“β˜… π‘ž = Quantile𝐡/100{ℓ𝑗 : 𝑗 ∈ P π‘ž} and obtain the adjusted reward β€” MiMo V2 6 technical report
  2. The indicator1[Β·] restricts the deduction to successful rollouts β€” MiMo V2 6 technical report

Filed under post-training β€Ί reinforcement learning algorithm