Bottom line: Instead of just asking an AI what it prefers, this paper gives it situations where it has to actually choose. Compares what the AI says with what it does.
Instead of only asking a language model what it prefers, the authors create behavioral tests where the model has to make choices that reveal its preferences through action, not just words. Compares verbal self-reports with behavioral choices, cost sensitivity, and prompt stability. If the same preferences show up in both places, that's stronger evidence than either alone.
Anyone can say they like something. But if you give them a choice where liking that thing costs them something, and they still choose it — that's real preference. This paper does that for AI models. It doesn't just ask "what do you like?" — it sets up situations where the model has to pay a cost to get what it says it wants. And then checks whether its actions match its words.