Model techniques map
Techniquespost-trainingpreference optimization

implementation detail · filed under post-training

Budget-based verbosity control

A reward-hacking mitigation that penalizes a candidate in binary comparison when its output exceeds a length budget estimated from the cold-start model.

source
1
model
1
lab adopt it
1
strongest
used

How sources treat it

One count per evidence span, weakest treatment to strongest.

used 1

Documented in

Evidence

1 span quoted from the sources, strongest treatment first.

To mitigate reward hacking toward increasingly verbose outputs, we apply a budget-based verbosity control analogous to the reasoning-effort control above: given an initial verbosity ℓ0 estimated from the cold-start model and a multiplier σ, a candidate whose output length exceeds σ·ℓ0 automatically loses the binary comparison.

usedunclearin Kimi K3Moonshot AI

Filed alongside

Other methods under post-training :: preference optimization.