When we interact with others, we constantly guess what they know and believe. A key part of this is deciding whether some new piece of information, like a sound or a smell, is strong enough to change their mind. This is called epistemic belief revision: judging if evidence is sufficient, not just noticeable.
The authors created a test called MuSeToM with 2,415 text scenarios covering six senses. They asked to decide whether a character should update a belief after receiving a sensory cue, and then to say what the new belief would be. Some cues were genuinely informative, others were just eye-catching but weak.
The main finding was that models often correctly sensed when a belief might change, but then struggled to identify the correct new belief and too frequently treated flashy cues as solid proof. This matters because AI assistants need to reason accurately about what people know to be helpful and trustworthy.