implementation detail · filed under post-training
Capping trajectory staleness
Limits the age of trajectories used for training, with the cited implementation using a ten-optimizer-step cap.
- 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.
We also cap the staleness of trajectories used for training at 10 optimizer steps, which we empirically found to be a good corner between speed and stability.
usedinference servingin LAGUNA M.1/XS.2Poolside
Filed alongside
Other methods under post-training :: rollout & RL infrastructure.
Asynchronous reinforcement learningPartial rolloutAsynchronous RL frameworks for large-scale agent scaffolds and environment orchestrationToken-in-token-out (TITO)Asynchronous reinforcement learning infrastructureCo-located RL trainingData SchedulerDecoupled control plane and data planeDecoupling agent rollout into sandbox and worker containerDeficit-corrected schedulingLarge-scale asynchronous RL in synthesized tasksOne-step off-policy asynchronous reinforcement learningPredictive Rollout DispatchSample-grained garbage collectionSeamless Rollout EngineSLIMEToken-granularity persistence of rollout statesToken-level interruptionTool ManagerToolboxAdaptive Rollout ConcurrencyAdaptive Rollout SchedulingAgent LoopAgent-centric rollout execution