cad-alpha-zero-standard-decoding

IN premise — summaries/2026/08/24/shi-2024-context-aware-decoding-s1-introduction.md

Created 2026-08-25T02:58:34+00:00

Setting α=0 in the CAD formula reduces the adjusted distribution exactly to standard autoregressive decoding (p_θ(y_t|c,x,y_

Summary

Setting alpha to zero in the CAD formula gives back exactly the same result as vanilla autoregressive decoding, which means the CAD framework is a strict generalization of the standard approach rather than a replacement. This guarantees the system can always fall back to known-good baseline behavior by dialing alpha down to zero, and that any departure from standard decoding is explicitly controlled by that single parameter.