implementation detail · filed under inference & serving
Trajectory summarization and rollout re-initiation
A strategy that summarizes an overflowed trajectory and then re-initiates the rollout.
Also called Summary.
- sources
- 2
- model
- 1
- labs adopt it
- 0
- strongest
- evaluated
How sources treat it
One count per evidence span, weakest treatment to strongest.
evaluated 2
Documented in
Evidence
2 spans quoted from the sources, strongest treatment first.
Summary, which summarizes the overflowed trajectory and re-initiates the rollout
evaluatedinference servingin DeepSeek-V3.2DeepSeek
Summary, which summarizes the overflowed trajectory and re-initiates the rollout
evaluatedinference servingin DeepSeek-V3.2DeepSeek
Filed alongside
Other methods under inference & serving :: context management.
Discard-all context managementPreserved thinking history modeExcluding prior thinking from conversation historyContext compactionContext foldingDiscard-75%Preserve thinkingThinking context management for tool useContext management methodContext management strategyDiscarding tool-call historyHierarchical context managementKeep-recent-kMemory compressionModality-specific deploymentSummary-based context compressionTest-time context management for extending token budgets