TY  - RPRT
TI  - A decodability criterion predicts when hidden-state selection beats majority voting in large language models
AU  - Zhixiang wang
AU  - Ziliang Hong
AU  - Ulas Bagci
PY  - 2026
UR  - https://arxiv.org/abs/2608.17124
ID  - 2608.17124
ER  -