Скачать или смотреть "LLM Factuality Scale Versus Rigor"

NotebookLM: " ... The FACTS Leaderboard [shown, is] a new, comprehensive benchmark created by Google DeepMind and Google Research to evaluate the factual accuracy of Large Language Models (LLMs) across diverse scenarios. The benchmark suite consists of four specialized sub-leaderboards designed to provide a holistic assessment: FACTS Multimodal tests factuality based on images and world knowledge; FACTS Parametric assesses recall of internal world knowledge through closed-book questions; FACTS Search measures performance when models must use a search API; and FACTS Grounding (v2) evaluates whether long-form responses are supported by provided context documents. The overall FACTS Score is an aggregate metric averaging performance across these four distinct tasks, and initial results show the top-performing model, Gemini 3 Pro, leading the benchmark but still leaving considerable room for improvement in LLM factuality."

https://arxiv.org/pdf/2512.10791
https://arxiv.org/abs/2510.09580