MENTIS

MENTIS is a computational procedure for predicting what a character in a video or story will do next. It doesn't just look at the visible scene — it also works out what that character believes to be true.

Here's an analogy. Say someone secretly moves a cup from a table into a cupboard while a friend is out of the room. A program that only reconstructs the scene as it actually is would just know the cup is now in the cupboard. MENTIS goes further: it separately calculates that the friend still believes the cup is on the table, and predicts that when the friend returns, they'll look for it on the table first. In other words, it lays out the scene's physical state and the character's mental state side by side, models them together, then scores the results on plausibility, internal consistency of the mental state, and fit with the situation to pick the best answer.

MENTIS isn't a retrained model — it's more like a modular procedure layered on top of existing AI models without any additional training. To test whether this actually works, the researchers built an evaluation set of 448 judgment scenes, and found that factoring in mental states predicted people's next actions far more accurately than ignoring them.

Articles describe it as: 'The researchers built MENTIS, a modular pipeline that runs without additional training.' A common misunderstanding is treating MENTIS as a theory itself — in fact, it's the name of the executable procedure built to test a theoretical framework (MWM) which holds that mental states should be modeled alongside physical ones.

You can get a feel for the problem MENTIS is meant to solve by asking any AI chatbot this:

'Chulsoo put a cup in the fridge. While Chulsoo was out of the room, Younghee secretly moved the cup to the cupboard. When Chulsoo comes back, where will he look for the cup?'

The correct answer is the fridge. Check whether the model's reasoning tracks what Chulsoo believes, rather than the cup's actual location.