1 related articles
MuSeToM is a human-validated textual benchmark with 2,415 instances covering six sensory cue types to evaluate LLM belief revision. Models sense belief change but struggle to identify the revised belief and over-update on salient yet insufficient cues.