Source-linked AI summary
FinSight: Towards Real-World Financial Deep Research
Jiajie Jin, Yuyao Zhang, Yimeng Xu, Hongjin Qian, Yutao Zhu, Zhicheng Dou
TL;DR
FinSight addresses the difficulty of producing accurate, deep, multimodal financial reports from large volumes of data. Its multi-agent workflow combines programmable data and tool integration, iterative chart refinement, and staged report writing; evaluations report that it outperforms deep research systems across key report-quality dimensions.
Problem
Producing accurate, analytically deep financial reports quickly remains difficult amid overwhelming data volumes and demands for multimodal content.
Source: §1 · p. 1Method
FinSight combines a multi-agent workflow for data collection, multi-turn analysis, and progressive report writing with CAVM, iterative chart refinement, and a two-stage writing framework.
Source: §1 · p. 2, §2.2 · p. 3, §2.4 · p. 5Results
FinSight significantly surpasses deep research systems in factual accuracy, analytical depth, and presentation quality across company- and industry-level tasks.
Source: §1 · p. 2Takeaways & Limitations
The experiments support FinSight's potential to generate multimodal financial reports approaching human-expert quality.
Source: §1 · p. 2
Abstract
from arXiv · showhide
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.
1 INTRODUCTION
FinSight targets the difficulty of producing accurate, deep, multimodal financial reports from overwhelming volumes of data. It combines a code-centered multi-agent workflow with iterative visualization refinement and staged report writing, and is evaluated on company- and industry-level tasks.
1 INTRODUCTION
- Current systems struggle to integrate heterogeneous financial data, produce visualizations, and adapt rigid workflows to intermediate findings.The paper frames these gaps against the volume of financial data and demand for rapid, high-quality analysis.Source: §1 · p. 1
- FinSight reconstructs expert research as data collection, interactive multi-round analysis yielding a Chain-of-Analysis, and report generation with chart and data references.The report follows a draft outline and is rendered in a professional style.Source: §1 · p. 2
- CAVM unifies data, tools, and agents in a programmable variable space that executable code can access and manipulate.This supports flexible, scalable work from bottom-up data operations to high-level workflow orchestration.Source: §1 · p. 2
- The Iterative Vision-Enhanced Mechanism refines code-generated charts through vision-language-model feedback, while the Two-Stage Writing Framework expands concise analysis into coherent reports with integrated visualizations and citations.The writing framework uses Chain-of-Analysis segments as a structured foundation for report generation.Source: §1 · p. 2
- FinSight significantly surpasses deep research systems across factual accuracy, analytical depth, and presentation quality on company- and industry-level financial tasks.The benchmark spans multiple markets and diverse industries, and the reported quality approaches that of human experts.Source: §1 · p. 2
2 METHOD
The method frames financial report generation as combining textual analysis, visualizations, and citations, then organizes FinSight into collection, analysis, and report-writing stages. CAVM, iterative chart feedback, and two-stage retrieval-based writing implement these stages.
2.1 PROBLEM FORMULATION
- A report comprises textual analysis, visualizations that support or extend it, and citations to data sources.Source: §2.1 · pp. 2–3
2.2 THE FRAMEWORK OF FINSIGHT
- FinSight organizes work into multi-source data collection, multi-turn analysis, and progressive report generation.Collection can be invoked dynamically during analysis and writing; report generation includes self-reflective optimization and post-processing.Source: §2.2 · pp. 3–4
2.3 CODE AGENT WITH VARIABLE MEMORY (CAVM)
- CAVM gives agents autonomy over a shared variable space of data, tools, and agents, dynamically updating context as code execution drives iterative work.At each step, execution feedback informs subsequent planning and modifies the variable space and interaction history.Source: §2.3 · p. 4
2.4 ITERATIVE VISION-ENHANCED MECHANISM FOR VISUALIZATION
- For each specified chart, executable plotting code produces an initial visualization; VLM feedback directs revisions until convergence or a preset quality threshold.Source: §2.4 · p. 5
2.5 TWO-STAGE WRITING WITH GENERATIVE RETRIEVAL
- Writing first creates parallel perspective-specific CoAs, then drafts outlined sections by retrieving relevant data and CoAs while following established figure and reference identifiers.Source: §2.5 · pp. 5–6
3 EXPERIMENTS
FinSight is evaluated on a benchmark of company- and industry-level financial research tasks against search-enabled language models and commercial deep-research agents. It achieves the highest overall score and leads across reported factuality, analytical-quality, and presentation measures, with consistency as an exception.
3.1 DATASET AND EVALUATION METRICS
- The benchmark addresses a lack of suitable financial-report evaluation resources with company- and industry-level targets paired with professional brokerage reports.The dataset contains 20 targets: 10 company-level and 10 industry-level; selected reference reports exceed 20 pages and contain more than 20 charts and visualizations.Source: §3.1 · p. 6
- Nine 0–10 automated metrics assess factual accuracy, information effectiveness, and presentation quality.They cover conclusion consistency, textual faithfulness, text-image coherence, information richness, coverage, analytical insight, structural logic, language professionalism, and chart expressiveness.Source: §3.1 · p. 6
3.2 BASELINES
- FinSight is compared with search-tool LLMs and commercial deep-research agents.The baselines include GPT-5, DeepSeek-R1, Claude-4.1-Sonnet, Gemini-2.5-Pro Deep Research, Grok Deep Search, OpenAI Deep Research, and Perplexity Deep Research.Source: §3.2 · p. 6
3.4 MAIN RESULTS
- 8.09 overall score puts FinSight ahead of Gemini Deep Research (6.82) and OpenAI Deep Research (6.11) on the financial research report generation benchmark.FinSight also has the best faithfulness-of-text-citations and text-image-consistency scores, while its consistency score (6.85) is below Gemini Deep Research (7.10).Source: Table 1 · p. 7, §3.4 · p. 7
3.3 IMPLEMENTATION DETAILS
- The implementation uses DeepSeek-V3 as its backbone and DeepSeek-R1 for writing, with Google Search API retrieval and Gemini-2.5-Pro evaluation.The input and output limits are 81,920 and 16,384 tokens, respectively; search retrieves the top 10 results in the China region.Source: §3.3 · p. 7
3.4 MAIN RESULTS
- FinSight scores highest in information richness, key-information coverage, insightfulness, and presentation measures, including visualization (9.00).Its consistency score is a reported exception: 6.85 versus 7.10 for Gemini Deep Research.Source: §3.4 · p. 7
4 ABLATION STUDIES
Ablations report declines when chart-feedback, two-stage writing, or dynamic search is removed. The analysis also describes report scale, iterative chart refinement, and the relationship between report length and quality across methods.
4 ABLATION STUDIES
- Removing iterative VLM feedback lowers Presentation Quality from 8.0 to 7.5 and Analytical Quality from 7.9 to 7.2.The authors attribute the decline in analysis to lower-quality visuals impeding insightful chart-based analysis.Source: Table 2 · p. 7, §4 · pp. 7–8
- Merging analysis and writing lowers analytical quality from 7.9 to 5.9 and factual accuracy from 7.0 to 6.4.The authors present this ablation as evidence for the two-stage analyze-then-write strategy.Source: Table 2 · p. 7, §4 · p. 8
- Removing dynamic search lowers factual accuracy from 7.0 to 5.9 and analytical depth from 7.9 to 5.7.The authors state that acquiring additional knowledge during analysis and writing is necessary for comprehensive and factually correct reports.Source: Table 2 · p. 7, §4 · p. 8
4.1 ANALYSIS
- Each Chain-of-Analysis averages 2,761 tokens and 5.3 images; a report synthesizes about 17.6 CoAs into 62,586 tokens and 51.2 images.The process averages 983.2 searches and 469.8 browsed pages per report.Source: Table 3 · p. 8, §4.1 · p. 8
- FinSight reports cluster in Figure 4’s top-right quadrant, typically exceeding 20,000 words while receiving high quality scores.OpenAI Deep Research and Perplexity Deep Research show wider scatter, and greater length does not reliably correspond to higher quality for them.Source: Figure 4 · p. 8, §4.1 · p. 8
- Across three iterations, VLM feedback refines a simple stock plot into a dual-axis price-and-volume chart with analytical overlays and contextual event markers.The feedback targets aesthetics, information density, and other aspects of the visualization.Source: §4.1 · p. 8, Figure 5 · p. 9
5 RELATED WORK
The related work describes deep-research systems as advancing information synthesis while remaining limited in multimodal processing and financial-domain adaptation. Financial report-generation agents also have reported gaps in depth, data breadth, and multimodal integration.
5.1 DEEP RESEARCH SYSTEMS
- Deep-research frameworks use iterative reasoning or agent collaboration, but existing systems remain limited in multimodal processing and financial-domain adaptation (Yang et al., 2025).The discussion also notes text-centric workflows and limited real-time market-data integration.Source: §5.1 · p. 9
5.2 LLM AGENTS IN FINANCIAL DOMAIN
- Financial report-generation systems remain constrained in depth, data breadth, and multimodal integration; FinTeam’s single-round process is described as lacking depth and comprehensiveness (Wu et al., 2025).The passage also distinguishes tools and data interfaces from systems with designed report-generation frameworks.Source: §5.2 · p. 9
6 CONCLUSION
FinSight’s experiments reportedly outperform existing baselines across financial research tasks, while the conclusion highlights its code-centric architecture and two report-generation mechanisms. The authors also describe LLM use as assistance for manuscript and implementation work, with human authors responsible for the research and final content.
6 CONCLUSION
- FinSight significantly outperforms existing baselines across a range of financial research tasks.The conclusion describes the approach as having potential to advance automated financial analysis.Source: §6 · p. 10
- FinSight combines dynamic, code-driven analysis with iterative chart refinement and a two-stage framework for expanding analysis chains into reports.The conclusion presents these components as part of its multi-agent, code-centric approach.Source: §6 · p. 10
A STATEMENT ON THE USE OF LARGE LANGUAGE MODELS (LLMS)
- LLMs assisted with manuscript clarity, paraphrasing, style refinement, and code implementation during preparation of the manuscript.The authors state that core research concepts, experimental design, data analysis, and conclusions were developed exclusively by the human authors; they checked and substantially edited LLM-generated content and take responsibility for the final paper.Source: A STATEMENT ON THE USE OF LARGE… · p. 14
B BASELINES DETAILS
The baseline comparison includes LLMs paired with search tools and commercial deep research agents, evaluated through official web applications using a shared report prompt.
B BASELINES DETAILS
- The authors organize comparison systems into LLMs with search tools and deep research agents.These are the two baseline types specified for the comparison.Source: B BASELINES DETAILS · p. 14
- The comparison covers GPT-5, Claude-4.1-Sonnet, and DeepSeek-R1 paired with search tools, plus four commercial deep research agents.The named agents are Grok Deep Search, Perplexity Deep Research, OpenAI Deep Research, and Gemini-2.5-Pro Deep Research.Source: B BASELINES DETAILS · p. 14
PROMPT
- Baselines were evaluated directly on their official web applications using a unified prompt for detailed corporate-finance reports with text, charts, and standardized citations.The prompt requests serial numbers and corresponding references at the end of each report.Source: B BASELINES DETAILS · p. 14, PROMPT · p. 14
C IMPLEMENTATION DETAILS OF FINSIGHT
Implementation details specify the model backbones, collection and retrieval components, and a three-iteration chart-critique process. The section also lists three ablation variants that remove individual mechanisms or strategies.
C IMPLEMENTATION DETAILS OF FINSIGHT
- The multi-source collection, deep-search, and data-analysis agents use DeepSeek-V3; report generation uses DeepSeek-R1.The maximum input and output lengths are 81,920 and 16,384 tokens, respectively.Source: C IMPLEMENTATION DETAILS OF FINSIGHT · p. 14
- Financial APIs use akshare 2, while web search uses Google Search API with China-region settings and the top 10 results; Playwright extracts webpage content.These are the specified data-collection components and settings.Source: C IMPLEMENTATION DETAILS OF FINSIGHT · p. 15
- Qwen3-Embedding-0.6B generates embeddings for data and Chain-of-Analysis segments, and cosine similarity selects relevant segments for each report section.Source: C IMPLEMENTATION DETAILS OF FINSIGHT · p. 15
- Chart generation uses Qwen2.5-VL-72B as a vision-language critic for three refinement iterations.The authors state that three iterations balance effectiveness and cost.Source: C IMPLEMENTATION DETAILS OF FINSIGHT · p. 15
- The ablation study uses five company questions: Cambricon Technologies, Li Auto-W, Pop Mart, 3SBio, and China Mobile.Source: C IMPLEMENTATION DETAILS OF FINSIGHT · p. 15
- The ablations remove chart iteration, the two-stage writing framework, or dynamic search enhancement from data collection and report generation.Without the writing framework, the variant concatenates Chain-of-Analysis segments directly into the report.Source: C IMPLEMENTATION DETAILS OF FINSIGHT · p. 15
D CONSTRUCTION OF THE FINANCIAL REPORT GENERATION BENCHMARK
The benchmark uses company- and industry-level financial research questions across multiple markets and sectors, with securities-firm reports serving as golden references.
D CONSTRUCTION OF THE FINANCIAL REPORT GENERATION BENCHMARK
- The benchmark questions cover five popular A-share companies, five Hong Kong-stock companies, and ten representative industries.The companies and industries are described as spanning diverse market sectors.Source: D CONSTRUCTION OF THE FINANCIAL REPORT… · p. 15
- Golden reference reports are collected from well-known Chinese securities firms for company and industry analyses across A-shares, Hong Kong stocks, and major industries.The references are the latest equity and industry research reports, as shown in Table 4.Source: D CONSTRUCTION OF THE FINANCIAL REPORT… · p. 15
E EVALUATION AND METRICS
The section defines report-quality metrics and describes a list-wise evaluation process using Gemini-2.5-Pro. It also specifies benchmark-based scoring instructions for evaluations with and without a golden reference.
E EVALUATION AND METRICS
- The metrics cover factual quality, analysis effectiveness, and presentation quality, including report structure, professional language, chart expressiveness, faithfulness, coherence, information richness, coverage, and analytical insight.Factual metrics include core-conclusion consistency, textual faithfulness, and text-image coherence; analysis-effectiveness metrics include information richness, coverage, and analytical insight.Source: E EVALUATION AND METRICS · p. 15
- Evaluation uses Gemini-2.5-Pro to score candidate reports list-wise alongside a golden reference, with separate prompts for metrics related and unrelated to that reference.The section also identifies Table 4 as golden referenced reports from Chinese securities firms.Source: E EVALUATION AND METRICS · p. 16, Table 4 · p. 16
Evaluation Instruction for Golden Report Relevant Metrics
- For golden-report-relevant metrics, the instructions compare each generated report with a benchmark across three dimensions and require a score and concise rationale for each.The benchmark is the standard for 7; higher scores require demonstrably better performance, while lower scores indicate a shortfall.Source: Evaluation Instruction for Golden… · pp. 17–18
- The golden-report-relevant criteria include core conclusion and data consistency, information coverage, and professional language and tone.The benchmark is defined as a high-quality professional report, and report IDs use generated-report filenames.Source: Evaluation Instruction for Golden… · p. 18
F A CASE OF COMPANY RESEARCH QUESTION
The case study presents SenseTime Technology as an example of FinSight’s company research and summarizes its business mix, technical position, and prospects. It also points to data-collection and analysis task examples in Tables 5 and 6.
SenseTime case
- Tables 5 and 6 are identified as examples of data-collection tasks and the data-analysis process with a Chain-of-Analysis output.The case section also introduces SenseTime Technology as the practical application example.Source: F A CASE OF COMPANY RESEARCH QUESTION · p. 19, Table 5 · p. 19, Table 6 · p. 20
- SenseTime’s 2024 revenue mix comprises Smart Business at 35%, Smart City at 30%, Smart Life at 20%, and Smart Auto at 15%.Smart Business is identified as the main revenue driver, while Smart Auto is described as a rapidly growing new growth engine.Source: F A CASE OF COMPANY RESEARCH QUESTION · p. 21
- The case describes SenseTime’s technical strengths as full-stack AI capabilities, more than 8,000 AI-related patents by 2024, and advantages in originality, platformization, ecosystem breadth, and talent.SenseCore is named as the platform for AI industrial-scale production.Source: F A CASE OF COMPANY RESEARCH QUESTION · p. 21
- SenseTime is described as a global computer-vision leader and a 2024 AI-maturity Leaders Quadrant company, with future directions including vertical applications and international expansion.Other listed directions are large-model enhancement and convergence of AI with IoT and blockchain.Source: F A CASE OF COMPANY RESEARCH QUESTION · p. 21
- The conclusion characterizes SenseTime’s development as a shift from computer-vision provider to AI platform enterprise and describes its competitive position as favorable.It attributes the trajectory to technology, product iteration, platform building, and ecosystem expansion.Source: F A CASE OF COMPANY RESEARCH QUESTION · p. 21
- The case says SenseTime must sustain R&D investment, deepen vertical applications, and strengthen international collaboration amid market and technological challenges.These actions are presented as forward-looking priorities for maintaining competitiveness.Source: F A CASE OF COMPANY RESEARCH QUESTION · p. 21
G REPORT GALLERY
The report gallery presents excerpts from company and industry reports generated by FinSight.
Report gallery
- Figures 6–9 show excerpts from reports on SenseTime, the semiconductor industry, 3SBio, and China Mobile.The complete report is stated to be available from the project link.Source: Figure 6 · p. 22, G REPORT GALLERY · p. 22, Figure 7 · p. 23