Source-linked AI summary
Fin-R1: A Large Language Model for Financial Reasoning through Reinforcement Learning
Zhaowei Liu, Xin Guo, Zhi Yang, Fangqi Lou, Lingfeng Zeng, Jinyi Niu, Mengping Li, Qi Qi, Zhiqiang Liu, Yiyang Han, Dongpo Cheng, Ronghao Chen, Huacan Wang, Xingdong Feng, Huixia Judy Wang, Chengchun Shi, Liwen Zhang
TL;DR
Financial LLMs remain limited by fragmented data, opaque reasoning, and weak transferability. Fin-R1 addresses these gaps with a financial CoT dataset and SFT-plus-GRPO training, achieving strong benchmark performance despite its compact scale. The paper also identifies future expansion toward financial multimodal data.
Problem
Financial LLMs face fragmented data, intransparent reasoning processes, and weak transferability to business applications.
Method
Fin-R1-Data is used to train Fin-R1 through supervised fine-tuning and GRPO reinforcement learning.
Results
85.0 on ConvFinQA and 76.0 on FinQA are reported for Fin-R1, which outperforms much larger LLMs.
Takeaways & Limitations
Fin-R1 provides a compact financial reasoning model trained with a high-quality financial CoT dataset and reports state-of-the-art performance.
Takeaways & Limitations
The authors plan to refine Fin-R1’s architecture to accommodate financial multimodal data.
Abstract
from arXiv · showhide
In recent years, general-purpose large language models (LLMs) such as GPT, Gemini, Claude, and DeepSeek have advanced at an unprecedented pace. Despite these achievements, their application to finance remains challenging, due to fragmented data sources, intransparent reasoning processes, and weak transferability to business applications. In response, we introduce Fin-R1, a reasoning LLM designed for financial scenarios. With a compact size of 7 billion parameters, Fin-R1 reduces deployment costs while addressing the aforementioned challenges. Its development follows a two-stage pipeline. First, we construct Fin-R1-Data, a high-quality financial dataset consisting of 60,091 chain-of-thought (CoT) samples, distilled and filtered from multiple authoritative benchmarks to ensure consistency and reliability. Second, we train Fin-R1 using Fin-R1-Data through supervised fine-tuning (SFT), followed by reinforcement learning (RL). This stage substantially improves the model's ability to solve complex financial reasoning tasks, yielding outputs that are both accurate and interpretable. Despite its relatively small parameter scale, Fin-R1 achieves competitive empirical performance across established financial benchmarks and demonstrates practical utility in compliance checking and robo-advisory. Our code is publicly available at https://github.com/SUFE-AIFLM-Lab/Fin-R1, and has already attracted over 700 stars.
1 Introduction
Financial LLMs face fragmented data, opaque reasoning, and weak transferability across high-stakes applications. Fin-R1 addresses these challenges through a financial reasoning dataset and post-training pipeline combining SFT with RL.
- Challenges: Financial data are fragmented across heterogeneous sources, making integration difficult and contributing to scarce high-quality financial CoT data.Contractual terms, regulatory requirements, macroeconomic indicators, and market signals may be inconsistent or contradictory.
- Challenges: Existing financial LLMs expose final outputs without transparent reasoning paths, conflicting with requirements for traceability and explainability.This restricts deployment in financial practice, where regulatory and compliance requirements are central.
- Challenges: High-stakes financial tasks challenge existing models’ transferability and generalization across evolving environments and scenarios.Models trained solely through SFT may rely on memorized examples rather than reasoning grounded in financial logic.
- Framework: Fin-R1-Data is a high-quality bilingual financial reasoning dataset with over 60,000 entries covering professional expertise, business practices, and numerical reasoning.DeepSeek-R1 generates reasoning traces, while Qwen2.5-72B-Instruct filters samples for logical consistency, coherence, and domain alignment.
- Framework: Fin-R1 applies SFT followed by GRPO reinforcement learning to Fin-R1-Data, enabling integrated reasoning across legal, economic, and quantitative domains.The framework uses explicit reasoning to support more reliable and interpretable outputs.
- Results: 75.2 average score on financial reasoning benchmarks placed Fin-R1 second overall, while its 7B-scale model surpassed existing same-scale state-of-the-art models by more than 17 points.The introduction reports nearly best performance on mainstream financial benchmarks such as FinQA.
2 Related Work
Related work spans pre-training, post-training reinforcement-learning methods, and financial LLMs. Existing financial models and general-purpose models still face challenges in domain-specific reasoning, accuracy, transparency, and transferability.
- Pre-training: Pre-training learns linguistic patterns and fundamental knowledge from massive text corpora, but requires substantial computational resources.Training PaLM with 540 billion parameters required 6,144 TPU v4 chips over several weeks.
- Pre-training: Pre-trained models often struggle to generalize to domain-specific reasoning tasks such as derivative pricing and risk-hedging strategies.The cited work identifies difficulty understanding complex financial logic as a consequence of this limitation.
- Post-training: Post-training methods include task-specific fine-tuning, reinforcement learning guided by feedback, and test-time scaling.These methods are presented as approaches for addressing limitations of pre-trained models.
- Post-training: PPO limits divergence between old and new policies but can suffer from instability, hyperparameter sensitivity, and high computational cost.Its computational burden includes learning a value function for the language model.
- Post-training: DPO removes value-function training and lowers computational cost, but remains dependent on the reference policy and has limited out-of-distribution generalization.DPO reformulates the reinforcement-learning problem as a classification task.
- Post-training: GRPO generates multiple candidate outputs and computes relative group-based advantages, eliminating value-network learning and improving computational efficiency.The related work describes advantages over PPO and DPO in reasoning-intensive tasks.
- Financial LLMs: Financial LLMs have expanded from text understanding toward reasoning and decision making, but accuracy, transparency, and transferability across financial scenarios remain open challenges.Existing financial models include BloombergGPT, DISC-FinLLM, PIXIU, XuanYuan, FinAgent, XuanYuan-FinX1-Preview, and Fino1.
3 Data Construction
Fin-R1-Data is a 60,091-entry bilingual financial dataset organized into four categories and constructed through distillation followed by filtering. Distillation generates reasoning paths and answers, while filtering checks answer correctness and reasoning quality before subsequent supervised fine-tuning.
- 3.1 Data Overview: Fin-R1-Data contains 60,091 bilingual Chinese-English entries organized into four categories.The four categories are financial advanced business knowledge, financial basic business knowledge, financial professional knowledge, and financial code.
- 3.1 Data Overview: Financial advanced business knowledge accounts for roughly 25% of the dataset and covers problems requiring explicit reasoning across numerical computation, causal inference, and contextual understanding.Representative tasks include numerical reasoning, sentiment classification, news categorization, and causal relationship extraction.
- 3.1 Data Overview: Financial basic business knowledge contributes over 50% of the dataset and covers regulatory compliance, domain knowledge acquisition, and financial text generation.This subset focuses on knowledge representation and content generation rather than complex reasoning.
- 3.1 Data Overview: Financial professional knowledge covers terminology explanations, specialized financial calculations, and postgraduate entrance-examination questions.The scarce examination-question component was supplemented with FinPEE, built from Shanghai University of Finance and Economics questions and expert-verified question-answer pairs.
- 3.1 Data Overview: Financial code contains scripts and programs for financial problems, including quantitative trading, risk modeling, and portfolio optimization.Its stated purpose is to enable automatic generation of financial code and quantitative-strategy scripts.
- 3.2 Data Processing: Data construction has two major steps: data distillation generates reasoning paths and answers from raw datasets, while data filtering reviews their correctness and quality.Figure 4 visualizes the workflow of these two steps.
- 3.2 Data Processing: Reasoning paths are filtered using seven criteria, and only high-quality paths are retained for subsequent supervised fine-tuning.The criteria include internal consistency, term overlap, reasoning-step count, logical coherence, diversity, domain relevance, and instruction alignment.
4 Model Training
Fin-R1 uses two-stage post-training: SFT teaches financial reasoning from structured traces, then GRPO refines outputs using format and accuracy rewards. The RL stage operates on objective questions with numerical solutions and updates the policy from group-relative rewards.
- Supervised Fine-tuning: SFT examples pair each query with an output containing a reasoning trace and an answer, often including a numerical solution.The trace and answer are enclosed in separate tags, and the objective minimizes cross-entropy on the output sequence.
- Supervised Fine-tuning: SFT guides the model toward correct reasoning and answer patterns but cannot learn behaviors absent from its demonstrations.The paper also notes that heterogeneous annotator styles may reduce SFT effectiveness.
- Group Relative Policy Optimization: GRPO uses objective-question and numerical-solution pairs, samples multiple candidate outputs, scores them, computes relative advantages, and updates the policy.Its reward design combines strict output formatting with semantic agreement between the candidate answer and the ground-truth solution.
- Group Relative Policy Optimization: The format reward requires a reasoning trace inside <think>...</think> and a concise answer inside <answer>...</answer>, with no extra content.A strictly matching output receives a score of 1; otherwise, it receives 0.
- Group Relative Policy Optimization: GRPO is described as statistically efficient as PPO while being considerably more computationally efficient.The method avoids computing value networks by using relative rewards within a group.
5 Experiment
Fin-R1 is evaluated on five financial datasets against eight language-model baselines using standardized sampling and an LLM-based answer judge. Despite its 7B scale, it performs strongly overall, leads FinQA and ConvFinQA, and benefits from the two-step training process.
- Evaluation Setup: Fin-R1 is evaluated across mainstream financial benchmarks despite having only 7B parameters.The evaluation reports datasets, methodology, baseline models, and comparative results.
- Evaluation Setup: The study uses five representative datasets and compares Fin-R1 with eight state-of-the-art language models spanning lightweight to large-scale architectures.For most objective datasets, 1,000 questions are sampled; Finance-Instruct-500k uses a stratified 10% test subset.
- Evaluation Methodology: The LLM-based judge scores each model on a 100-point scale reflecting the percentage of questions answered correctly, while accommodating valid numerical-format differences.The prompt template was tuned through numerical experiments to improve judge reliability.
- Results: 75.2 is Fin-R1’s average score, ranking second overall and placing it ahead of all similarly sized models.It trails DeepSeek-R1 at 78.2 by 3 points and exceeds DeepSeek-R1-Distill-Llama-70B at 69.2 by 6 points.
- Results: 76.0 on FinQA and 85.0 on ConvFinQA place Fin-R1 first among all competing models on both reasoning tasks.On the remaining tasks, it achieves comparable or better performance than baselines such as Qwen2.5-7B-Instruct.
- Ablation Study: GRPO alone improves the base model modestly but often produces incoherent outputs, whereas SFT alone achieves much higher scores across most datasets.The results support the effectiveness of combining SFT with GRPO rather than relying on either stage alone.
6 Conclusion
Fin-R1 addresses fragmented financial data, opaque reasoning, and weak business generalization through a dedicated dataset and SFT-plus-GRPO training pipeline. The compact model achieves strong benchmark performance and is presented as advancing LLM applications in finance.
- Fin-R1 is introduced to address fragmented financial data, intransparent reasoning, and weak business generalization in finance.
- Fin-R1-Data is a high-quality financial reasoning CoT dataset used to train the model with SFT and GRPO.
- Fin-R1 achieves state-of-the-art performance, outperforming much larger language models with over one-hundred-times more parameters.
- 85.0 and 76.0 are the best scores on the ConvFinQA and FinQA benchmark datasets, respectively.
- The authors state that Fin-R1 considerably advances the application of LLMs in finance.
- Future work will refine Fin-R1 for financial multimodal data and pursue deeper integration with risk management and regulatory compliance.
A Details for raw datasets
The raw-data appendix catalogs financial datasets by business scope, data composition, source characteristics, and task coverage. The listed resources span expertise, calculation, reasoning, sentiment, compliance, text generation, and financial coding.
- The appendix provides information on the sources of the datasets used in the experiments.
- Tables 3 and 4 classify financial data by business scope and detail source openness, language, data type, and original and processed sizes.
- Finance-500K covers professional financial content, including terminology explanation and financial-expertise question answering.
- FinanceIQ supports terminology explanation, financial-expertise question answering, and financial calculations.
- FinPEE is a financial-calculation dataset constructed from original postgraduate entrance-exam questions at Shanghai University of Finance and Economics.
- Ant-Finance covers non-complex financial business tasks such as content generation and compliance management, while FinCorpus supports financial knowledge acquisition and text generation.
- FinQA and ConvFinQA support financial reasoning and question-answering tasks, while TFNS supports sentiment analysis in financial news.
- FinCUGE includes numerical analysis and causal relationship extraction, and FinanceQT focuses on financial code and quantitative strategy scripts.
B The Prompt of Data Construction
The data-construction appendix specifies prompt templates for distillation and filtering. It defines structured inputs, evaluation rules, and binary selection criteria for producing and retaining reasoning data.
- B The Prompt of Data Construction: The appendix introduces prompt templates for data distillation and filtering in Fin-R1-Data construction.
- B.1 The Prompt of Data Distillation: Data distillation follows the official DeepSeek-R1 prompt setting and uses a template shown in Figure 10.
- B.1 The Prompt of Data Distillation: The distillation prompt specifies the task description, input, instruction, execution directive, and normative reasoning notes.
- B.2 The Prompt of Answer Check: The answer-check prompt frames the model as a scoring assistant and supplies both ground truth and model answer for comparison.
- B.2 The Prompt of Answer Check: Answer checking applies numerical-equivalence and rounding rules, then constrains the output to a binary decision.
- B.3 The Prompt of Reasoning Selection: Reasoning selection evaluates inputs, reasoning, and standard answers against criteria structured around seven key dimensions.
- B.3 The Prompt of Reasoning Selection: Each reasoning criterion is worth one point, and reasoning receives 1 only when all seven points are met; otherwise it receives 0.
C Model Selection for Data Construction
The appendix reports supplementary experiments intended to illustrate the model-selection process used during data construction.
- Supplementary experiments are provided to illustrate model selection for data construction.
- The experiments belong to the data-construction process rather than the main model-performance evaluation.
- The stated purpose of this section is illustration of the selection process, without a reported quantitative outcome in these passages.
C.1 Model Selection for Answer Check
The study evaluates how prompt strategies affect answer-checking models and selects Qwen2.5-72B-Instruct as the judge model based on accuracy, regularity, and human consistency.
- Experimental setup: Five prompt templates were tested using 100 randomly selected FinQA instances and five repeated experiments per strategy.The templates included OF, CIE, WQ, CIE-WQ, and ZH formats, producing 500 comparative results per prompt group.
- Evaluation metrics: Classification Inaccuracy measures disagreement with human annotations, while Format Irregularity measures violations of binary 0/1 output constraints.
- Prompt-strategy results: GPT-4o reaches 2.0% inaccuracy under CIE, whereas Qwen2.5-72B-Instruct adapts best to OF prompting.The supplied passage identifies Qwen2.5-72B-Instruct's OF accuracy as extremely high but does not provide its exact value.
- Prompt-strategy results: WQ causes 3.6% Irregularity for GPT-4o and 3.2% for Qwen2.5-72B-Instruct, with Qwen's inaccuracy rising from 0.4% to 8.0%.Manual verification attributes most format deviations to long-text samples and possible input-sequence-induced hallucinations.
- Cross-lingual results: Chinese prompts produce higher classification errors than optimal English strategies, although Qwen2.5-72B-Instruct adapts better to Chinese than GPT-4o.
- Model selection: Qwen2.5-72B-Instruct was selected as the judge model because the chosen prompt template provides high accuracy and regularity.
C.2 Model Selection for Reasoning Selection
The study compares language-model reasoning scores with human annotations using a seven-criterion binary framework and heatmap-based alignment analysis. Qwen2.5-72B-Instruct shows higher concordance with human judgments than GPT-4o and is selected as the scoring model.
- Alignment analysis: Heatmaps compare the scoring distributions of Qwen2.5-72B-Instruct, GPT-4o, and human annotations to assess alignment and discrepancies.
- Evaluation framework: Twenty filtered data points were evaluated by Qwen2.5-72B-Instruct and GPT-4o using seven predefined reasoning criteria.Each criterion received 1 when satisfied and 0 otherwise, yielding total scores from 0 to 7.
- Evaluation framework: Human annotators independently scored the same reasoning data to establish a reference baseline.
- Results: GPT-4o shows larger discrepancies from human judgments than Qwen2.5-72B-Instruct.
- Model selection: Qwen2.5-72B-Instruct was selected as the scoring model for reasoning evaluation.
D Case Study
A financial-securities case study compares Fin-R1 with Qwen2.5-7B-Instruct. Fin-R1 produces more focused, correct answers and higher-quality reasoning, while the base model struggles with the required reasoning-task format.
- Case-study setup: The case study uses an interactive financial-securities investment scenario to compare Fin-R1 with Qwen2.5-7B-Instruct.
- Base-model behavior: Without a specified output format, Qwen2.5-7B-Instruct fails to follow the required thought-first, then-answer structure.
- Base-model behavior: Qwen2.5-7B-Instruct remains suboptimal even when the task output format is specified.
- Fin-R1 results: Fin-R1 outperforms Qwen2.5-7B-Instruct in the case study.
- Fin-R1 results: Fin-R1 provides high-quality reasoning processes alongside answers that are correct and more focused.