Tongliang Liu

Persona skills expose significant privacy risks, with existing defenses failing to adequately protect against attribute disclosure and impersonation across diverse agent architectures.

TC-UAP is the first method to effectively safeguard videos from both reference- and tuning-based customization, achieving unprecedented identity protection.

Deep transformers can encode complex grammatical structures in low-dimensional spaces, supporting the linear representation hypothesis.

Aggregating insights from diverse causal discovery experts with LLM-guided reweighting leads to significantly improved causal graph accuracy, even in ambiguous scenarios.

FlowBP redefines reward backpropagation by transforming the backward trajectory into a design object, leading to significant improvements in model alignment with human preferences.

Even minor real-world environment corruptions can cripple MLLM-powered computer-use agents, revealing a surprising fragility in their ability to execute desktop tasks.

DPO's reliance on a reference policy can backfire, prematurely halting learning when the reference is pessimistically wrong, but a simple one-line fix can significantly improve performance.

LLMs learn better from AI *reward* than AI *preference*, leading to higher human-AI agreement and improved performance compared to standard online AI feedback and RLHF.