Agentar-Fin-R1 - A large-scale financial inference model launched by Ant Financial.

What is Agentar-Fin-R1?

Agentar-Fin-R1 is a large-scale language model developed by Ant Financial specifically for the financial sector, enhancing reasoning capabilities, credibility, and domain expertise in financial scenarios. Based on the Qwen3 foundation model, the model offers 8-bit and 32-bit parameter versions and is optimized through a refined financial task labeling system and a multi-layered credibility assurance framework. In data construction, a label-driven three-stage pipeline ensures the credibility of data sources, synthesis, and governance. The model performs exceptionally well on financial benchmarks (such as Fineva, FinEval, and FinanceIQ) and general reasoning tasks (such as MATH-500 and GPQA-diamond), demonstrating its superior performance and general reasoning capabilities in the financial domain.

Main functions of Agentar-Fin-R1

• Complex reasoning abilityAgentar-Fin-R1 can handle complex financial tasks involving multi-step analysis, risk assessment, and strategic planning.
• Decision supportThrough deep reasoning and data analysis, we provide financial institutions with precise decision support, helping them make wiser choices in complex and ever-changing financial markets.
• Intent recognitionIt accurately identifies users' intentions in financial scenarios, such as investment consultation, product inquiry, and risk assessment, and provides users with personalized services.
• Slot identification and information extractionIt can accurately identify and structure key information in financial texts, such as fund names, insurance products, and stock codes, providing a foundation for subsequent analysis and processing.
• Tool planning and recommendationRecommend suitable financial tools based on user needs, such as portfolio analysis tools and market comparison tools, to improve user experience and work efficiency.
• Expression generationGenerate accurate, reliable, and regulatory-compliant professional financial statements to ensure transparency and compliance of information.
• Security risk identificationTo identify and prevent security threats such as malicious input, data breaches, and system abuse, and to ensure the stable operation of the financial system.
• Compliance verificationWe have a deep understanding of and strict compliance with regulatory requirements such as anti-money laundering laws, data privacy protection, investor protection, and risk disclosure, ensuring that the model output complies with legal and ethical standards.

The technical principle of Agentar-Fin-R1

• A refined financial task labeling systemAgentar-Fin-R1 has built a sophisticated financial task labeling system that breaks down the financial field into several precisely defined categories, including different business scenarios (such as banking, securities, and insurance) and task types (such as intent recognition, slot recognition, and risk assessment). This guides data processing and training workflows, achieving systematic task-oriented optimization and ensuring comprehensive coverage of financial inference scenarios.
• Multi-dimensional credibility guaranteeTo ensure high data quality and credibility, Agentar-Fin-R1 employs a multi-dimensional credibility assurance framework:
  • The source is credibleData is obtained from authoritative financial institutions and regulatory documents, and its authenticity and relevance are ensured through knowledge engineering processes.
  • Synthesis TrustA multi-agent collaboration framework is introduced to generate high-quality synthetic data through mutual discussion and review among agents.
  • Trustworthy governanceData security and quality are ensured through manual sampling and labeling, deduplication, de-duplication, and filtering based on a self-developed reward model.
• Weighted training frameworkAgentar-Fin-R1 employs a dynamically weighted training framework, dynamically adjusting sample weights based on task difficulty. Specifically:
  • Difficulty perception weightingBy calculating the pass@k score for each task, the weights of the tasks are dynamically adjusted to ensure that the model invests more resources in complex tasks.
  • Exponential smoothing and lower bound clippingBy employing exponential smoothing and weight lower bound pruning, the stability and convergence of the training process are ensured.
• Two-stage training strategyAgentar-Fin-R1 employs a two-stage training strategy to balance the comprehensive infusion of financial knowledge with optimization for complex tasks.
  • Phase 1: By injecting financial knowledge through large-scale supervised fine-tuning (SFT), we ensure that the model has comprehensive knowledge of the financial field.
  • Phase TwoBy combining reinforcement learning (GRPO) and targeted fine-tuning, the model's performance on complex tasks can be further improved.
• Attribution loopAgentar-Fin-R1 introduces an attribution loop mechanism to optimize model performance through error attribution and targeted improvements.
  • Misattribution: Classify prediction errors using a two-dimensional labeling framework to identify performance gaps.
  • Dynamic resource allocationBased on performance gaps and learning efficiency, training resources are dynamically allocated to ensure continuous optimization of the model on key tasks.
• Innovative evaluation benchmarks FinovaTo comprehensively evaluate the model's performance in real-world financial scenarios, Agentar-Fin-R1 proposes a new evaluation benchmark. FinovaIt covers the following three key dimensions:
  • Intelligent agent capabilities: Assess core capabilities such as financial intent recognition, slot recognition, tool planning, and expression generation.
  • Complex reasoning abilityCombining financial mathematics, code understanding, and multi-step reasoning, it simulates real-world financial decision-making scenarios.
  • Safety and ComplianceThe evaluation model assesses its performance in security risk identification and regulatory compliance.
• High-efficiency data synthesis and validationAgentar-Fin-R1 employs a dual-track data synthesis strategy, combining task-oriented knowledge-guided generation and instruction evolution mechanisms to generate high-quality inference triples. 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

• Intelligent financial customer serviceThrough multi-round dialogue management, Agentar-Fin-R1 can continuously understand user needs and gradually guide users to complete complex financial operations, such as opening an account, transferring funds, and providing financial advice.
• Risk assessment and managementThe model can assess the risk level of an investment portfolio, provide risk warnings and management advice, and help investors make more informed decisions.
• Market Trend AnalysisAgentar-Fin-R1 can analyze market data, identify trends and patterns, and provide financial institutions with real-time analysis of market dynamics.
• Financial Statement AnalysisUsing natural language processing technology, Agentar-Fin-R1 can parse and analyze financial statements, providing detailed financial analysis reports to help analysts quickly obtain key information.
• Personalized recommendationsAgentar-Fin-R1 can recommend suitable financial products, such as funds, insurance, and wealth management products, based on users' historical data and preferences.