IdeaGene-Bench (IG-Bench)
收藏资源简介:
IdeaGene-Bench是由上海交通大学等多家机构联合构建的科学谱系推理评估基准,其核心框架将学术论文抽象为可遗传的'思想基因组'对象。该数据集涵盖10个科学领域的1,961条黄金谱系轨迹、1,085个精心标注的思想基因组对象及920对基因组差异记录,数据规模达到1,029个评估实例,主要来源于自然语言处理、计算机视觉等领域的经典文献。数据构建过程采用专家提名种子论文、基于引用的谱系扩展、多轮大语言模型辅助提取与专家审计相结合的四阶段流程,通过50名跨学科研究生进行严格质量验证。该数据集旨在评估人工智能系统在机制继承层面的科学谱系推理能力,解决当前评估体系难以检验研究提案是否真正继承核心机制、修复已知局限的深层次问题。
IdeaGene-Bench is a scientific lineage reasoning evaluation benchmark jointly constructed by Shanghai Jiao Tong University and several other institutions. Its core framework abstracts academic papers into heritable "thought genome" objects. This dataset covers 1,961 golden lineage trajectories across 10 scientific disciplines, 1,085 meticulously annotated thought genome objects, and 920 pairs of genomic difference records, with a total of 1,029 evaluation instances. It is primarily sourced from classic literature in fields such as Natural Language Processing (NLP) and Computer Vision (CV). The dataset construction process adopts a four-stage workflow combining expert-nominated seed papers, citation-based lineage expansion, multi-round Large Language Model (LLM)-assisted extraction, and expert auditing, with strict quality verification performed by 50 interdisciplinary graduate students. This benchmark aims to evaluate the scientific lineage reasoning ability of AI systems in terms of mechanism inheritance, and address the in-depth problem that current evaluation systems struggle to verify whether research proposals truly inherit core mechanisms and rectify known limitations.

- 1Ideas Have Genomes: Benchmarking Scientific Lineage Reasoning and Lineage-Grounded Idea Generation上海交通大学; 卡内基梅隆大学; 中国科学院大学; 上海人工智能实验室; 中国科学技术大学; 华东师范大学; 微软公司 · 2026年




