implementation detail · filed under optimization
Per-token regularization for off-policy RL
Constrains policy updates to a localized neighborhood to tolerate highly stale data and sustain training stability.
- 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.
Our policy optimization algorithm inherently tolerates such an extreme off-policy regime through a per-token regularization. By constraining policy updates within a localized neighborhood, this regularization enables the algorithm to robustly handle highly stale data and sustains training stability.
usedunclearin Kimi K3Moonshot AI
Filed alongside
Other methods under optimization :: training stability.
Weight paddingActivation or logit clippingAnti-hallucination trainingBehavioral regularizationCross-replica model-weight hash consistency checksElevated constant-learning-rate training stability stress testExponential moving average of checkpointsGradient clippingLearning-rate elevation for training stability stress testingLoss masking for excessively stale tokensOff-policy sample filteringSafeguards against training drift and reward hackingSoft droppingTraining stability stress testingWeight clippingWeight decay coupled to learning-rate squared