Tag: AI alignment
Browse our exclusive articles!
Jailbreaking Vision-Language Models via Visual Attacks
Discover how visual modality exploits vulnerabilities in vision-language models and learn key strategies to enhance AI safety and alignment.
TUR-DPO: Enhanced Preference Optimization for AI Models
Discover TUR-DPO, a topology- and uncertainty-aware method that improves AI preference optimization with better reasoning and calibration.
Debiasing Reward Models with Causal Inference Intervention
Reduce biases in reward models using causally motivated inference-time interventions to improve alignment with human preferences without losing performance...
Emergent Misalignment in AI: Consistency & Safety Insights
Explore emergent misalignment in large language models and its impact on AI safety, behavior, and reliability in critical applications.
Addressing Demographic Bias in LLM Safety Alignment
Explore the Selective Safety Trap in LLMs and discover how MiJaBench audits demographic biases to improve AI safety for all groups.