Next Token Prediction for Generalizable Human Activity Recognition (NextHAR; CIKM 2026)

Tokenizer: frequency-space sensor tokens ("motion-semantic tokens"). Pretraining: next-token-group prediction, modelling inter-channel and temporal structure jointly. Benchmark: SBHAR (=UCI HAPT), HHAR, Shoaib, Motion (MotionSense), DSADS, PAMAP2, TNDA, USC-HAD, KU-HAR; three settings of increasing difficulty, four datasets each. Protocol: fine-tune on a few labelled source samples, evaluate cross-dataset.

• Form factor: phone/wrist/body (benchmark mix)

One major challenge of sensor-based Human Activity Recognition (HAR) is data heterogeneity, where domain shifts between training and testing data could significantly degrade recognition performance. Recently, large language models (LLMs) have demonstrated strong cross-domain generalization capabilities by modeling language with next-token prediction, offering a promising paradigm for mitigating domain shifts in HAR. In this paper, we propose NextHAR, a novel self-supervised sensor pretraining framework for generalizable HAR. NextHAR first tokenizes sensor data in the frequency space, converting them into motion-semantic tokens. Then, to learn human motion dynamics, NextHAR jointly models inter-channel correlations and temporal dependencies by predicting the next sensor token groups. Furthermore, we introduce a comprehensive cross-domain HAR benchmark to enable systematic evaluation. Extensive experiments on this benchmark show that NextHAR outperforms state-of-the-art methods by 13.83% in accuracy and 15.00% in F1-score. Moreover, NextHAR achieves an average inference latency of 0.033 seconds on wearable devices, highlighting its efficiency for real-world applications. Our code and organized benchmark are publicly available at https://github.com/saiketa/NextHAR.

Verbatim from Semantic Scholar (opens in a new tab) · Abstract via Semantic Scholar; © the authors/publisher

No results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.

• Daily and Sports ActivitiesPretrainingEvaluation8 subjects · IMU
• HHARheterogeneity HAR, smartphones + smartwatches (n=9)PretrainingEvaluation9 subjects · IMU
• KU-HARKhulna University smartphone accel+gyro HAR (n=90, 18 activities, 100 Hz)Evaluation90 subjects · IMU
• MotionSenseiPhone 6s IMU HAR + attribute inference (n=24, 6 activities)PretrainingEvaluation24 subjects · IMU
• PAMAP2Physical Activity Monitoring (n=9, 3× IMU + HR)PretrainingEvaluation9 subjects · IMU · Temp · HR
• Shoaib Sensors 2014smartphone IMU HAR, 10 subjects, 7 activities, 5 body positions (50 Hz)Evaluation10 subjects · IMU
• TNDA-HARwrist/ankle/back IMU, 8 activities (n=23)PretrainingEvaluation23 subjects · IMU
• UCI HAPTSmartphone-Based Recognition of Human Activities and Postural TransitionsPretrainingEvaluationIMU
• USC-HADhip MotionNode acc/gyro/mag daily activities (n=14)PretrainingEvaluation14 subjects · IMU

• NextHARtransformernext-token-group prediction over frequency-space IMU tokens for cross-domain HARFoundation modelIMU2026