Breaking01 / AI
AI benchmark tests nonverbal reasoning in early childhood
Researchers reported that vision-language models recognised context well but struggled to interpret specific behavioural and interpersonal cues.

Researchers introduced a benchmark using roughly 12-second interaction clips to assess how well AI models understand young children's nonverbal communication, according to a paper published on arXiv on 11 Nov.
The paper describes gaze, gestures and physical contact as cues through which children build meaning. The researchers said they selected clips to cover a broad range of natural settings, cultural contexts and developmental stages.
The evaluation framework builds from individual interaction cues towards social reasoning. Its categories cover context, behaviour, interpersonal interaction and open-ended descriptions.
In researcher-published findings, vision-language models, which process visual content and language together, showed strength in recognising context. The researchers reported substantial performance gaps in interpreting precise behavioural and interpersonal cues.
The researchers said existing benchmarks prioritised adults' conversation-driven interactions. They argued that broad evaluation frameworks mixed basic perception with higher-level reasoning.
The paper's abstract does not disclose scores for individual models or the total number of video clips.