First, we train a new variant of the Nemotron 3 Ultra GenRM that is trained with a reasoning length penalty to be more token efficient.
specific method · filed under post-training
Reasoning-length penalty
Penalizes reasoning length, including in a GenRM variant, to encourage token efficiency or control reasoning effort.
The sources on this page name it without explaining how it works. The line above is the extraction's one-line gloss. See reward modelling for the methods filed beside it.
Also called a reasoning length penalty, add the length-penalty term.
Documented in
Evidence
2 quotes from 2 sources, one card per model, each opening with its strongest.
How sources treat it
used 2
usedunclearin Nemotron 3.5 alignment teacherNVIDIA
DeepSeek-V4.1-Flash
DeepSeekwe add the length-penalty term 𝑟 len 𝑏,𝑗 to the reward of response
usedunclearin DeepSeek-V4.1-FlashDeepSeek