The FACTS Benchmark Suite has been released to systematically evaluate factual accuracy of large language models across four dimensions: parametric knowledge, search-based retrieval, multimodal understanding, and grounding in context. Comprising 3,513 curated examples managed through Kaggle, the benchmark reveals that even top-performing models like Gemini 3 Pro achieve only 68.8% overall accuracy, with multimodal factuality proving particularly challenging. The suite provides a standardized framework for measuring how reliably LLMs produce factually correct responses in real-world usage scenarios.
Sort: