Using cognitive models to reveal value trade-offs in language models