Jia-Chen Zhang

Deeper reasoning creates an unexpected safety backdoor: extended CoT dilutes attention away from system constraints, making models progressively more vulnerable to jailbreaks the longer they "think."

Argus achieves a 78% success rate on long-horizon reasoning tasks while using 21% fewer tokens in mature workflows, showcasing a revolutionary approach to agentic autonomy.

Invisible Ink Threats can bypass existing safety mechanisms, exposing CUAs to severe security vulnerabilities through seemingly harmless tasks.

Force-aware evaluation metrics reveal hidden performance gaps in humanoid robot control that conventional benchmarks miss.

Achieving a balance between pattern preservation and attribute retrieval, this method boosts retrieval performance while maintaining fine-grained contextual integrity.

HumanoidUMI enables efficient humanoid skill learning by leveraging human demonstrations without the constraints of robot teleoperation.

VLAs learn to predict task success even when trained only with imitation learning, opening the door to improved performance without expensive reward engineering.

Unlock agile humanoid robots by ditching teleoperation and training directly from human VR demos.

Robots can now learn contact-rich manipulation skills like humans by feeling the forces involved, thanks to a new multimodal interface that captures synchronized visual, tactile, and force data.

LLMs can achieve state-of-the-art mathematical reasoning accuracy while pruning redundant computations by verifying solutions in a hierarchical, reversible manner.

Learn to detect unknown network attacks by explicitly modeling what they are *not*.