SuperAPI AI Directory
Partner with us中文AI newsFavorites
DirectoryNews and tutorials

Agentar-Fin-R1 - A financial reasoning large model launched by Ant Group Digital Technologies

Author: AI小集 · Original source: AI工具集

What is Agentar-Fin-R1

Agentar-Fin-R1 is a large language model launched by Ant Group Digital Technologies specifically for the financial domain, enhancing reasoning capabilities, trustworthiness, and domain expertise in financial scenarios. The model is developed based on the Qwen3 foundation model and is available in 8B and 32B parameter versions, optimized through a refined financial task label system and a multi-layered trustworthiness assurance framework. In terms of data construction, a label-driven three-stage pipeline is adopted to ensure trustworthy data sources, trustworthy synthesis, and trustworthy governance. The model performs excellently on financial benchmarks (such as Fineva, FinEval, FinanceIQ) and general reasoning tasks (such as MATH-500, GPQA-diamond), demonstrating its outstanding performance in the financial domain and general reasoning capabilities.

Agentar-Fin-R1

Main features of Agentar-Fin-R1

  • Complex reasoning capability: Agentar-Fin-R1 can handle complex financial tasks involving multi-step analysis, risk assessment, and strategic planning.
  • Decision support: Through deep reasoning and data analysis, it provides precise decision support for financial institutions, helping them make wiser choices in complex and ever-changing financial markets.
  • Intent recognition: It accurately recognizes user intent in financial scenarios, such as investment consulting, product inquiries, risk assessment, etc., providing personalized services for users.
  • Slot recognition and information extraction: It can accurately identify and structure key information in financial texts, such as fund names, insurance products, stock codes, etc., providing a foundation for subsequent analysis and processing.
  • Tool planning and recommendation: It recommends suitable financial tools based on user needs, such as portfolio analysis tools, market comparison tools, etc., improving user experience and work efficiency.
  • Expression generation: It generates accurate, reliable, and regulatory-compliant professional financial expressions, ensuring information transparency and compliance.
  • Security risk identification: It identifies and prevents security threats such as malicious input, data leakage, and system abuse, ensuring the stable operation of financial systems.
  • Compliance verification: It deeply understands and strictly complies with regulatory requirements such as anti-money laundering regulations, data privacy protection, investor protection, and risk disclosure, ensuring that model outputs comply with legal and ethical standards.

Technical principles of Agentar-Fin-R1

  • Refined financial task label system: Agentar-Fin-R1 has built a refined financial task labeling system, decomposing the financial domain into multiple precisely defined categories, including different business scenarios (such as banking, securities, insurance, etc.) and task types (such as intent recognition, slot recognition, risk assessment, etc.). It guides data processing and training workflows, achieving systematic task-oriented optimization and ensuring comprehensive coverage of financial reasoning scenarios.
  • Multi-dimensional trustworthiness assurance: To ensure the high quality and trustworthiness of data, Agentar-Fin-R1 adopts a multi-dimensional trustworthiness assurance framework:
    • Trustworthy sources: Data is obtained from authoritative financial institutions and regulatory documents, and knowledge engineering processing is used to ensure the authenticity and relevance of the data.
    • Trustworthy synthesis: A multi-agent collaboration framework is introduced to generate high-quality synthetic data through mutual discussion and review among agents.
    • Trustworthy governance: Through manual sampling annotation, deduplication, detoxification, and filtering based on a self-developed reward model, ensure data safety and quality.
  • Weighted Training Framework: Agentar-Fin-R1 adopts a dynamic weighted training framework, dynamically adjusting sample weights according to task difficulty. Specifically:
    • Difficulty-Aware Weighting: By calculating the pass@k score for each task, dynamically adjust the task weights to ensure the model invests more resources in complex tasks.
    • Exponential Smoothing and Lower Bound Clipping: Through an exponential smoothing mechanism and weight lower bound clipping, ensure the stability and convergence of the training process.
  • Two-Stage Training Strategy: Agentar-Fin-R1 adopts a two-stage training strategy, balancing the comprehensive injection of financial knowledge and the optimization of complex tasks:
    • First Stage: Inject financial knowledge through large-scale supervised fine-tuning (SFT) to ensure the model has comprehensive financial domain knowledge.
    • Second Stage: Combine reinforcement learning (GRPO) and targeted fine-tuning to further improve the model's performance on complex tasks.
  • Attribution Loop: Agentar-Fin-R1 introduces an attribution loop mechanism to optimize model performance through error attribution and targeted improvements:
    • Error Attribution: Classify prediction errors through a two-dimensional label framework to identify performance weak spots.
    • Dynamic Resource Allocation: Dynamically allocate training resources based on performance gaps and learning efficiency to ensure continuous optimization of the model on key tasks.
  • Innovative Evaluation Benchmark Finova: To comprehensively evaluate the model's performance in real financial scenarios, Agentar-Fin-R1 proposes a new evaluation benchmark Finova, covering the following three key dimensions:
    • Agent Capabilities: Evaluate core capabilities such as financial intent recognition, slot recognition, tool planning, and expression generation.
    • Complex Reasoning Capabilities: Combine financial mathematics, code understanding, and multi-step reasoning to simulate real financial decision-making scenarios.
    • Safety and Compliance: Evaluate the model's performance in safety risk identification and regulatory compliance.
  • Efficient Data Synthesis and Verification: Agentar-Fin-R1 adopts a dual-track data synthesis strategy, combining task-oriented knowledge-guided generation and instruction evolution mechanisms to generate high-quality reasoning triples. Through multi-model consistency verification and manual sampling annotation, ensure the accuracy and reliability of the data.

Agentar-Fin-R1 Project Address

  • arXiv Technical Paper：https://arxiv.org/pdf/2507.16802

Application Scenarios of Agentar-Fin-R1

  • Financial Intelligent Customer Service: Through multi-turn dialogue management, Agentar-Fin-R1 can continuously understand user needs and gradually guide users to complete complex financial operations, such as account opening, transfers, wealth management consultation, etc.
  • Risk Assessment and Management: The model can assess the risk level of investment portfolios, provide risk warnings and management suggestions, and help investors make wiser decisions.
  • Market Trend Analysis: Agentar-Fin-R1 can analyze market data, identify trends and patterns, and provide financial institutions with real-time analysis of market dynamics.
  • Financial Statement Analysis: Through natural language processing technology, Agentar-Fin-R1 can parse and analyze financial statements, provide detailed financial analysis reports, and help analysts quickly obtain key information.
  • Personalized recommendations: Agentar-Fin-R1 can recommend suitable financial products based on the user's historical data and preferences, such as funds, insurance, wealth management products, etc.