HALLUCINOGEN is a novel benchmark that tests Large Vision-Language Models' vulnerability to hallucinations through contextual reasoning prompts.

• Evaluates models' tendencies to fabricate non-existent visual entities across diverse scenarios
• Provides systematic methodology to measure hallucination severity in multimodal AI systems
• Identifies specific contextual patterns that trigger unreliable AI responses
• Offers insights for developing more robust vision-language models

Security Implications: This research is critical for securing AI deployments by identifying when systems might confabulate information, helping prevent misinformation propagation and ensuring trustworthy AI in high-stakes applications like medical diagnostics and security systems.

Towards a Systematic Evaluation of Hallucinations in Large-Vision Language Models