Why it matters

Min-p is emerging alternative. Understanding shapes modern sampling.

Advertisement

The architecture

Find max prob p_max.

Threshold = min_p * p_max.

Keep tokens above threshold.

Min-p sampling flowMax probpeak of distThresholdmin_p * maxSample aboverenormalizeAdaptive to distribution shape; often 0.05-0.1 for min_p
Min-p.
Advertisement

How it works end to end

Adaptive to peak.

Prevents low-prob nonsense.

Preserves reasonable diversity.