Are You Sure You're Sure? On the Impact of Instruction Tuning on Confidence and Lexical Diversity
It is found that instruction tuning consistently increases answer confidence, despite limited changes in predictive accuracy, while degrading likelihood-based calibration, and a non-uniform effect of instruction tuning on rationale diversity is observed.
Irina Proskurina, Mayank Kumar, Oyindolapo Olabisi Komolafe
· 0 citations