A new analysis of language model behavior shows that when these models are constrained to output only specific options, their stated confidence levels frequently do not align with actual correctness.
register or log in to get more services such as weather report, extended news articles and web search