A Korean research team developed an artificial intelligence (AI) model that analyzes mouse movements to determine the meaning of behavior.
A team led by Kim Dae-su, a professor in the Department of Brain and Cognitive Sciences at KAIST, said on the 1st that it implemented an AI model, "BehaVERT," that learns animal gesture data to detect social behavior abnormalities in autism model mice. The findings were published in March in the international journal in the Computer Vision field, the International Journal of Computer Vision.
The team recorded movements of body parts such as the mouse's nose, ears, spine, legs, and tail as coordinate data, then converted them into "tokens" to train the AI. Like the word pieces that split sentences in Natural Language Processing (NLP), tokens are the basic units AI uses to understand information. The researchers turned mouse movements into a kind of behavioral words so the model could read the flow of behavior over time.
BehaVERT applies a BERT-based transformer model used in Natural Language Processing (NLP) to animal behavior analysis. BERT is an AI model that determines meaning by examining the relationships of words before and after within a sentence. The team applied this principle to behavioral data so the AI could learn the context and meaning between behaviors, going beyond the level of merely classifying that "a behavior occurred."
In experiments, BehaVERT outperformed existing models on five international standard evaluations, including social interaction, multi-individual behavior, 3D movement, and autism behavior analysis. It also allowed researchers to see which movements the AI focused on to make its judgments, enabling interpretation of the results.
In particular, in experiments distinguishing autism model mice from typical mice, the team found that BehaVERT focuses on "mouth-to-mouth contact." Autism model mice are known to approach other individuals but have deficits in actual social interaction. The AI identified this feature from behavioral data alone, without learning separate biological knowledge.
The researchers said the results suggest that animal behavior, like language, may have a certain structure and meaning. This implies a shift from classifying behavior by preset human criteria to allowing AI to autonomously find important patterns within data.
Kim Dae-su said, "BehaVERT is an AI model that goes beyond classifying behavior to interpreting its meaning," adding, "It could be used as a tool to precisely analyze animal behavior in new drug development, mental illness research, and behavioral genetics."
References
International Journal of Computer Vision (2026), DOI: https://doi.org/10.1007/s11263-026-02834-y