In a groundbreaking advance at the intersection of artificial intelligence and neuroscience, researchers at KAIST have unveiled an AI model that deciphers animal behavior with a linguistic-like understanding. The model, named BehaVERT, represents a pioneering step toward translating complex behavioral movements into a structured, interpretable language. This technological breakthrough opens new pathways for studying neurological conditions and behavioral patterns with unprecedented granularity.
At the core of BehaVERT is a transformative idea: animal movements contain latent structures analogous to human language. Rather than viewing animal behavior as isolated actions, the researchers approached it as sequences imbued with contextual meaning. By transforming skeletal keypoints—such as the movements of a mouse’s nose, ears, spine, limbs, and tail—into discrete tokens, BehaVERT treats behavioral sequences like sentences formed from words, enabling deep semantic analysis.
BehaVERT employs a BERT-based transformer architecture, a state-of-the-art deep learning model originally developed for natural language processing tasks. Transformers excel at capturing relationships across sequences by self-attention mechanisms, which allow the model to weigh the importance of different tokens contextually. This architecture empowers BehaVERT to not only classify behaviors frame by frame but also to comprehend overarching behavioral states spanning across entire sequences, thereby mimicking the way language models infer meaning beyond individual words.
To train this model, the KAIST team annotated skeletal movements from video footage using a web-based tool, converting them into high-dimensional tokens—each frame yielding 768-dimensional representations. Leveraging a self-supervised learning framework, BehaVERT was able to learn patterns directly from behavioral data without requiring extensive manual labeling. This approach reduced bias and enabled the system to uncover intrinsic behavioral semantics inherent in the data.
The robustness of BehaVERT was validated across five international benchmark datasets, demonstrating its versatility in analyzing social interactions, multi-animal dynamics, complex three-dimensional motions, and autism-related behavioral profiles. Notably, the model’s performance surpassed existing behavioral classification methods, reflecting its capacity to understand nuanced motion patterns that are often imperceptible to human observers.
One of BehaVERT’s most compelling achievements is its ability to provide interpretability. The model’s attention mechanisms can be visualized along behavioral timelines, indicating which specific motions influenced decision-making. This interpretability is crucial for bridging the gap between AI predictions and biological understanding, allowing neuroscientists to validate or discover new behavioral markers.
A striking example of BehaVERT’s interpretive power came from its analysis of an autism mouse model (Shank3B knockout mice). The AI independently highlighted oral-oral contact behaviors as key differentiators between autistic and healthy mice. This is particularly significant because it aligns with established biological findings—autism model mice exhibit social interaction deficits despite normal approaches, a subtlety that the AI rediscovered solely from raw behavioral cues without prior biological knowledge.
Further analysis showed that BehaVERT’s internal representation space intrinsically structured behavioral attributes such as mobility, attention, and social engagement into coherent clusters. This discovery suggests that animal behavior may possess an underlying semantic architecture reminiscent of linguistic grammar and syntax, shedding new light on how complex behaviors are organized and understood.
Interdisciplinary collaboration was pivotal in the development of BehaVERT. Remarkably, the lead researchers, including first author Dr. Seungjae Shin, were primarily trained in biological sciences rather than AI, exemplifying a successful fusion of domain expertise. Their dedicated self-education in transformer architectures and deep learning underpinned the creation of customized models and training approaches tailored for decoding motion semantics in rodents.
BehaVERT represents not just a new tool for behavioral classification but a platform for scientific discovery. By enabling interpretable behavioral decoding, it heralds applications in drug discovery, psychiatric disorder research, and behavioral genetics. The capacity to interpret the “language” of motion allows researchers to formulate hypotheses grounded in data-driven insights, accelerating the pace of experimental biology.
An additional technological feat demonstrated by the team is the adaptability of BehaVERT across species. A model trained on rat behavior successfully generalized to mice, highlighting the potential to develop behavioral foundation models with cross-species applications. This adaptability could revolutionize comparative behavioral studies, providing scalable analytic frameworks for diverse animal models.
KAIST’s ongoing commitment to AI-driven behavioral science is evidenced by their prior work on AVATAR, which reconstructs rodent behavior in virtual environments to facilitate detailed study. BehaVERT builds on this legacy by adding semantic understanding to motion analysis, marking a significant milestone toward holistic behavioral neuroscience.
In summary, BehaVERT is a visionary AI system that reads animal behavior as if it were a language, uncovering hidden meanings and structures within seemingly erratic motions. Its integration of skeletal tokenization, transformer architectures, and self-supervised learning delivers a potent tool bridging computational science and biology. As behavioral data continue to expand in scale and complexity, BehaVERT promises to revolutionize how we decode the fundamental language of life.
Subject of Research: Not explicitly stated
Article Title: BehaVERT: A Transformer-Based Motion Language Model for Decoding Behavioral Semantics in Mice
News Publication Date: 24-Mar-2026
Web References: http://dx.doi.org/10.1007/s11263-026-02834-y
References:
BehaVERT: A Transformer-Based Motion Language Model for Decoding Behavioral Semantics in Mice, International Journal of Computer Vision, DOI: 10.1007/s11263-026-02834-y
Image Credits: KAIST
Keywords: AI, animal behavior, transformer, BERT, motion language model, behavioral semantics, neuroscience, autism mouse model, deep learning, self-supervised learning, skeletal keypoints, behavioral analysis
