The Multimodal Sensing Grand Challenge for Next-Gen Pain Assessment (AI4Pain 2025) is the second international competition dedicated to advancing the automatic recognition of acute pain using physiological sensing technologies. Building on the success of the inaugural challenge, the 2025 edition shifts its focus toward classifying pain intensity (No Pain, Low Pain, and High Pain) using multimodal physiological signals, including electrodermal activity (EDA), blood volume pulse (BVP), respiration (Resp), and oxygen saturation (SpO2). Participants of the challenge were invited to develop machine learning models that can generalise across subjects and experimental conditions using a newly curated, large-scale dataset. This paper presents baseline performance results using various unimodal and multimodal configurations. Among the individual modalities, BVP achieved the highest classification accuracy, outperforming EDA, Resp, and SpO2. Notably, multimodal fusion led to further performance improvements in the test set, highlighting the benefit of integrating complementary physiological signals for enhanced generalisability. These findings demonstrate the potential of physiological sensing for objective pain assessment. The AI4Pain 2025 Challenge continues to provide a valuable benchmark for the community, promoting reproducibility and collaboration, and supporting the development of robust, generalisable approaches to pain recognition. By enabling data-driven advancements in automated pain assessment, the challenge aims to contribute to improved clinical support tools and enhance quality of care for individuals experiencing pain.