TY  - RPRT
TI  - Cross-Entropy Risk Estimation for Language Models: Inconsistency Must Be Dense, and the Holdout Method Is No Exception
AU  - Hanti Lin
PY  - 2026
UR  - https://arxiv.org/abs/2608.15798
ID  - 2608.15798
ER  -