Bits per sample

Definition

An information-theoretic lens decomposing RL efficiency into samples-per-FLOP times bits-per-sample; supervised learning extracts -log(p) bits per labeled example via cross-entropy, while sparse-reward RL extracts near zero at low pass rates, and soft targets carry far more bits than hard labels.

Cast within

No topic-region cast yet — this appears once Bits per sample's compiled claims are aligned into a topic region's argument tree.

Claims mentioning Bits per sample (3)

Your notes

Loading notes...