Large Language Model Agents in Finance: A Survey Bridging Research, Practice, and Real-World Deployment

In one sentence

Summary

The paper addresses the disconnect between cutting-edge LLM capabilities and the stringent, interdependent workflows of real-world financial institutions. It argues that while LLMs show promise in automating tasks like parsing filings and gauging sentiment, they struggle with systemic obstacles such as static benchmarks, poor numerical reasoning, and lack of real-time adaptability. The authors propose a dual-perspective framework: a practitioner-centric taxonomy mapping five financial departments (Data Analysis, Investment Research, Trading, Investment Management, Risk Management) to specific agent roles, and a research-focused analysis of modeling challenges, including prompt sensitivity and multi-agent fragility. This holistic view contextualizes technical progress within institutional constraints like privacy and regulatory auditing. The methodology involves a systematic review of over 30 financial benchmarks and 20 representative LLM models, categorized by modality, task, and deployment limitations. The authors analyze datasets ranging from SEC filings and earnings calls to time-series price data, evaluating models like BloombergGPT, FinMA, and FinGPT. They detail the pipeline for each agent type, from data extraction and sentiment analysis to strategy execution and portfolio optimization. The study highlights specific limitations in current benchmarks, such as the lack of multi-asset coverage, real-time data integration, and consideration of transaction costs or liquidity constraints. It also examines multi-agent collaboration frameworks that mimic institutional workflows, such as investment committees and risk review processes. Key findings indicate that while LLMs excel in text summarization and entity recognition, they face significant hurdles in numerical precision, long-horizon logic, and dynamic market adaptation. The survey identifies open challenges including the need for continual adaptation, coordination-aware multi-agent systems, and privacy-compliant deployment. It emphasizes that current models often rely on centralized data and lack built-in regulatory auditing mechanisms. The paper concludes by advocating for deeper researcher-practitioner collaboration and transparent model architectures to enable safer, scalable AI adoption in finance, noting that most existing solutions are conceptual or limited to single-asset scenarios without real-world execution constraints.

Research problem

The paper addresses the gap between LLM research capabilities and the complex, regulated, and interdependent workflows of real-world financial institutions. It highlights systemic obstacles such as static benchmarks, poor numerical reasoning, lack of real-time adaptability, and insufficient attention to privacy and regulatory compliance. The core problem is how to align LLM agents with institutional demands across five key financial domains while ensuring safety, scalability, and interpretability.

Core contributions

• Proposes a dual-perspective framework bridging practitioner-centric financial workflows and research-focused LLM modeling challenges.
• Introduces a functional taxonomy mapping five core financial domains (Data Analysis, Investment Research, Trading, Investment Management, Risk Management) to specific agent roles and sub-tasks.
• Catalogs and compares over 30 financial benchmarks and 20 representative LLM models across modalities, tasks, and deployment limitations.
• Analyzes multi-agent collaboration frameworks that mirror institutional structures like investment committees and risk review processes.
• Identifies critical open challenges in numerical reasoning, real-time adaptability, and privacy-compliant deployment for financial AI.

Method

• Dual-perspective framework combining practitioner-centric taxonomy with research-focused model analysis.
• Systematic literature review of financial LLM applications, benchmarks, and models.
• Categorization of tasks into five financial domains: Data Analysis, Investment Research, Trading, Investment Management, Risk Management.
• Analysis of multi-agent collaboration mechanisms mimicking institutional workflows.
• Evaluation of model limitations regarding numerical reasoning, real-time adaptability, and regulatory compliance.
• Review of dataset characteristics including modality, size, and temporal coverage.

Results & findings

• LLMs show strong performance in text summarization and named entity recognition but struggle with numerical precision and long-horizon logic.
• Current benchmarks are often static, unimodal, and lack real-time data integration, limiting their applicability to dynamic markets.
• Multi-agent systems exhibit fragility under real-world stress and lack robust coordination mechanisms for conflict resolution.
• Privacy and compliance remain underexplored, with most pipelines relying on centralized data and lacking regulatory auditing.
• Models like FinMA and BloombergGPT demonstrate domain-specific adaptation but face limitations in online updates and scalability.
• Multi-agent frameworks like FinCon and TradingAgents show promise in simulating institutional workflows but suffer from prompt sensitivity.

Data & experiments

• Systematic review of over 30 financial benchmarks including ECT-Sum, FinQA, StockNet, and InvestorBench.
• Analysis of 20 representative LLM models such as BloombergGPT, FinMA, FinGPT, and FinTral.
• Evaluation of datasets covering text (SEC filings, earnings calls), tables, time series, and audio modalities.
• Assessment of model performance on tasks like text summarization, sentiment analysis, event classification, and strategy execution.
• Review of multi-agent systems including FinCon, TradingAgents, and Cryptoagents for collaborative decision-making.

Limitations & caveats

• Most benchmarks rely on static, archived data rather than real-time market feeds.
• Current models struggle with numerical precision and long-horizon logical reasoning.
• Multi-agent systems lack robust coordination mechanisms and are sensitive to prompt engineering.
• Privacy and regulatory compliance are underexplored, with limited auditing mechanisms.
• Many solutions are conceptual or limited to single-asset scenarios without real-world execution constraints.
• High computational overhead and energy consumption are significant barriers to deployment.

Curator notes

Discovery status: round1_promoted_seed_for_round2. Source links: A Multimodal Foundation Agent for Financial Trading: Tool-Augmented, Diversified, and Generalist; A Survey of Large Language Models for Financial Applications: Progress, Prospects and Challenges; A Survey of Large Language Models in Finance: FinLLMs; A Survey on Large Language Models for Critical Societal Domains: Finance, Healthcare, and Law; AlphaFin: Benchmarking Financial Analysis with Retrieval-Augmented Stock-Chain Framework; BloombergGPT: A Large Language Model for Finance; ConvFinQA: Exp