TuringEnterprises/Multimodal-STEM-HLE-plus-plus

multimodal-STEM-HLE++是一个高价值的多模态STEM数据集，旨在并经过实证证明，能够推动最先进的大型语言模型超越其当前极限。该数据集包含50个任务，覆盖物理学、数学、化学、生物学、计算机科学和工程学等6个领域及41个子领域，所有任务均需结合图像和文本进行联合推理，涉及图表、绘图、表格、方程和科学图形。数据集设计用于强化学习后训练，在现有最先进模型上具有约20%的通过率，处于最优难度区间，既能暴露真正的推理差距，又能提供可学习的奖励信号。每个任务都包含真实答案和详细的分步原理，适用于可验证奖励的强化学习流程，并已证明能显著提升前沿多模态STEM推理能力。multimodal-STEM-HLE++ is a high-value multimodal STEM dataset designed and empirically proven to push state-of-the-art LLMs beyond their current limits. It contains 50 tasks across 6 domains (Physics, Math, Chemistry, Biology, Computer Science, Engineering) and 41 subdomains, all requiring joint reasoning over images and text, including diagrams, plots, tables, equations, and scientific figures. The dataset is optimized for RL post-training, with approximately 20% pass@1 on SOTA models, placing it in the optimal difficulty regime to expose genuine reasoning gaps while providing learnable reward signals. Each task includes a ground-truth answer with a detailed step-by-step rationale, making it suitable for reinforcement learning with verifiable rewards pipelines, and it has proven utility in driving measurable improvements in frontier multimodal STEM reasoning.