BEAR-Bench is a new benchmark for multimodal models that evaluates their ability to reason about text-dense documents. The benchmark includes 1000 human-annotated questions based on English and Russian business and scientific documents. It aims to address limitations in existing benchmarks, which often focus on information extraction or require external knowledge. BEAR-Bench is a self-contained, complex benchmark that can help assess the strengths and weaknesses of multimodal models. --- Why it matters: This matters because it provides a more comprehensive evaluation of multimodal models' ability to reason about professional documents, particularly in languages like Russian which are underrepresented in existing benchmarks. Source: https://arxiv.org/abs/2608.17895