FinTeam: Multi-Agent Financial Report System

• FinTeam is a collaborative financial multi-agent system that decomposes report generation into four specialized LLM agents.
• It employs a sequential pipeline—Document Analyzer, Analyst, Accountant, and Consultant—to process, analyze, compute, and synthesize financial data.
• Empirical evaluations demonstrate its enhanced performance in report quality, calculation accuracy, and overall user acceptance in financial scenarios.

Searching arXiv for the cited FinTeam papers to ground the article in the current arXiv record. FinTeam is a financial multi-agent collaborative system proposed for comprehensive financial scenarios, especially financial report generation tasks that range from macro- to micro-economics analysis and require extensive data analysis (Wu et al., 5 Jul 2025). It is motivated by two premises stated in the paper: existing LLM models are usually fine-tuned on simple QA tasks and cannot comprehensively analyze real financial scenarios, and financial companies often distribute tasks among departments. FinTeam operationalizes that analogy through four specialized LLM agents—document analyzer, analyst, accountant, and consultant—trained with specific financial expertise and arranged in a scenario-driven workflow that converts raw financial text and a user query into a professional report. The project is released at the repository specified by the authors (Wu et al., 5 Jul 2025).

1. Problem Setting and Design Rationale

FinTeam targets financial report generation in settings that combine qualitative interpretation, retrieval over heterogeneous knowledge bases, and quantitative computation. The paper evaluates these capabilities on comprehensive financial tasks constructed from real online investment forums, spanning macroeconomic, industry, and company analysis (Wu et al., 5 Jul 2025).

The system’s core design choice is role specialization. Rather than asking a single model to execute document parsing, retrieval, computation, and report writing in one pass, FinTeam decomposes the task into four narrower functions. This decomposition is explicit rather than latent: each agent consumes structured output from its predecessor, appends its own result to a shared buffer, and hands off a restricted intermediate representation. This suggests a workflow-oriented view of financial intelligence in which specialization is enforced at the interface level, not only through prompting.

2. Agent Topology and I/O Contracts

FinTeam arranges four specialized LLM agents in a pipelined architecture. The pipeline is:

User Query → [Document Analyzer] → DA_Output → [Analyst] → AN_Output → [Accountant] → AC_Output → [Consultant] → Final_Report (Wu et al., 5 Jul 2025).

Agent	Responsibility	Output
Document Analyzer (DA)	ingest unstructured financial text; intent recognition, entity extraction, sentiment analysis, keyword/key-sentence extraction	JSON
Analyst (AN)	retrieval-augmented reasoning over macroeconomic indicators, industry databases, and financial statements	Markdown steps
Accountant (AC)	accurate numeric computation via external tools	inline tool calls plus narrative
Consultant (CO)	integrate upstream outputs into a polished financial report	plain text, ~300–500 tokens

The Document Analyzer takes raw text of at most 4,096 tokens and returns JSON with fields for entities, sentiment, and keywords. Its prompt template is: “Analyze the following financial document. Extract entities, intent, and sentiment. Return JSON.” The Analyst receives the DA JSON plus the user question and emits Markdown steps organized as “Retrieval Query → Retrieved Passages → Analytical Answer.” Its prompting follows a Chain-of-Retrieval sequence: QGen, Ret, and Ans. The Accountant receives AN sub-answers that require quantitative work and produces tool-mediated outputs such as [Calculator(expression) → result], followed by integrated narrative. Its prompt is Toolformer-style: “If you need to compute, wrap your expression in Calculator(…) and pause.” The Consultant concatenates the DA, AN, and AC outputs with the user query and returns a final narrative report with sections such as Summary, Key Drivers, Risks, and Recommendations; the template is “1. Executive Summary; 2. Analysis Details; 3. Calculations Summary; 4. Conclusion & Advice.” (Wu et al., 5 Jul 2025)

A common misunderstanding is to treat this architecture as a standard MoE. The paper instead specifies a sequential multi-agent pipeline with explicit intermediate artifacts and append-only state. Routing occurs at the workflow level through ordered agent invocation, not through token-level gating.

3. Data Construction and Fine-Tuning Regime

The four agents are individually fine-tuned on task-specific datasets totaling approximately 241k samples (Wu et al., 5 Jul 2025). The training-set statistics reported in the paper are: Consultant, 63k samples from translated FiQA plus self-chat on JingGuan forum; Document Analyzer, 95k samples from open NLP corpora such as FPB, CCKS, and SmoothNLP plus ChatGPT-annotated financial reports; Accountant, 64k instrumented Q&A pairs from exam questions, BELLE School Math, and self-instruction via ChatGPT; Analyst, 20k CoR pairs generated by ChatGPT.

The preprocessing pipeline is also role-specific. Text is normalized to UTF-8, sentence-segmented, and tokenized with the Qwen tokenizer. For DA, 80k unlabeled reports are auto-annotated via ChatGPT and then sampled down to balance per task. For AN, knowledge bases are indexed with m3e-base embeddings. For AC, numeric formats such as commas and currency symbols are unified and placeholder tokens are injected for tool calls. For CO, FiQA is translated to Chinese and paired with forum dialogues (Wu et al., 5 Jul 2025).

All agents are LoRA-tuned on Qwen2.5-7B-Chat with Deepspeed ZeRO-0. The reported hyperparameters are batch size $= 1$ per GPU, gradient accumulation $= 4$, maximum sequence length $= 4{,}096$, epochs $= 2$, learning rate $\eta = 5 \times 10^{-5}$ with cosine decay, and LoRA rank $r = 8$, $\alpha = 16$, target_modules = "all" (Wu et al., 5 Jul 2025). The training objective for each agent is standard next-token cross-entropy. This training regime indicates that specialization in FinTeam is implemented through separate instruction distributions and separate fine-tuning targets rather than through a single jointly routed backbone.

4. Inter-Agent Collaboration and Conflict Resolution

FinTeam’s collaboration mechanism is defined through a shared “conversation buffer.” At turn $k$,

$$\text{buffer}_k = \text{buffer}_{k-1} \,\|\, \text{Agent}_k(\text{buffer}_{k-1}, \text{user\_query})$$

where each agent appends its output verbatim to the accumulated context (Wu et al., 5 Jul 2025). No global embedding sharing is needed. This is a strong architectural commitment: the system relies on textual state transfer rather than hidden-state fusion.

The paper also specifies a deterministic procedure for resolving redundant or conflicting numeric facts. Let $F = \{f_1, \ldots, f_m\}$ be numeric facts from upstream agents. The Consultant selects

$= 4$0

breaking ties by “latest timestamp first” (Wu et al., 5 Jul 2025). The mechanism is therefore a simple majority-vote followed by timestamp precedence, not a learned verifier.

The orchestration is summarized in the paper with the following pseudocode:

$= 4{,}096$1

This design has a clear methodological implication. Because the upstream outputs are appended rather than abstracted away, intermediate reasoning artifacts remain inspectable. For finance-oriented evaluation, this supports decomposition into document understanding, retrieval quality, calculation correctness, and final report quality.

5. Benchmarking and Empirical Results

FinTeam is benchmarked using automated LLM metrics, GPT-4o scoring, and human preference (Wu et al., 5 Jul 2025). At the agent level, the paper reports a 7.43-point average improvement on FinCUGE for the Document Analyzer. In the tabulated results, Sentiment (Acc) is 66.99 for FinTeam versus 66.19 for Qwen2.5-Instruct, for a gain of $= 4$1; Event QA (F1) is 47.44 versus 37.33, for a gain of $= 4$2; and the average is 47.20 versus 39.77, for a gain of $= 4$3. On FinEval, the Consultant achieves 68.48% versus 66.42% for Qwen2.5-Instruct, a $= 4$4 boost. For financial calculation accuracy, the Accountant scores 0.70 versus 0.61, a gain of $= 4$5 (Wu et al., 5 Jul 2025).

At the system level, GPT-4o scores four dimensions from 1 to 5—Accuracy, Thoroughness, Clarity, and Professionalism—and defines overall score as the mean of the four. FinTeam obtains Accuracy 4.54, Thoroughness 4.94, Clarity 4.84, Professionalism 4.96, and Overall 4.86. The corresponding overall scores are 4.83 for GPT-4o and 4.66 for Xuanyuan-13B. The paper further states that paired $= 4$6-tests confirm that the improvements in Thoroughness, Professionalism, and Overall are significant with $= 4$7 (Wu et al., 5 Jul 2025).

Human evaluation is framed as an acceptance-rate test. Undergraduates selected the best answer among five models for each of 150 queries, with

$= 4$8

FinTeam was chosen 93 times, corresponding to 62.00%, compared with 9.33% for Qwen2.5 and 19.33% for Xuanyuan (Wu et al., 5 Jul 2025). Within the paper’s evaluation design, this is the strongest evidence that the multi-agent composition improves perceived utility on comprehensive financial scenarios rather than only on isolated subtasks.

6. Scenario Walk-Throughs and Relation to FinTeamExperts

The paper illustrates FinTeam with three representative scenarios (Wu et al., 5 Jul 2025). In macroeconomic analysis, the query “How will rising U.S. interest rates affect global equity markets?” yields DA extraction of entities such as “U.S. interest rates” and “global equity markets,” neutral sentiment, AN retrieval over Fed dot-plot and historical equity returns, and AC computation [Calculator((1.02)^4 - 1) → 0.0824], interpreted as “8.24% annualized impact.” The Consultant then produces a report whose summary states that “Global equities tend to lag rate hikes by 3–6 months,” with recommendations including “Underweight rate-sensitive sectors; consider havens.” In industry analysis, the query on EV charging infrastructure leads AN to note 30% CAGR to 2028 and AC to produce [Calculator(0.30 * 5 – 1) → 0.5], summarized as “50% total growth over 5 years.” In company analysis, the Tesla example uses revenue $= 4$9B from the Tesla Q1 report, computes ROE as [Calculator(3.3/23 * 100) → 14.35]% and Debt/Equity as [Calculator(11.5/22) → 0.52], and synthesizes these with a SWOT outline.

FinTeam should also be distinguished from the related framework “FinTeamExperts” (Yu et al., 2024). FinTeamExperts is a role-specialized sparse Mixture-of-Experts for financial analysis with three decoder-only LLMs—Macro Analyst, Micro Analyst, and Quantitative Analyst—and token- or example-level routing via a gating function such as $= 4{,}096$0. It is trained in two phases: continued autoregressive pretraining on role-specific corpora and instruct-tuning with MoE routing. By contrast, FinTeam uses four LoRA-tuned agents in a pipelined workflow with a shared conversation buffer, explicit tool calls, and a final consultant layer (Yu et al., 2024). A common misconception is therefore to collapse the two systems under a single architectural label because of the shared naming scheme. The arXiv record instead presents two distinct approaches to role specialization in financial LLM systems: one based on sequential collaboration and one based on sparse expert routing.