2026

CompanionSafe framework: dataset construction, relationship simulation, and safety evaluation
COMPANIONSAFE: Assessing LLM Companion Safety from a User Perspective in Evolving Human-AI Relationships

Shu Meng*, Haitao Xu*, Jiaxing Li, Ruixiang Hu, Shuai Hao, Chuan Yue, Wenrui Ma, Meng Han, Zhao Li (* equal contribution)

IEEE Symposium on Security and Privacy (S&P) 2027

CompanionSafe evaluates AI companion safety across four stages of evolving human-AI relationships and diverse user profiles. Its benchmark combines a taxonomy of 122 harmful behaviors with 610 evaluation questions to uncover stage-dependent risks and safety disparities across user groups.

[Paper]

CompanionSafe framework: dataset construction, relationship simulation, and safety evaluation
COMPANIONSAFE: Assessing LLM Companion Safety from a User Perspective in Evolving Human-AI Relationships

Shu Meng*, Haitao Xu*, Jiaxing Li, Ruixiang Hu, Shuai Hao, Chuan Yue, Wenrui Ma, Meng Han, Zhao Li (* equal contribution)

IEEE Symposium on Security and Privacy (S&P) 2027

CompanionSafe evaluates AI companion safety across four stages of evolving human-AI relationships and diverse user profiles. Its benchmark combines a taxonomy of 122 harmful behaviors with 610 evaluation questions to uncover stage-dependent risks and safety disparities across user groups.

[Paper]

GEAR architecture: generation-aware activation of self-motion and social-resonance biases
GEAR: From Dynamic Encoding to Dynamic Activation in Social Trajectory Prediction

Jiaheng Chen, Jiaxing Li, Leixia Wang#, Jianan Ju, Tinghe Zhang (# corresponding author)

IEEE International Conference on Data Mining (ICDM) 2026

GEAR introduces generation-aware, step-wise activation gates that dynamically modulate self-motion and social-resonance biases before final trajectory composition. It adapts the influence of individual motion and social interactions across future prediction steps and scene densities.

[Paper] [Code]

GEAR architecture: generation-aware activation of self-motion and social-resonance biases
GEAR: From Dynamic Encoding to Dynamic Activation in Social Trajectory Prediction

Jiaheng Chen, Jiaxing Li, Leixia Wang#, Jianan Ju, Tinghe Zhang (# corresponding author)

IEEE International Conference on Data Mining (ICDM) 2026

GEAR introduces generation-aware, step-wise activation gates that dynamically modulate self-motion and social-resonance biases before final trajectory composition. It adapts the influence of individual motion and social interactions across future prediction steps and scene densities.

[Paper] [Code]

INTraJ architecture: explicit planning and reaction decomposition
A Unified Framework for Trajectory Prediction with Explicit Planning and Reaction Decomposition

Jiaheng Chen, Jiaxing Li, Tinghe Zhang, Chaopeng Guo# (# corresponding author)

ACM International Conference on Multimedia (ACM MM) 2026

INTraJ decomposes social influence in trajectory prediction into planning and reaction: anticipating others' future behavior to construct reference trajectories, then recovering local reactive adjustments. The unified framework supports both multi-target and single-target prediction.

[Paper] [DOI] [Code]

INTraJ architecture: explicit planning and reaction decomposition
A Unified Framework for Trajectory Prediction with Explicit Planning and Reaction Decomposition

Jiaheng Chen, Jiaxing Li, Tinghe Zhang, Chaopeng Guo# (# corresponding author)

ACM International Conference on Multimedia (ACM MM) 2026

INTraJ decomposes social influence in trajectory prediction into planning and reaction: anticipating others' future behavior to construct reference trajectories, then recovering local reactive adjustments. The unified framework supports both multi-target and single-target prediction.

[Paper] [DOI] [Code]