[KongchangAI]
Conceptδ阈值 / 噪声阈值

Noise floor δ

A statistical threshold in benchmarking used to determine whether a score improvement is genuine. It is determined by the range of score fluctuation caused by benchmark randomness (sampling temperature, execution environment, test case sampling, etc.). Only differences between two runs exceeding this threshold are regarded as non-noise.

Timeline (last 90 days)

Oct 12

The noise floor δ is the statistical threshold for judging whether an improvement in a benchmark test is real, used to distinguish statistical significance from reaching a target.

Unverified50%

All Facts (1)

Source Articles