Open Models Almanac
Methodsinference & servingreasoning control

single source · inference & serving

Effort-dependent exponential token-penalty schedule

An exponentially effort-dependent penalty on reasoning-token length is stronger at lower effort and relaxes as requested effort increases.

The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See reasoning control for the methods filed beside it.

Also called effort-dependent token-penalty coefficient, exponential penalty schedule, exponential token penalty schedule.

Filed under inference & serving › reasoning control