2026 * Equal contribution
Why Agents Compromise Safety Under Pressure
ACL'26 ACL 2026 Findings
Hengle JIANG, Ke Tang
Why benign agents fail without adversarial attacks? We found that LLM-based Agents exhibit normative drift under compounding task constraints in real-world applications. We provide detailed pressure taxonomy and a safety framework to mitigate such behavior.
MoViD: View-Invariant 3D Human Pose Estimation via Motion-View Disentanglement
SenSys'26 ACM/IEEE SenSys 2026 (51/271 accepted)
Yejia Liu*, Hengle JIANG*, Haoxian Liu, Runxi Huang, Xiaomin Ouyang
MoViD enables robust, cross-view 3D human pose estimation and real-time edge deployment by explicitly disentangling motion from view information and employing a view-aware inference strategy.
WearBCI Dataset: Understanding and Benchmarking Real-World Wearable Brain-Computer Interfaces Signals
SenSys'26 ACM/IEEE SenSys 2026 (51/271 accepted)
Haoxian Liu, Hengle JIANG, Lanxuan Hong, Xiaomin Ouyang
WearBCI introduces the first multimodal wearable BCI dataset featuring synchronized EEG, IMU, and egocentric video to systematically benchmark denoising methods under real-world motion artifacts and explore the potential of IMU-assisted signal enhancement.
Agentic Pressure: The Endogenous Entropy of Reliable Autonomy
ICLR'26 ICLR 2026 Workshop on Reliable Autonomy
Hengle JIANG, Ziying Luo, Ke Tang
A phenomenological formulzation of Agentic Pressure, starting to study the endogenous risk of reliable agent systems.
2025
FreePose: Real-Time View-Invariant 3D Human Pose Estimation via Motion-View Disentanglement
MobiCom'25 ACM MobiCom 2025 Demo
Yejia Liu*, Hengle JIANG*, Xiaomin Ouyang
A demonstration of real-time view-invariant 3D human pose estimation running on edge devices with under 300ms latency, targeting IoT and mobile robotics applications.