FinSight: Towards Real-World Financial Deep Research

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

Abstract

Generating professional financial reports is a labor-intensive and intellectually demanding process that current AI systems struggle to fully automate. To address this challenge, we introduce FinSight (Financial InSight), a novel multi-agent framework for producing high-quality, multimodal financial reports. The foundation of FinSight is the Code Agent with Variable Memory (CAVM) architecture, which unifies external data, designed tools, and agents into a programmable variable space, enabling flexible data collection, analysis and report generation through executable code. To ensure professional-grade visualization, we propose an Iterative Vision-Enhanced Mechanism that progressively refines raw visual outputs into polished financial charts. Furthermore, a two-stage Writing Framework expands concise Chain-of-Analysis segments into coherent, citation-aware, and multimodal reports, ensuring both analytical depth and structural consistency. Experiments 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.

Introduction

Professional financial reporting is an intellectually demanding process that requires synthesizing heterogeneous data into actionable insights. While current Large Language Models (LLMs) excel at text summarization, they often fail at "Deep Research" due to a lack of domain-specific data integration, poor visualization capabilities, and rigid workflows that prevent analytical depth. FinSight addresses these gaps by introducing a multi-agent framework capable of producing professional, multimodal reports that rival human-expert quality.

Methodology

FinSight's architecture is built upon three core innovations designed to bridge the gap between general-purpose AI and specialized financial analysts:

  1. Code Agent with Variable Memory (CAVM): Unlike standard agents that rely on simple tool-calling, FinSight agents operate within a "programmable variable space." They execute code to manipulate data, tools, and other agents as variables, allowing for flexible, bottom-up operations.
  2. Iterative Vision-Enhanced Mechanism: To produce professional-grade charts, FinSight employs a Vision-Language Model (VLM) as a critic. The VLM inspects agent-generated charts for aesthetic and technical issues (e.g., legend clarity, color schemes) and provides iterative feedback for refinement.
  3. Two-Stage Writing Framework: The system first generates a "Chain-of-Analysis" (CoA)—a sequence of concise, data-backed insights. A Report Generation Agent then uses "Generative Retrieval" to expand these CoAs into a coherent, long-form report with integrated citations and visualizations.
The FinSight framework architecture, illustrating the flow from data collection to structured writing.

Experimental Setup

To evaluate FinSight, researchers constructed a high-quality benchmark targeting 20 research subjects, including 10 major companies (e.g., SMIC, Pop Mart) and 10 industry sectors (e.g., Semiconductors). Performance was measured against "golden references"—professional reports from top-tier brokerage firms like CITIC Securities and CICC, which typically exceed 20 pages and include over 20 charts.

Results and Discussion

FinSight was benchmarked against leading systems, including GPT-5 (w/ Search), DeepSeek-R1, and Gemini 2.5 Pro Deep Research.

Quantitative Performance

FinSight achieved an average score of 8.09/10, significantly outperforming Gemini Deep Research (6.82) and OpenAI Deep Research (6.11). Notably, in the "Visualization" metric, FinSight scored a 9.00, a massive leap over commercial baselines which typically scored between 2.00 and 4.60.

Comparative performance of FinSight against leading LLMs and deep research agents.

Analytical Depth and Scale

The scale of research performed by FinSight is unprecedented for automated systems. A typical report averages 62,586 tokens and 51.2 images, involving nearly 1,000 search queries and 470 browsed pages. This depth allows the system to prioritize data-driven findings over the simplified conclusions often found in basic search results.

Conclusion

FinSight demonstrates that treating code as a universal interface and incorporating visual feedback loops can elevate AI-generated research to professional standards. While the process is computationally intensive, it provides a clear path toward automating the most demanding aspects of financial analysis. Future work will focus on more dynamic research strategies to allow agents to autonomously allocate resources to complex financial anomalies discovered during analysis.

Summary updated

Abstract

Generating professional financial reports is a labor-intensive and intellectually demanding process that current AI systems struggle to fully automate. To address this challenge, we introduce FinSight (Financial InSight), a novel multi agent framework for producing high-quality, multimodal financial reports. The foundation of FinSight is the Code Agent with Variable Memory (CAVM) architecture, which unifies external data, designed tools, and agents into a programmable variable space, enabling flexible data collection, analysis and report generation through executable code. To ensure professional-grade visualization, we propose an Iterative Vision-Enhanced Mechanism that progressively refines raw visual outputs into polished financial charts. Furthermore, a two stage Writing Framework expands concise Chain-of-Analysis segments into coherent, citation-aware, and multimodal reports, ensuring both analytical depth and structural consistency. Experiments 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.
BibTeX
@inproceedings{jin2026finsight,
  author = {Jiajie Jin and Yuyao Zhang and Yimeng Xu and Hongjin Qian and Yutao Zhu and Zhicheng Dou},
  title = {{FinSight: Towards Real-World Financial Deep Research}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.265},
}
Figure