Scholarlink

In the field of finance, understanding the abundant information encapsulated in comprehensive financial materials is fundamental for analysts to make well-informed decisions and optimize strategies. However, it demands a profound understanding of the financial intricacies. In this study, we introduce FORA, an efficient Financial One-foR-All cross-task reasoning framework via large language models (LLMs). To optimize information processing, FORA begins by refining raw data, generating a clarified and context-rich analysis via an information refinement module. Subsequently, FORA utilizes an inference enhancement module to seamlessly integrate information from various sources, including the refined signals, facilitating the accurate mapping from input to output. By leveraging the powerful capabilities of LLMs, these modules can efficiently adapt to diverse task types with minimal data requirements and zero training costs. FORA is tested across four representative tasks, demonstrating an average performance increase of 6.33%. In light of this and our subsequent analysis, we argue that FORA represents a significant stride forward in the exploration of advanced cross-task financial reasoning frameworks.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FORA: An Efficient One-for-All Cross-Task Reasoning Framework for Financial Domain via LLMs

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Zhichao Duan,
  • Tengyu Pan,
  • Zhenyu Li,
  • Bowen Dong,
  • Xiuxing Li,
  • Jianyong Wang

摘要

In the field of finance, understanding the abundant information encapsulated in comprehensive financial materials is fundamental for analysts to make well-informed decisions and optimize strategies. However, it demands a profound understanding of the financial intricacies. In this study, we introduce FORA, an efficient Financial One-foR-All cross-task reasoning framework via large language models (LLMs). To optimize information processing, FORA begins by refining raw data, generating a clarified and context-rich analysis via an information refinement module. Subsequently, FORA utilizes an inference enhancement module to seamlessly integrate information from various sources, including the refined signals, facilitating the accurate mapping from input to output. By leveraging the powerful capabilities of LLMs, these modules can efficiently adapt to diverse task types with minimal data requirements and zero training costs. FORA is tested across four representative tasks, demonstrating an average performance increase of 6.33%. In light of this and our subsequent analysis, we argue that FORA represents a significant stride forward in the exploration of advanced cross-task financial reasoning frameworks.