INVESTLOGICBENCH2026
收藏资源简介:
INVESTLOGICBENCH2026是一个面向大型语言模型投资逻辑评估的基准数据集,由论文作者构建。该数据集包含201,247条结构化决策记录,源自151位真实投资者的投资逻辑链,详细标注了市场事件、推理过程、决策行动及市场验证结果。数据集通过多源数据采集、双重逻辑提取(包括显性陈述分析与隐性模式推断)、基于真实市场数据的检索增强验证等四阶段流水线创建。其应用领域为金融智能体评估,旨在解决现有模型在动态市场环境中缺乏真实投资逻辑推理能力的问题,弥补静态问答基准与市场绩效之间的鸿沟。
INVESTLOGICBENCH2026 is a benchmark dataset for evaluating the investment reasoning capabilities of large language models (LLMs), developed by the authors of the corresponding research paper. This dataset contains 201,247 structured decision-making records derived from the investment reasoning chains of 151 real-world investors, with detailed annotations covering market events, reasoning processes, decision-making actions, and market validation outcomes. It was constructed via a four-stage pipeline encompassing multi-source data collection, dual logic extraction (including explicit statement analysis and implicit pattern inference), retrieval-augmented validation based on real market data, and additional standardized processing steps. Its targeted application is the evaluation of financial AI Agents, aiming to address the issue that existing models lack authentic investment logical reasoning capabilities in dynamic market environments, and bridge the gap between static question-answering benchmarks and actual market performance.




