Human social perception is not a single snapshot. We infer intent before an utterance finishes, revise our interpretation as new cues arrive, and attend to specific facial, gestural, or prosodic signals that shape what something means. A new arXiv paper argues that current multimodal language models may not mirror this incremental process, and introduces GesBench as a way to measure it.

GesBench appears to be a benchmark designed to ask when and where models perceive social meaning. Rather than evaluating only final outputs, it likely tracks the temporal dynamics of model attention and prediction, comparing them to human patterns of cue use. The abstract highlights facial, gestural, and prosodic signals as key inputs, suggesting the benchmark includes video or audio-plus-text data.

The paper's abstract is truncated, so specific results or model comparisons are not yet available. Still, the framing is significant: if models are to interact naturally with people, they may need to update social interpretations on the fly, just as humans do. GesBench could provide a common test for that capability, though its full design and findings remain to be seen.