Changing language models’ confidence affects whether they answer, study author says
The author describes a Nature Machine Intelligence paper as causal evidence that large language models use confidence to guide behavior, rather than just report it.
TLDR
Boosting or suppressing confidence changes whether large language models answer or abstain, according to one author of an open-access Nature Machine Intelligence paper. The author says this provides causal evidence that confidence guides model behavior.
Combined views
43.4K
3 Sources, first seen 24d ago
likes