implementation detail · filed under inference & serving
stage-wise curriculum over reasoning-effort budget multiplier
- 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.
Training follows a stage-wise curriculum over the budget multiplier τ. We first train a max-budget variant with a relatively large τ, while still capping the maximum budget to suppress excessive overthinking. We then anneal τ to smaller values to obtain the high- and low-effort expert models.
usedunclearin Kimi K3Moonshot AI
Filed alongside
Other methods under inference & serving :: reasoning control.
Configurable reasoning effortChain-of-thought reasoningDefault thinking modeDeployment-time scalar effort controlDisabling reasoning via chat-template configurationInference-time reasoning budget controlInterleaved thinking between tool callsThinking mode selectionCross-turn persistent reasoning historyMaximum thinking effortReasoning parserAlways-on thinking modeclear_thinking chat-template parameterControl-token-enabled thinking modeEffort-dependent exponential token-penalty scheduleGenerate-verify-refine loopMedium-effort reasoning modeParallel-fewest-step samplingQwen3 soft thinking switchTask- and mode-specific sampling parameter recommendationsTest-time compute scalingAdaptive reasoningCapped linear reasoning-token length deductionConfigurable thinking or reasoning mode