FINSIGHT: TOWARDS REAL-WORLD FINANCIAL DEEP RESEARCH

Jiajie Jin, Yuyao Zhang, Yimeng Xu, Yutao Zhu, Hongjin Qian, Zhicheng Dou

Keywords: Financial, deep research, report generation

Abstract:

Financial research reports are vital for investment decisions, yet their creation is a complex process that current AI methods struggle to fully automate due to challenges in domain knowledge, multimodal generation, and analysis depth. To address these limitations, we introduce FinSight, a novel multi-agent system designed to generate high-quality, multimodal financial reports.
At its core, FinSight features the Code Agent with Variable Memory (CAVM) architecture, which unifies external data, designed tools, and agents into a programmable variable space, driven by code to empower the system to process large-scale, multi-source data in an agentic manner. Recognizing the limitations of traditional methods in chart generation, we propose an Iterative Vision-Enhanced Mechanism to progressively refine visualizations into professional-grade charts. Furthermore, we have developed a Two-Stage Writing Framework that expands concise Chains-of-Analysis into comprehensive and coherent multimodal reports.
Experiments conducted on various company and industry-level tasks demonstrate that FinSight significantly outperforms all baselines, including leading deep research systems in terms of factual accuracy, analytical depth, and presentation quality, demonstrating a clear path toward generating reports that approach human-expert quality.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 7624