Duohe Ma

reinforcement learning (1) adversarial robustness (1) direct preference optimization (1) preference optimization (1) safety alignment (1) multimodal large language model (1) hallucination reduction (1) adversarial prompt generation (1) noise-aware learning (1) large language model (1) modality bia (1) noise-aware preference optimization (1) modality bias mitigation (1) jailbreak attack defense (1) llm safety alignment (1) refusal direction ablation (1) internal defense mechanism (1)