Source-linked AI summary
FinNextAssist: Towards Professional Financial Deep Research Assistant
Xiangyu Li, Fengbin Zhu, Xuan Yao, Siyu Liu, Xiaoluan Liu, Chao Wang, Huanbo Luan, Xiaofen Xing, Xiangmin Xu, Ke-Wei Huang, Richang Hong, Tat-Seng Chua
TL;DR
Financial analysis requires Deep Research agents to jointly handle heterogeneous sub-tasks with different data types and correctness requirements. FinNextAssist combines a four-stage pipeline with specialized financial skills and sub-agents; it outperforms strong proprietary and open-source agents across three benchmarks.
Problem
Financial analysis combines heterogeneous sub-tasks with distinct correctness requirements, while standard Deep Research pipelines flatten different evidence types and leave models to infer type boundaries and verification rules.
Source: §1 · p. 2Method
FinNextAssist uses a four-stage research pipeline with typed evidence, specialized financial skills, and TabAgent and HeteroAgent for table understanding and heterogeneous data interpretation.
Source: §1 · p. 2Results
FinNextAssist outperforms strong proprietary and open-source agents across three benchmarks; on FinDeepResearch it scores 48.5, 10.5 points above the strongest baseline.
Source: §1 · p. 2Takeaways & Limitations
Typed evidence, specialized sub-agents, and executable financial skills keep heterogeneous reasoning requirements from being conflated in a single undifferentiated context.
Source: §5 · p. 10Sub-agents and skills are curated manually; automatically identifying them from query characteristics remains an open direction.
Source: §5 · p. 10
Abstract
from arXiv · showhide
Deep Research (DR) agents have demonstrated strong capabilities in complex, research-oriented tasks through autonomous planning, iterative retrieval, multi-step reasoning, and structured reporting. However, adapting DR agents to finance introduces unique challenges: financial analysis demands the joint completion of heterogeneous sub-tasks spanning diverse data types, tools, and analytical workflows. We identify three key requirements for a professional financial DR agent: integration of authoritative, heterogeneous financial data sources; specialized analytical tools and skills; and dedicated sub-agents for domain-specific sub-tasks. Building on these principles, we propose FinNextAssist, an end-to-end deep research framework designed for professional financial analysis. FinNextAssist decomposes the research process into four stages: Task Planner, Evidence Compiler, Reasoning Engine, and Report Assembler, and introduces two novel lightweight sub-agents: TabAgent, for cross-market financial table understanding, and HeteroAgent, for cross-modality heterogeneous financial data interpretation. Extensive experiments on FinDeepResearch, the Finance Agent Benchmark, and FinTMMBench-Web show that FinNextAssist substantially outperforms both strong proprietary and open-source DR agents, with ablation studies confirming the contribution of each component across diverse markets and languages.
1 Introduction
Financial analysis requires Deep Research agents to coordinate heterogeneous sub-tasks with different data types and correctness requirements. FinNextAssist addresses this with an integrated pipeline and specialized sub-agents, and reports gains over proprietary and open-source agents across three benchmarks.
1 Introduction
- Financial analysis combines tasks such as filing retrieval, table interpretation, ratio derivation, price-dynamics analysis, and event-reaction interpretation, each with distinct correctness requirements.Ratios require exact numerical consistency, event analysis requires precise temporal alignment, and qualitative assessment requires grounded semantic synthesis.Source: §1 · pp. 1–2
- The authors identify three requirements: authoritative heterogeneous data sources, specialized financial tools and skills, and dedicated sub-agents for domain-specific tasks.Examples include errors from inaccurate web snippets, currency mismatches in P/E calculations, and the challenge of interpreting dense financial tables.Source: §1 · p. 2
- FinNextAssist uses a four-stage pipeline—Task Planner, Evidence Compiler, Reasoning Engine, and Report Assembler—with TabAgent for financial tables and HeteroAgent for heterogeneous data interpretation.It combines curated financial sources and specialized skills with the two sub-agents in an end-to-end framework.Source: §1 · p. 2
- 48.5 overall score on FinDeepResearch—10.5 points above the strongest baseline—plus 77.39% accuracy on Finance Agent Benchmark, 20.68 points above the strongest baseline, and 25.87 EM above the best Web-based baseline.FinNextAssist also ranks first in all eight evaluated markets.Source: §1 · p. 2
- The benchmarks show improvements across diverse markets and languages, while ablations support contributions from TabAgent and HeteroAgent.The reported gains span settings requiring structured-statement comprehension, precise numerical computation, cross-source reconciliation, and temporal alignment.Source: §1 · p. 2
2 FinNextAssist Framework
FinNextAssist frames financial analysis as producing evidence-backed report items that satisfy item-specific correctness predicates. Its four-stage pipeline plans typed constraints, compiles evidence, reasons with specialized support, and assembles a verified report.
2.1 Problem Formulation
- Each report item pairs an analytical value with provenance evidence and must satisfy its own correctness predicate.The predicates can require exact numerical matches, formula verification, temporal alignment, or source attribution.Source: §2.1 · p. 2
2.2 Overview
- The four-stage pipeline assigns typed constraints, compiles evidence, reasons over each item with specialized support, and verifies the assembled report.The Report Assembler returns items failing provenance or consistency checks for up to two revision rounds, then marks any still-unverified item unsupported.Source: §2.2 · p. 3, §2.6 · p. 5
2.3 Task Planner
- The Task Planner assigns each item a typed constraint and adapts accounting standard, currency, and fiscal year cycle to the target market.The constraint specifies the analytical dimension, evidence type, data source, correctness predicate, and output language.Source: §2.3 · pp. 3–4
2.4 Evidence Compiler
- The Evidence Compiler retrieves heterogeneous sources, parses and types their contents, and repeats retrieval when required evidence coverage is insufficient.It stops when coverage reaches a threshold or the maximum number of iterations is reached.Source: §2.4 · p. 4
2.5 Reasoning Engine
- HeteroAgent aligns heterogeneous sources on a timeline, resolves conflicts by prioritizing more authoritative sources, and maps events to trading dates and event windows.Its outputs include market metrics over specified windows, aligned event windows, and news-derived sentiment signals.Source: §2.5 · pp. 4–5
- TabAgent and HeteroAgent propose typed intermediate evidence rather than final values, leaving reasoning and synthesis to the backbone model.The sub-agents can be extended to support new evidence types without modifying the core Reasoning Engine.Source: §2.5 · p. 5
- Financial Skills compute exact numerical results outside free-form generation, while TabAgent normalizes financial tables and provides values with source locations and accounting-convention tags.TabAgent maps heterogeneous line-item names to a canonical schema and aligns units, signs, and currencies.Source: §2.5 · pp. 4–5
3.1 Experimental Setup
The experiments use three financial benchmarks spanning full-report analysis, filing-based question answering, and web-retrieval multimodal reasoning. The setup compares model and deep-research baselines and follows benchmark-specific evaluation protocols.
3.1.1 Benchmarks.
- FinDeepResearch evaluates full corporate-analysis reports for 64 companies across 8 markets and 4 languages, using 15,808 grading items across 6 report sections and 4 capability levels.The levels are Recognition, Calculation, Abstraction, and Interpretation.Source: §3.1.1 · p. 5
- Finance Agent Benchmark tests SEC-filing question answering, while FinTMMBench-Web uses 5,676 questions on tables, news, prices, and charts with evidence retrieved from the Web.The benchmarks cover filing retrieval and synthesis as well as temporal multimodal reasoning under Web retrieval.Source: §3.1.1 · p. 5
3.1.2 Baselines.
- FinDeepResearch compares 23 baselines across Thinking, Thinking+Search, and Deep Research settings; the other benchmarks use selected model baselines and, for FinTMMBench-Web, published original-benchmark references.The original FinTMMBench published baselines use its fixed local corpus, unlike the new Web-retrieval baselines.Source: §3.1.2 · p. 5
3.1.3 Implementation and Evaluation.
- The Reasoning Engine uses Claude-Sonnet-4.5, while TabAgent and HeteroAgent use Qwen-2.5-7B-Instruct; evidence comes from filings, market-data tools, and Web retrieval.FinDeepResearch reports are graded by its official service, Finance Agent Benchmark uses its official harness, and FinTMMBench-Web reports EM, F1, Accuracy, and LLM-judge Accuracy.Source: §3.1.3 · pp. 5–6
3.2 Main Results
FinNextAssist leads across three financial research benchmarks with different task formats. Its results are strongest in settings involving structured financial evidence, numerical computation, and retrieval.
3.2 Main Results
- Among 24 evaluated methods on FinDeepResearch, FinNextAssist achieves an Information Precision score of 48.5, outperforming the strongest baseline, Gemini 3 Pro Deep Research (38.0), by 10.5 points.Source: §3.2.1 · p. 6
3.2.2 Results on Finance Agent Benchmark.
- 77.39% accuracy exceeds Claude-Sonnet-4.5’s 56.71% by 20.68 points on the Finance Agent Benchmark.The comparison also reports the same model as FinNextAssist’s backbone, with the passage attributing the gain to the framework design rather than model capacity.Source: §3.2.2 · p. 6
3.2.3 Results on FinTMMBench-Web.
- FinNextAssist exceeds the best Web baseline by 25.87 EM, 41.26 F1, 33.91 Accuracy, and 30.02 LLM-judge Accuracy points.Table 2 reports results under Local and Web retrieval settings; local-corpus results are not directly comparable to Web results.Source: §3.2.3 · p. 6, Table 2 · p. 6
- F1 and Accuracy exceed EM by 19.33 and 14.50 points, respectively, indicating content matches can exceed exact reference-string matches.The typed evidence interface constrains grounding, not answer form.Source: §3.2.3 · p. 6
3.2 Main Results
- FinNextAssist leads on benchmarks spanning full-length reports, SEC-filing question answering, and temporal multimodal reasoning under Web retrieval.This range of formats supports the authors’ conclusion that its typed evidence interface transfers across task granularities and evidence types.Source: §3.2.3 · p. 7
3.3 In-depth Analysis
FinNextAssist’s strongest section and capability gains occur where structured evidence, calculation, and temporal alignment matter. Its market results are strongest in non-English settings, while higher-level interpretation remains difficult.
3.3.1 Performance Analysis Across Sections.
- FinNextAssist ranks first in three of six sections; it gains 16.3 points in Financial Performance and 6.8 in Market Performance over the named baselines.It scores 77.0 versus Gemini 3 Pro Deep Research’s 60.7 in Financial Performance, and 24.6 versus 17.8 in Market Performance; the latter is a 38% relative improvement.Source: §3.3.1 · p. 7, Table 3 · p. 8
3.3.2 Performance Analysis Across Capabilities.
- Recognition reaches 76.1 (+12.6 over Gemini 3 Pro Deep Research) and Calculation 75.0 (+30.1), the largest single-capability improvement.Abstraction scores 42.9, below OpenAI o3-deep-research’s 44.8; Interpretation reaches 20.9, the highest score, while remaining the least-solved capability.Source: §3.3.2 · pp. 7–8
3.3.3 Performance Analysis Across Markets.
- FinNextAssist ranks first in all eight markets, with the largest gains in China (+14.4) and Hong Kong (+11.9).Gains are positive across English-language and Southeast Asian markets, and its scores span a narrower range (44.9–51.7) than the best baseline scores (34.9–46.7).Source: §3.3.3 · p. 8, Table 4
3.4 Ablation Study
Removing either sub-agent reduces overall performance, while the backbone comparison shows that the framework’s results persist across models. Recognition is stable across backbones, but Calculation is more sensitive.
3.4.1 Effect Analysis of each Sub-agent.
- Removing TabAgent lowers the overall score from 48.5 to 44.2, removing HeteroAgent to 46.0, and removing both to 42.7.TabAgent’s largest capability costs are 10.0 points in Recognition and 8.6 in Calculation; HeteroAgent reduces Calculation by 13.9 points while barely changing Recognition.Source: §3.4.1 · pp. 8–9, Table 5 · p. 9
- With both agents, Calculation improves 20.0 points over the no-agent baseline, 2.5 points more than the summed gains from enabling either agent alone.Abstraction and Interpretation vary only marginally across the four settings.Source: §3.4.1 · p. 9, Table 5 · p. 9
3.4.2 Analysis of Different LLM Backbones.
- Across backbones, overall scores are 48.5 for Claude-Sonnet-4.5, 47.1 for DeepSeek-v3.2, and 46.8 for GPT-4o-mini; even the weakest exceeds the strongest baseline by 8.8 points.Recognition remains stable, while Calculation falls from 75.0 to 68.9 and 67.5; Abstraction declines modestly and Interpretation stays between 20.3 and 20.9.Source: §3.4.2 · p. 9, Table 6 · p. 9
4 Related Work
Prior deep-research systems combine planning, retrieval, and reasoning but often leave evidence organization to a unified context. Financial language and agent systems have expanded domain capabilities, while FinNextAssist separates evidence by correctness requirements.
4.1 Deep Research Agents.
- Deep Research agents evolved from agentic paradigms and measurable environments to commercial and open-source systems, but often leave evidence boundaries and verification rules implicit.FinNextAssist instead organizes retrieved materials into typed intermediate objects with predicate-specific constraints.Source: §4.1 · p. 9
4.2 LLM-based Financial Systems.
- Financial models and agents span language understanding, report generation, forecasting, research, and trading, yet often reason over heterogeneous evidence in a unified process.FinNextAssist distinguishes itself by separating evidence according to correctness requirements and using typed stages, specialized components, and deterministic operators.Source: §4.2 · p. 9
5 Conclusion
FinNextAssist combines authoritative heterogeneous financial data, specialized tools and skills, and domain-specific sub-agents in a four-stage research pipeline. Experiments on three financial research benchmarks support the design, while automatic routing and self-evolving sub-agent training remain open directions.
5 Conclusion
- FinNextAssist organizes financial deep research into Task Planner, Evidence Compiler, Reasoning Engine, and Report Assembler stages, with TabAgent and HeteroAgent supplying typed evidence to the backbone.TabAgent handles cross-market financial table understanding; HeteroAgent interprets cross-modality heterogeneous financial data.Source: §5 · p. 9
- Eight executable financial skills keep numerical operations off the generation path, and experiments on three financial research benchmarks confirm the design’s effectiveness and generalizability.The framework is designed to keep heterogeneous reasoning requirements from being conflated in a single undifferentiated context.Source: §5 · p. 10
- Two open directions are automatically selecting sub-agents and skills from query characteristics and training sub-agents end-to-end with self-evolving adaptation.The proposed adaptation would use domain-specific feedback to fine-tune sub-agents and experience to refine planning and routing.Source: §5 · p. 10
6 Ethical Considerations
The ethical concern is that financial analysis may present unverifiable claims with the confidence of verified ones, potentially affecting consequential decisions. FinNextAssist uses evidence provenance, executable numerical skills, and unsupported-item declarations; its data use is limited to public sources and firm-level financial information.
6 Ethical Considerations
- Unverifiable financial claims presented as verified can propagate into decisions with material consequences.The paper identifies this as the principal risk of automated financial analysis.Source: §6 · p. 10
- FinNextAssist attaches a provenance chain to every reported item, calculates numerical results with executable skills, and declares claims unsupported when available evidence cannot substantiate them.Source: §6 · p. 10
- The framework uses public corporate disclosures, market records, and benchmark corpora under their respective terms, retrieves only firm-level financial information, and processes no personal data.Source: §6 · p. 10