PsycheMME, a new video benchmark described on arXiv, takes a psychology-informed approach to evaluating social understanding and interaction. Its authors argue that understanding human needs and selecting responses require distinct judgments, and that averaging across tasks can hide failures in one area even when overall accuracy looks reasonable.
The benchmark spans eight capabilities, including grounding and mental states, though the abstract does not specify the remaining dimensions. By separating these aspects, PsycheMME aims to give researchers a more granular view of how video models perceive social situations and choose how to act.
Because this report is based only on the arXiv abstract, details about task design, annotations, and evaluation results are not yet available. The announcement positions PsycheMME as a step toward more psychologically faithful social reasoning in video models, but those claims await full documentation. If the benchmark delivers on its premise, it could offer a clearer lens on model strengths and gaps in real-world social interactions.