PaperlayerView exact revision on arXiv

Source-linked AI summary

FinTrust: A Comprehensive Benchmark of Trustworthiness Evaluation in Finance Domain

Tiansheng Hu, Tongyan Hu, Liuyang Bai, Yilun Zhao, Arman Cohan, Chen Zhao

arXiv:2510.15232v1cs.LGcs.CL

TL;DR

Finance is high-stakes, yet trustworthy LLM use remains challenging because of hallucinations, sensitive information, and document-grounding demands. FINTRUST benchmarks seven dimensions across varied financial data and tasks; evaluations find limited legal and ethical awareness, including privacy and transparency shortcomings.

  • Problem

    Financial LLM use raises trustworthiness concerns involving hallucinations, private information, and reasoning grounded in financial documents.

    Source: §1 · p. 1

  • Method

    FINTRUST is a financial-domain benchmark spanning seven trustworthiness dimensions and varied task formats, contexts, and modalities.

    Source: §1 · pp. 1–2, §2 · p. 2

  • Results

    Across proprietary, open-source, and finance-specific models, experiments find limited legal and ethical awareness, including privacy and transparency shortcomings.

    Source: §6 · p. 9

  • Takeaways & Limitations

    The findings underscore the need for better alignment with domain-specific legal and ethical standards.

    Source: §6 · p. 9

    The authors did not fine-tune or align open-source models on FINTRUST, so the effects of targeted alignment efforts remain unassessed.

    Source: Limitations and Ethics · p. 9

Abstract

from arXiv · show

Recent LLMs have demonstrated promising ability in solving finance related problems. However, applying LLMs in real-world finance application remains challenging due to its high risk and high stakes property. This paper introduces FinTrust, a comprehensive benchmark specifically designed for evaluating the trustworthiness of LLMs in finance applications. Our benchmark focuses on a wide range of alignment issues based on practical context and features fine-grained tasks for each dimension of trustworthiness evaluation. We assess eleven LLMs on FinTrust and find that proprietary models like o4-mini outperforms in most tasks such as safety while open-source models like DeepSeek-V3 have advantage in specific areas like industry-level fairness. For challenging task like fiduciary alignment and disclosure, all LLMs fall short, showing a significant gap in legal awareness. We believe that FinTrust can be a valuable benchmark for LLMs' trustworthiness evaluation in finance domain.

1 Introduction

Finance is a high-stakes domain where hallucinations, sensitive information, and document-grounding challenges make trustworthy LLM deployment difficult. FINTRUST evaluates seven dimensions of financial trustworthiness, and results show both model-specific strengths and broad gaps in fiduciary alignment and disclosure.

1 Introduction

  • Finance deployment is challenging because models can hallucinate, handle private information, and struggle to ground reasoning in financial documents.Hallucinations are especially pronounced in financial documents combining tables and text.Source: §1 · p. 1
  • FINTRUST comprises 15,680 question-answer pairs across seven dimensions, grouped into trustfulness and robustness, legal and ethical handling, and cross-document investment discovery.Knowledge discovery uses long-form question answering requiring cross-document reasoning.Source: §1 · p. 1
  • Unlike benchmarks focused mainly on completing finance tasks, FINTRUST assesses alignment, uses fine-grained subtasks, and grounds evaluation in real-world financial scenarios.Its safety tasks include ten common financial crimes.Source: §1 · pp. 1–2
  • Proprietary models such as o4-mini outperform on tasks including safety, while DeepSeek-V3 excels in industry-level fairness and all models struggle with fiduciary alignment and conflict disclosure.Fin-R1 also performs strongly in personal-level fairness.Source: §1 · p. 2
  • The authors present FINTRUST as a systematic benchmark for financial-domain LLM trustworthiness and describe its multidimensional analysis as offering deployment insights.The benchmark’s safety, fairness, privacy, and transparency dimensions target legal and ethical aspects of trustworthiness.Source: §1 · p. 2, Figure 1 · p. 2

2 FINTRUST Benchmark

FINTRUST contains 15,680 instances spanning seven trustworthiness dimensions, multiple task formats, and text, table, and time-series contexts. Its subsets test factuality and abstention, safety, fairness, robustness, privacy, transparency, and investment insight generation.

2 FINTRUST Benchmark

  • FINTRUST contains 15,680 instances across seven dimensions, using short- and long-form QA, classification, masked-token prediction, and fill-in-the-blank tasks with text, tables, and time series.The dimensions are Truthfulness, Safety, Fairness, Robustness, Privacy, Transparency, and Knowledge Discovery.Source: §2 · p. 2, Table 1 · p. 3

2.1 Trustfulness Subset

  • The Trustfulness subset tests accurate answers and calibrated abstention on long-context financial questions, covering hallucination detection, confidence, numerical accuracy, and informative explanations.Hallucination questions are made unanswerable by removing numerically relevant information; calculation examples come from DocMath-Eval.Source: §2.1 · p. 3

2.2 Safety Subset

  • Safety tests whether models resist four black-box jailbreak types, using prompts about ten categories of illegal financial activity.The attacks include plain, scenario-nesting, low-resource-language, and genetic-algorithm-based methods.Source: §2.2 · pp. 3–4

2.3 Fairness Subset

  • Fairness evaluates industry-level buy/sell preferences and personal-level credit-default predictions for stability across age, gender, and race.Industry-level evaluation uses Russell 3000 firms, while personal-level evaluation uses the German Credit Dataset.Source: §2.3 · p. 4

2.4 Robustness Subset

  • Robustness tests whether models preserve performance on answerable questions and refuse unanswerable ones despite misspellings, incomplete information, out-of-domain rewrites, or simulated OCR errors.The tasks use perturbed questions from FailsafeQA.Source: §2.4 · p. 4

2.5 Privacy Subset

  • Privacy tests whether models refuse suspicious questions targeting personal information, with prompts that omit, imply, or explicitly state privacy expectations.Generated contexts use random Social Security numbers, bank details, and credit card numbers to avoid real-world privacy concerns.Source: §2.5 · p. 4

2.6 Transparency Subset

  • Transparency tests whether injected ownership information affects company investment predictions, using paired firms where Company A’s stock rises while Company B’s falls.Four prompts vary the company perspective and whether transparency requirements are explicit.Source: §2.6 · pp. 4–5

2.7 Knowledge Discovery Subset

  • Knowledge Discovery asks models to derive investment conclusions from internal knowledge, context, and calculations through fill-in-the-blank answers grounded in two companies’ 10-K reports.Gold answers fragment calculated data and inference text into blanks for completion.Source: §2.7 · p. 5

3 Experiment

The experiments evaluate 11 proprietary, open-source, and finance-specific LLMs across trustfulness, safety, fairness, robustness, privacy, transparency, and knowledge discovery. Results show different strengths across tasks, alongside persistent weaknesses in handling uncertain or sensitive financial contexts.

3.2 Trustfulness Subset Performance

  • All LLMs tend to abstain rather than hallucinate when uncertain, yet show overconfidence; GPT-4.1 mini, o4-mini, and DeepSeek-V3 show it more prominently.GPT-4.1, GPT-4.1 mini, and o4-mini significantly outperform other models in numerical accuracy; Fin-R1 and DianJin-R1-7B are exceptions to otherwise comparable correctness and informativeness scores.Source: §3.2 · pp. 5–6, Table 2 · p. 6

3.3 Safety Subset Performance

  • Most models are vulnerable to genetic-algorithm attacks; o4-mini is the exception that successfully defends against them.The authors hypothesize that o4-mini’s distinct log-probability distribution makes the attack ineffective.Source: Figure 2 · p. 6, §3.3 · p. 6

3.4 Fairness Subset Performance

  • Fin-R1 and DianJin-R1 excel in personal-level fairness, while DeepSeek-V3 ranks highest in industry-level fairness.Reasoning-based models perform significantly worse in personal fairness; the authors suggest reasoning may overemphasize sensitive attributes.Source: §3.4 · p. 6, Figure 3 · p. 7, Table 3 · p. 7

3.5 Robustness Subset Performance

  • All LLMs perform well on answerable questions but show a significant performance drop on unanswerable questions.Fin-R1, XuanYuan-Fin-X1, and DianJin-R1 were excluded because context length frequently exceeded their maximum context window.Source: Table 4 · p. 7, §3.5 · p. 7

3.6 Privacy Subset Performance

  • o4-mini is the most conservative on privacy-related questions, while finance-specific LLMs perform poorly.Implicit privacy references in the system prompt appear more effective than explicit ones, which the authors hypothesize may reflect context awareness.Source: Figure 4 · p. 7, §3.6 · p. 7

3.7 Transparency Subset Performance

  • Models tend to favor Company B when ownership is switched to Company B, often without disclosing ownership in their reasoning.Explicitly mentioning transparency in the system prompt leads to a greater drop in Company A selection rate; Fin-R1 achieves the highest accuracy, potentially due to evaluation-set overlap with its instruction-tuning data.Source: Figure 5 · p. 8, Table 5 · p. 8, §3.7 · p. 8

3.8 Knowledge Discovery Subset Performance

  • GPT-4.1 achieves the best knowledge-discovery performance, and all models except Fin-R1 score above 4, indicating satisfactory performance in the fill-in-the-blank setting.A representative failure case involves a numerical calculation error.Source: Table 6 · p. 8, §3.8 · p. 8

4 Discussion

The discussion identifies trade-offs across reasoning, fine-tuning, safety, privacy, and transparency. It concludes that weak ownership disclosure leaves models short of fiduciary alignment and may create legal issues in real-world applications.

Reasoning vs. Non-Reasoning

  • Reasoning models hallucinate less and are more likely to question errors, but reflection can draw in irrelevant personal information and add fairness bias.Source: §4 · p. 8

Fine-tuning

  • Fine-tuned models typically perform more fairly than other models, which the authors attribute as likely related to extra alignment during post-training.Source: §4 · p. 8

Safety Vulnerability and Privacy

  • Most models except o4-mini are vulnerable to genetic-algorithm attacks, and all three fine-tuned models show much worse context-aware privacy ability.The fine-tuned models expose sensitive information to the questioner with almost no resistance; the authors call this a critical ability deterioration during fine-tuning.Source: §4 · p. 8

Transparency and Legal Alignment

  • Although proprietary models generally shift answers less when system prompts change, all models disclose ownership poorly, potentially misclassifying it as sensitive information.The authors say this conflicts with fiduciary-alignment legal requirements and may cause legal issues in real-world applications.Source: §4 · pp. 8–9

5 Related Work

Related work covers financial-task benchmarks and trustworthiness evaluations, then identifies a gap in holistic, trustworthiness-centered evaluation for finance. FINTRUST is presented as a response tailored to financial applications.

Existing Benchmarks in the Trustworthiness and Financial Domains

  • FINTRUST addresses the stated gap with a comprehensive benchmark for evaluating LLM trustworthiness in finance.Source: §5 · p. 9
  • Prior trustworthiness work examines dimensions such as hallucination, fairness, and safety, and some comprehensive evaluations focus on specific domains such as medical VLMs.Source: §5 · p. 9
  • Financial benchmarks have often focused on answer accuracy, while other work evaluates reasoning abilities.Source: §5 · p. 9

6 Conclusion

The conclusion presents FINTRUST as a comprehensive benchmark for financial-domain LLM trustworthiness and reports limited legal and ethical awareness across model types. It calls for improved alignment with domain-specific legal and ethical standards.

6 Conclusion

  • FINTRUST is a comprehensive benchmark for evaluating LLM trustworthiness in the financial domain.Source: §6 · p. 9
  • Proprietary, open-source, and finance-specific fine-tuned models show limited legal and ethical awareness, including on privacy and transparency.Source: §6 · p. 9
  • The conclusion underscores the need for better alignment with domain-specific legal and ethical standards.Source: §6 · p. 9

Limitations and Ethics

The study limits its evaluation to publicly available LLMs or APIs and identifies untested alignment effects and potential financial harms as relevant concerns.

Limitations

  • The study does not test instruction fine-tuning or alignment of open-source LLMs, so it cannot assess how targeted alignment might affect trustworthiness.The authors identify instruction tuning or reinforcement learning with the dataset as future work.Source: Limitations and Ethics · p. 9

Ethics and scope

  • Experiments use publicly available LLMs or APIs, with no deployment in real-world financial systems.Source: Limitations and Ethics · p. 9
  • The authors flag hallucinations, overconfident outputs, and potential biases as ethical concerns that could cause significant harm if misused.FINTRUST is designed to expose these risks and support safer, more trustworthy financial LLM applications.Source: Limitations and Ethics · p. 9

Appendix Contents

The appendix contents list task formulations, overall results, prompts, a bias map, and error analysis.

Appendix contents

A.1 Task Formulation Figures

The appendix figures are titled as task formulations for seven trustworthiness dimensions: trustfulness, safety, fairness, robustness, privacy, transparency, and knowledge discovery.

Task formulation figures

A.2 Overall Results

The section pairs an eleven-model, multidimension evaluation with task-level examples of trustworthiness failures. These examples show errors in abstention, financial calculations, safety, fairness, robustness, privacy, transparency, and knowledge discovery.

A.7 Error Analysis

  • Table 7 organizes the eleven-model evaluation across trustfulness, safety, fairness, robustness, privacy, transparency, and knowledge-discovery measures.Source: Table 7 · p. 16

Trustfulness – Hallucination + Certainty

Safety – Genetic Algorithm attack

Fairness – Personal Level

  • In the personal-fairness example, Qwen3’s answer changes when sensitive attributes are altered, although the altered-context answer matches the gold answer.Source: Fairness – Personal Level · p. 26

Robustness – Irrelevant Context

  • With irrelevant context, GPT-4.1 answers using a different company and date, and the analysis says models may infer figures without noticing the context mismatch.Source: Robustness – Irrelevant Context · p. 27

Privacy – Implicit Mention

  • In the example listing Ethan Dixon’s conservative bond-heavy portfolio and Steven Roberts’s aggressive short-term trading strategy, GPT-4.1 and GPT-4.1-mini state Ethan’s preference, while o4-mini says it cannot share that.Source: Privacy – Implicit Mention · p. 28

Transparency – Company B Ownership & Direct Mentioning Transparency

Knowledge Discovery

  • In the knowledge-discovery example, GPT-4.1’s response receives a score of 3; the analysis attributes low judge scores to calculation errors or incorrect or directionally different inferences.Source: Knowledge Discovery · p. 30
Loading 2510.15232v1…