Researchers have created a benchmark to measure how well AI models can translate their understanding of social situations into coordinated actions. The benchmark, called MOSAIC, involves two AI agents interacting in scenarios that require them to consider each other's mental states and respond accordingly. In testing, most vision-language models failed to produce behaviors consistent with the expected outcomes, highlighting a gap between social inference and action. One model, PCM-LLM, performed better due to its explicit inclusion of a module for translating beliefs into actions. --- Why it matters: This research matters because it highlights the limitations of current AI models in understanding and responding to social cues, which is crucial for developing more human-like intelligence. The findings suggest that simply being able to reason about social situations is not enough; AI models need to be able to translate those understandings into coordinated actions. Source: https://arxiv.org/abs/2608.20975