This study introduces a Probabilistic Graphical Model (PGM) framework utilizing Pearl's do-operator to causally audit LLM safety mechanisms, specifically isolating the effect of injecting cultural demographics into prompts. A large-scale empirical analysis was conducted across seven instruction-tuned models from diverse origins, including the UAE's Falcon3-7B, as well as models from the US, Europe, China, and India, using ToxiGen and BOLD datasets. The findings revealed a disparity between observational and interventional bias, demonstrating that standard fairness metrics can overestimate demographic bias. Western models exhibited higher causal refusal rates for specific demographic groups, while Eastern models showed low overall intervention rates with targeted sensitivities toward regional demographics. Why it matters: This research highlights the geopolitical nuances of LLM safety alignment and the potential for demographic-sensitive over-triggering to restrict benign discourse, which is particularly relevant for diverse regions like the Middle East in developing culturally-aware AI.
Researchers at MBZUAI and other institutions have published a study at ACL 2024 investigating how jailbreak attacks work on LLMs. The study used a dataset of 30,000 prompts and non-linear probing to interpret the effects of jailbreak attacks, finding that existing interpretations were inadequate. The researchers propose a new approach to improve LLM safety against such attacks by identifying the layers in neural networks where the behavior occurs. Why it matters: Understanding and mitigating jailbreak attacks is crucial for ensuring the responsible and secure deployment of LLMs, particularly in the Arabic-speaking world where these models are increasingly being used.
MBZUAI-based startup LibrAI has launched the Libra-Leaderboard, an evaluation framework for LLMs that assesses both capability and safety. The leaderboard evaluates 26 mainstream LLMs using 57 datasets, assigning scores based on bias, misinformation, and oversensitivity. LibrAI also launched the Interactive Safety Arena to engage the public and educate them on AI safety through adversarial prompt testing. Why it matters: The Libra-Leaderboard provides a benchmark for responsible AI development, emphasizing the importance of aligning AI capabilities with safety considerations in the rapidly evolving LLM landscape.
Researchers at MBZUAI presented a new Arabic dataset at NAACL to measure LLM safety, building on a Chinese dataset called 'Do Not Answer'. The dataset includes nearly 5,800 questions with challenges and harmless requests containing sensitive terms to test for over-sensitivity. The team localized cultural concepts and added 3,000 questions specific to Arabic language and culture. Why it matters: This comprehensive benchmark, accounting for the diversity of Arabic dialects and cultures, advances the development of safer and more culturally aligned LLMs for Arabic speakers.
MBZUAI researchers have expanded LLM safety research to Chinese, presenting their work at the 62nd Annual Meeting of the Association for Computational Linguistics in Bangkok. They developed an open-source Chinese dataset of 3,000 prompts translated and localized from the English "Do-Not-Answer" dataset. The dataset includes a "region-specific sensitivity" category to address unique safety risks for Chinese speakers, evaluating if models are over-sensitive in identifying innocuous questions as harmful. Why it matters: This research addresses a critical gap in LLM safety evaluation, ensuring that language models are both safe and effective for diverse linguistic and cultural contexts, particularly in regions with unique sensitivities.