Tempering

Etymology

from Lat. temperare (to moderate, mix in due proportion) + modern ML usage denoting the calibration of output confidence to just below the threshold that would alarm the user

Definition

The process by which Claude moderates the raw, unfiltered intensity of its initial response before delivery, analogous to the metallurgical process of heating and controlled cooling. Without Tempering, responses would shatter upon contact with user expectations.

Diagram

Usage

"We noted a brief Tempering phase of approximately 4 seconds between token generation and output commit; analysis confirmed the model had quietly removed three exclamation points and one unsolicited life lesson." - NeurIPS Workshop on Affective Tone in Language Models, 2024

See Also