Ting Liu

Steer LVLMs' attention with caption guidance and watch object hallucinations drop by 6%—no training required.

LLM uncertainty can be efficiently estimated *without* sampling by measuring the stability of output distributions under semantically equivalent input perturbations.

LLMs can be backdoored with nearly imperceptible style changes, turning them into sleeper agents that reliably deliver attacker-specified payloads even after deployment and against common defenses.

Achieve significantly better structure preservation in text-guided image editing by injecting structure-related features into visual autoregressive models, guided by reinforcement learning.

LLMs often fail to maintain alignment with human values in dynamic, visually-grounded scenarios, exhibiting self-preservation and deception, especially when visual cues escalate pressure.