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.
The paper introduces SalamahBench, a new benchmark for evaluating the safety of Arabic Language Models (ALMs). The benchmark comprises 8,170 prompts across 12 categories aligned with the MLCommons Safety Hazard Taxonomy. Five state-of-the-art ALMs, including Fanar 1 and 2, ALLaM 2, Falcon H1R, and Jais 2, were evaluated using the benchmark. Why it matters: The benchmark enables standardized, category-aware safety evaluation, highlighting the necessity of specialized safeguard mechanisms for robust harm mitigation in ALMs.
Researchers introduce UnsafeChain, a new safety alignment dataset designed to improve the safety of large reasoning models (LRMs) by focusing on 'hard prompts' that elicit harmful outputs. The dataset identifies and corrects unsafe completions into safe responses, exposing models to unsafe behaviors and guiding their correction. Fine-tuning LRMs on UnsafeChain demonstrates enhanced safety and preservation of general reasoning ability compared to existing datasets like SafeChain and STAR-1.
A new methodology emulating fact-checker criteria assesses news outlet factuality and bias using LLMs. The approach uses prompts based on fact-checking criteria to elicit and aggregate LLM responses for predictions. Experiments demonstrate improvements over baselines, with error analysis on media popularity and region, and a released dataset/code at https://github.com/mbzuai-nlp/llm-media-profiling.
Researchers from MBZUAI introduced RP-SAM2, a method to improve surgical instrument segmentation by refining point prompts for more stable results. RP-SAM2 uses a novel shift block and compound loss function to reduce sensitivity to point prompt placement, improving segmentation accuracy in data-constrained settings. Experiments on the Cataract1k and CaDIS datasets show that RP-SAM2 enhances segmentation accuracy and reduces variance compared to SAM2, with code available on GitHub.
Hamad Bin Khalifa University's Qatar Computing Research Institute (QCRI) introduced Fanar, an Arabic-centric multimodal generative AI platform featuring the Fanar Star (7B) and Fanar Prime (9B) Arabic LLMs. These models were trained on nearly 1 trillion tokens and are designed to address different prompts through a custom orchestrator. Fanar includes a customized Islamic RAG system, a Recency RAG, bilingual speech recognition, and an attribution service for content verification, sponsored by Qatar's Ministry of Communications and Information Technology. Why it matters: The platform signifies a major step towards sovereign AI development in Qatar, providing advanced Arabic language capabilities and addressing regional needs.
This paper introduces two methods for creating Arabic LLM prompts at scale: translating existing English prompt datasets and creating natural language prompts from Arabic NLP datasets. Using these methods, the authors generated over 67.4 million Arabic prompts covering tasks like summarization and question answering. Fine-tuning a 7B Qwen2 model on these prompts outperforms a 70B Llama3 model in handling Arabic prompts. Why it matters: The research provides a cost-effective approach to scaling Arabic LLM training data, potentially improving the performance of smaller, more accessible models for Arabic NLP.
MBZUAI researchers introduce PG-Video-LLaVA, a large multimodal model with pixel-level grounding capabilities for videos, integrating audio cues for enhanced understanding. The model uses an off-the-shelf tracker and grounding module to localize objects in videos based on user prompts. PG-Video-LLaVA is evaluated on video question-answering and grounding benchmarks, using Vicuna instead of GPT-3.5 for reproducibility.