A new arXiv paper, Training Language Models To Be Coherent Decision-Makers, argues that reliable decision-making requires more than accurate prediction. The authors frame the problem around three distinct requirements: a model must preserve its beliefs, apply the relevant utilities, and recognize when the information needed to justify an action is missing.
The paper's abstract suggests that these capabilities go beyond standard predictive accuracy. Without them, a language model may make predictions correctly yet still fail to act coherently, because it may lose track of what it knows, misapply what it values, or proceed without realizing that it lacks critical information.
As of the available abstract, the paper's empirical findings are not presented — the text cuts off after stating that the authors "study whether language…" Thus, the source establishes a conceptual framework rather than a reported result. Readers should look to the full paper for evidence on whether current language models meet these criteria.