From General-Purpose Large Language Models to Domain-Specific Financial Models: A Structured Review and Comparative Case Analysis

Abstract

The use of large language models (LLMs) in finance has expanded rapidly, yet general-purpose systems face recurring problems in high-risk and regulated applications: Unsupported factual generation, weak reasoning across text, tables, charts, and numerical data, and limited traceability of model-assisted decisions. This study conducts a Preferred Reporting Items for Systematic Reviews and Meta Analyses (PRISMA)-informed structured review of 67 studies, principally published between January 2020 and March 2026, with three foundational pre-2020 studies retained through backward citation tracking. It develops a qualitative analytical and diagnostic framework with three pathways: Specialized knowledge modeling, task-constrained multimodal reasoning, and Trust-by-Design. A study-level evidence map and an explicit coding protocol show how the pathways were derived. The framework is used to compare BloombergGPT, FinGPT, FinTral, Fin-o1, Fin-R1, and Agentar-Fin-R1 through common criteria for knowledge controls, modality handling, lifecycle governance, evidence quality, and deployment maturity. The case analysis distinguishes model-paper design descriptions, developer reported benchmark results, platform-level deployment reports, and prospective applications. The analysis indicates that financial LLMs differ materially in how they organize domain knowledge, coordinate heterogeneous inputs, and govern model use across the lifecycle. Agentar-Fin-R1 illustrates the framework; the case does not validate the framework or establish independently observed production performance. The paper provides a structured basis for assessing domain-specific financial LLMs and connects responsible model design with sustainable financial innovation through durable governance arrangements, controlled model and compliance risk, responsible scaling, financial inclusion, and institutional resilience. Received: 1 April 2026 | Revised: 2 September 2026 | Accepted: 20 September 2026 Conflicts of Interest The authors declare that they have no conflicts of interest to this work. Data Availability Statement Tables S1–S4 in the Supplementary Materials provide the prospective replication protocol, the study-level evidence map for the 67-study corpus, the code-to-pathway derivation and threshold rationale, and the register of seven sources outside the review corpus. The original query-level exports, source-specific yields, and coder-level screening logs were not retained. Data sharing is not applicable to this article as no new data were created or analyzed in this study. Author Contribution Statement Fumin Niu: Conceptualization, Methodology, Investigation, Writing – original draft, Writing – review & editing, Visualization. Fengshuo Liu: Conceptualization, Investigation, Writing – original draft. Zekang Zhang: Methodology, Writing – original draft. Yishi Li: Writing – original draft. Yufei Ren: Supervision, Project administration.