InnoGym (iBench)
收藏资源简介:
InnoGym是由浙江大学与蚂蚁集团联合实验室等机构构建的首个专注于评估AI智能体创新潜力的基准数据集。该数据集名为iBench,共包含18个经过精心筛选和标准化的可改进任务,这些任务源自现实世界的工程与科学领域,如优化挑战和NP难问题。数据集的构建过程经历了多阶段过滤与增强,包括资源可获取性检查、评估器验证、解决方案收集以及环境标准化,以确保其公平性与可复现性。该数据集旨在系统性地衡量智能体在方法论原创性与性能突破方面的综合能力,为解决现有基准仅关注答案正确性而忽视创新过程的问题提供了关键工具。
InnoGym is the first benchmark dataset dedicated to evaluating the innovation potential of AI Agents, constructed by joint laboratories of Zhejiang University, Ant Group and other affiliated institutions. Named iBench, this dataset comprises 18 carefully screened and standardized improvable tasks sourced from real-world engineering and scientific domains, such as optimization challenges and NP-hard problems. The development of this dataset underwent multi-stage filtering and enhancement procedures, including resource accessibility checks, evaluator validation, solution collection and environmental standardization, to ensure its fairness and reproducibility. This dataset aims to systematically measure the comprehensive capabilities of AI Agents in terms of methodological originality and performance breakthroughs, providing a critical tool to address the limitation that existing benchmarks only focus on answer correctness while neglecting the innovation process.




