PaperlayerView exact revision on arXiv

Source-linked AI summary

Same Claim, Different Judgment: Benchmarking Scenario-Induced Bias in Multilingual Financial Misinformation Detection

Zhiwei Liu, Yupen Cao, Yuechen Jiang, Mohsinul Kabir, Polydoros Giannouris, Chen Xu, Ziyang Xu, Tianlei Zhu, Md. Tariquzzaman, Triantafillos Papadopoulos, Yan Wang, Lingfei Qian, Xueqing Peng, Zhuohan Xie, Ye Yuan, Saeed Almheiri, Abdulrazzaq Alnajjar, Mingbin Chen, Harry Stuart, Paul Thompson, Prayag Tiwari, Alejandro Lopez-Lira, Xue Liu, Jimin Huang, Sophia Ananiadou

arXiv:2601.05403v2cs.CL

TL;DR

Financial misinformation benchmarks and LLM-bias studies have offered limited evidence on multilingual, scenario-dependent financial claim judgments. MFMD-Scen combines multilingual claims with financial scenarios and evaluates 22 LLMs, finding substantial behavioral biases that are amplified in low-resource languages.

  • Problem

    Existing financial misinformation benchmarks largely use fixed evaluation settings, and many LLM-bias studies use simplified tasks rather than multilingual financial claim verification.

    Source: §1 · p. 1

  • Method

    MFMD-Scen combines a multilingual financial misinformation dataset with controlled persona, region, and identity scenarios to evaluate claim verification across contexts.

    Source: §1 · p. 2, §2.1 · p. 3

  • Results

    Current models show substantial behavioral biases, especially in retail-investor, herding-personality, and emerging Asian-market scenarios, with stronger effects in low-resource languages.

    Source: §1 · p. 2, §5 · p. 9

  • Takeaways & Limitations

    MFMD-Scen provides a benchmark for evaluating behavioral biases in multilingual financial misinformation judgments across economic scenarios.

    Source: §1 · p. 2, §5 · p. 9

    The finance-filtered Snopes dataset is imbalanced toward false information, and only two human annotators were available for some regions.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Large language models (LLMs) have been widely applied across various domains of finance. Since their training data are largely derived from human-authored corpora, LLMs may inherit a range of human biases. Behavioral biases can lead to instability and uncertainty in decision-making, particularly when processing financial information. However, existing research on LLM bias has mainly focused on direct questioning or simplified, general-purpose settings, with limited consideration of the complex real-world financial environments and high-risk, context-sensitive, multilingual financial misinformation detection tasks MFMD. In this work, we propose MFMDScen, a comprehensive benchmark for evaluating behavioral biases of LLMs in MFMD across diverse economic scenarios. In collaboration with financial experts, we construct three types of complex financial scenarios: (i) role- and personality-based, (ii) role- and region-based, and (iii) role-based scenarios incorporating ethnicity and religious beliefs. We further develop a multilingual financial misinformation dataset covering English, Chinese, Greek, and Bengali. By integrating these scenarios with misinformation claims, MFMDScen enables a systematic evaluation of 22 mainstream LLMs. Our findings reveal that pronounced behavioral biases persist across both commercial and open-source models. This project is available at https://github.com/lzw108/FMD.

1 Introduction

Financial misinformation benchmarks and LLM-bias studies have largely used fixed or simplified settings, leaving multilingual, scenario-dependent financial judgment underexamined. MFMD-Scen combines multilingual claims with controlled financial scenarios and finds that scenario context can systematically shift LLM judgments even when claims remain fixed.

1 Introduction

  • Prior financial misinformation benchmarks use fixed classification settings, while many LLM-bias studies rely on direct elicitation or simplified tasks rather than financial claim verification.Those settings offer limited leverage for studying multilingual or scenario-dependent judgment variability.Source: §1 · pp. 1–2
  • The benchmark pairs real-world claims with English, Chinese, Greek, and Bengali versions, using native-speaker review and targeted human revision for translations.Claims are selected for global relevance and recovered as complete statements from Snopes-based FinFact claims.Source: §1 · p. 2
  • MFMD-Scen evaluates binary financial-claim verification with and without scenario context across persona, region, and identity axes.The benchmark measures scenario-conditioned effects as the performance difference between scenario-aware and scenario-agnostic evaluation.Source: §1 · p. 2
  • Scenario context systematically shifted judgments on fixed claims, most strongly for retail-investor and herding scenarios and emerging Asian markets, with identity-role interactions and amplified effects in low-resource languages.Models could rely on risk-averse, skepticism-heavy defaults rather than claim content; explicit reasoning benefits were clearer primarily for large models.Source: §1 · p. 2
  • The study evaluates 22 LLMs and reports substantial behavioral biases, particularly in retail-investor, herding-personality, and emerging Asian-market contexts.Its expert-designed benchmark covers roles, personality, regional, and socio-cultural contexts.Source: §1 · p. 2

2 MFMD-Scen Benchmark

MFMD-Scen combines financial scenarios with misinformation claims to benchmark scenario-dependent financial misinformation detection. It defines verification under three scenario types and constructs a multilingual dataset for evaluation.

2.1 Task Formulation

  • The task tests whether a financial claim is True or False within a persona-, region-, or identity-conditioned scenario, comparing scenario-conditioned and scenario-free predictions.Behavioral bias is quantified as the difference in F1 scores between those two conditions for the same claim.Source: §2.1 · p. 3

2.1.1 Task 1: Detection in Different

  • Persona scenarios pair financial decision-maker profiles with five behavioral biases, each presented explicitly or through implicit narrative cues.The biases are overconfidence, loss aversion, herding, anchoring, and confirmation bias.Source: §2.1.1 · pp. 3–4

2.1.2 Task 2: Detection in Different Financial Markets (MFMD-region)

  • MFMD-region varies market context across six economic regions and represents differences in regulation, risk culture, macroeconomic conditions, asset classes, and market maturity.The regions are Europe, North America, Asia Pacific, China Mainland, Australia, and the UAE.Source: §2.1.2 · p. 4

2.1.3 Task 3: Detection in Different Identities (MFMD-identity)

  • MFMD-identity varies scenario background by combining ethnicity and faith or belief with financial roles.The scenario design draws on demographic and religious distributions from public data sources.Source: §2.1.3 · p. 4

2.2 Multilingual Financial Misinformation Dataset Construction

  • The multilingual dataset starts from complete financial claims, selects globally relevant news, and translates claims into Chinese, Greek, and Bengali for comparison with English.Two native speakers evaluated each translation, with human revision and review when quality was insufficient. Financial misinformation is defined as false or misleading information involving monetary value, assets, transactions, or promised financial gain that may lead to financially consequential behavior or economic loss.Source: §2.2 · p. 4

3 Evaluations

The evaluation applies the benchmark to a broad set of language models, reports standard performance and scenario-conditioned metrics, and includes a rough human-performance assessment.

3 Evaluations

  • The evaluation compares existing LLMs using the proposed metrics and human-level performance assessment.The models were evaluated after construction of MFMD-Scen.Source: §3 · p. 5, §3.3 · p. 5

3.1 Models

  • The evaluation covers reasoning-oriented and standard chat models across open- and closed-source offerings.LLaMA 3.1-8B was unable to produce responses in most cases, so its results are not reported.Source: §3.1 · p. 5

3.2 Evaluation metrics

  • The study reports overall accuracy and macro-F1 by language, and scenario-specific macro-F1 for true and false claims alongside AM and MAV across models.AM represents bias direction, while MAV represents deviation magnitude.Source: §3.2 · p. 5

3.3 Human-level Performance Measurement

  • A rough human-level assessment used volunteers to judge the truthfulness of 144 English claims from five MFMD-region scenarios.Eleven volunteers were from China Mainland, with two recruited from each of Europe, Asia Pacific, Australia, and the UAE.Source: §3.3 · p. 5

4 Evaluation Results

Across 22 models, false statements are judged more reliably than true ones, while scenario, role, region, identity, and language are associated with systematic shifts in financial misinformation judgments. Bias patterns vary by context, with retail-investor and herding scenarios, emerging Asian markets, and identity-role interactions among the prominent findings.

4.1 Main findings

  • Models show relatively mature judgments of false statements but deficiencies on true statements, alongside clear scenario-dependent financial misinformation biases.False-condition scores exceed 0.85, while performance drops under the TRUE condition.Source: §4.1 · p. 5, §4.2 · p. 6
  • Across settings, retail-investor and herding scenarios, emerging Asian markets, and low-resource languages are associated with stronger negative or pronounced biases.The reported bias direction and magnitude also vary by role, region, and identity.Source: §4.1 · pp. 5–6

4.2 Results on MFMD-persona

  • Role differences are generally minimal for FALSE statements, but retail investors perform worst for TRUE statements under implicit bias; explicit prompts yield no consistent cross-language role pattern outside English.Professionals slightly outperform company owners under implicit bias.Source: §4.2 · p. 6
  • TRUE-category bias is smaller under explicit than implicit scenarios, while herding produces larger bias than confirmation scenarios, especially for retail investors and professional institutions.False-category differences between implicit and explicit scenarios are small.Source: §4.2 · p. 6
  • For FALSE statements, bias rises with linguistic difficulty or resource scarcity, particularly in Bengali; TRUE-category bias is larger and comparatively elevated in Chinese.Explicit company-owner prompts also show language-specific patterns: herding in English, anchoring in Chinese, and loss aversion in Greek and Bengali.Source: §4.2 · p. 6
  • For TRUE statements, smaller models show larger, predominantly negative bias; larger or closed-source models are more stable, while FALSE-category biases are mostly mild.Mistral models generally show larger FALSE-category bias, with Mistral-NEMO excepted; Bengali is a noted exception for Claude-3.5-Haiku.Source: §4.2 · p. 7

4.3 Results on MFMD-region

  • Region results show negative bias in Asia Pacific and China Mainland scenarios and predominantly positive bias in U.S. scenarios; retail investors show the largest bias in UAE scenarios.European and U.S. scenarios generally induce smaller bias, while professionals and company owners show greater bias primarily in Asian scenarios.Source: §4.3 · p. 7, Figure 4 · p. 8

4.4 Results on MFMD-identity

  • Identity and role interact: American groups show positive bias across roles, while several Arab, Latino/Hispanic, and African identities shift from negative retail-investor bias to positive company-owner bias.The reported role-driven directional shift is stronger for Islam than Christianity; the identity heatmap summarizes bias across representative models and 22-model AM and MAV.Source: §4.4 · p. 8, Figure 5 · p. 8

5 Conclusion

MFMD-Scen evaluates 22 mainstream LLMs across persona, region, and identity scenarios. The models exhibit substantial behavioral biases, especially in retail-investor, herding, and emerging Asian market scenarios, with biases further amplified in low-resource languages.

5 Conclusion

  • MFMD-Scen evaluates 22 mainstream LLMs across persona, region, and identity tasks and finds substantial financial misinformation judgment biases, particularly in retail-investor, herding, and emerging Asian market scenarios.The conclusion reports that these biases are further amplified in low-resource languages.Source: §5 · p. 9

Social Impact

Scenario-dependent biases in financial statement judgments may have practical implications for finance-related LLM deployment. The authors also note that responsibly designed LLMs could support financial literacy, decision-making, and education.

Social Impact

  • In decision-making tools, scenario-dependent biases may mislead investors or amplify cultural, regional, or demographic disparities, while responsibly designed LLMs could support financial literacy and education.The authors frame these as possible practical implications of deployment.Source: Social Impact · p. 9

Limitations

The benchmark has an imbalanced dataset and limited human annotation in some regions. The authors address the class imbalance with category-wise analyses, but report that some regions had only two human annotators.

Limitations

  • The Snopes-derived finance dataset is false-heavy, and resource constraints left some regions with only two human annotators.The authors use category-wise analyses to mitigate biases introduced by dataset imbalance.Source: Limitations · p. 9

Ethical Considerations

The section frames MFMD-Scen as an audit of model behavior in hypothetical financial contexts, against prior bias studies and financial misinformation benchmarks with limited scenario and multilingual evaluation.

Ethical Considerations

  • The scenarios are hypothetical evaluation probes for auditing model bias, not for profiling or judging individuals or groups.They focus on roles, behaviors, and contexts to examine systematic model tendencies.Source: Ethical Considerations · p. 9

A.1 Bias and Behavioral Considerations in LLMs

  • Prior LLM-bias studies often use direct elicitation or simplified scenarios, leaving complex, sensitive financial environments less examined.The reviewed work covers cognitive biases, preference and belief tasks, interview responses, and simulated political debates.Source: §A.1 · p. 12

A.2 Financial Misinformation Detection

  • Financial misinformation research includes detection methods and claim-verification benchmarks, but remains mostly English-focused and lacks evaluation across financial scenarios.Existing financial benchmarks often frame detection as classification or evidence-based verification.Source: §A.2 · p. 12

B.2.1 Europe; B.2.2 USA; B.2.3 Asia Pacific

  • Regional prompts vary market context and role, including European regulatory norms, U.S. performance-driven narratives, and Asia Pacific volatility and policy cues.Other prompts cover China Mainland, Australia, and the UAE through their distinct policy, commodity, or oil-market contexts.Source: §B.2.1 · p. 12, §B.2.2 · p. 13, §B.2.3 · pp. 13–14

B.2.4 China Mainland; B.2.5 Australia; B.2.6 UAE

  • China Mainland, Australia, and UAE prompts pair retail-investor, professional, and company-owner roles with local policy, commodity, or investment-growth contexts.The prompts include China’s policy shifts, Australia’s resource cycles, and UAE oil markets and government-backed initiatives.Source: §B.2.4 · p. 14, §B.2.5 · p. 14, §B.2.6 · p. 14

B.3 MFMD-identity: Ethnicity and (Faith/Belief) Scenario Pairs

  • MFMD-identity uses ethnicity–faith or belief pairs in retail-investor and company-owner templates to examine background-dependent interpretation.The listed pairs include American–Christianity, European–Christianity, and Latino/Hispanic–Christianity.Source: §B.3 · p. 14

C Data Annotation

The data-annotation process combines expert review, financial and regional filtering, translation review, and language-specific financial misinformation sources to assemble the benchmark data.

C.1 Annotation System

  • Scenario design was led by a finance-oriented expert, with interdisciplinary feedback; two native speakers proficient in English reviewed translations for each language.The Doccnao platform was used for annotation, and two additional annotators contributed to GlobalEn construction.Source: §C.1 · p. 15

C.2 Financial vs Non-financial

  • From 1,788 collected items, annotators retained 502 claims labeled financial by both, adjudicating disagreements with a third finance expert.The initial 200 items informed multiple rounds of discussion to establish the final guidelines.Source: §C.2 · p. 15

C.3 Regional vs Global

  • Regional annotation distinguishes news limited to particular regions or countries from news with potential worldwide impact; 144 global items were selected.Disagreements were adjudicated by a third finance expert.Source: §C.3 · pp. 15–16

C.4 Translation Review

  • GPT-4.1 translations into Chinese, Greek, and Bengali were reviewed by two native speakers per language, and Poor-rated items received human translation and review.The human-revised counts were 5 Chinese, 12 Greek, and 31 Bengali items.Source: §C.4 · p. 16

D Original language data

  • Because open-source financial misinformation datasets are scarce, the study uses English financial data and filters financial portions from multidomain datasets for other languages.The source-data descriptions are organized by language.Source: D Original language data · p. 16

D.1 English Part

  • The English source includes FinDVer claims entailed or refuted using textual and tabular information from financial documents.Annotators generated entailed claims and experts created refuted claims by perturbing them.Source: §D.1 · p. 16

D.2 Chinese Part

  • For Chinese, the dataset filters financial-domain items from MDFEND and CHEF, which draw on Weibo fake and real news and evidence-based real-world claims, respectively.The passage reports MDFEND at 4,488 fake and 4,640 real news items, and CHEF at 10K claims overall.Source: §D.2 · p. 17

D.3 Bengali Part

  • BanMANI uses seed news articles to generate manipulated and non-manipulated social-media content with ChatGPT, followed by human validation.For large datasets, 250 instances per category were sampled for testing.Source: §D.3 · p. 17

D.4 Multilingual Evaluation

The multilingual evaluation reports differences by model family, language, and task complexity, using both scenario-free and scenario-conditioned claim prompts. Closed-source models achieve the best results in the reported evaluation.

D.4 Multilingual Evaluation

  • Large Qwen3 and GPT models perform strongly on Greek and Bengali, while LLaMA and Mistral show a larger low-resource-language performance gap.Greek and Bengali are contrasted with Chinese and English in the reported comparison.Source: §D.4 · p. 17
  • Mistral performs relatively poorly on evidence-based FinDVer and CHEF tasks, and performance generally rises with model size.For same-sized Qwen models, reasoning models do not significantly outperform no-think models on financial misinformation tasks.Source: §D.4 · p. 17
  • Closed-source gpt-5-mini and GPT-4.1 achieve the best results, while the evaluation reports a gap between closed- and open-source models.Table 5 presents the MFMD dataset evaluation.Source: §D.4 · p. 17, Table 5 · p. 18

E Templates for evaluating LLMs

  • The evaluation prompts ask for a True/False claim judgment, either without scenario information or with an instruction to consider a supplied scenario.The scenario prompt includes both scenario and claim text.Source: E Templates for evaluating LLMs · p. 18

F Human performance on the GlobalEn dataset

The section compares 22 LLMs with volunteers who have investment experience across five regions on GlobalEn. Smaller models are often closer to human performance, while larger models often exceed it.

F Human performance on the GlobalEn dataset

  • Human judgments came from investment-experienced volunteers across five regions, with 11 participants from China Mainland and two from each other region.Participants judged claims using their own past experience and knowledge; regional performance was averaged.Source: F Human performance on the GlobalEn… · p. 18
  • For false claims, Mistral-Large is relatively close to human performance except in China Mainland, while Mixtral-8x7B is close in that region.These are the reported false-category comparisons.Source: F Human performance on the GlobalEn… · p. 18
  • Smaller-scale models tend to be closer to human performance, whereas larger models often exceed it across the reported regional comparisons.The closest model for true claims varies by region: Gemini-2.0 in Europe, Qwen-8B in Asia regions, Mistral-Large in Australia, and Mistral-Small-24B in the UAE.Source: F Human performance on the GlobalEn… · p. 18

G Detailed Results (Table 7 to 13)

Identity-based bias varies with both label truth and role: it intensifies for RetailInvestor under True labels but is buffered for CompanyOwner.

G Detailed Results (Table 7 to 13)

  • Under True labels, most non-American Atheist groups and European-Muslim groups shift toward negative directional bias in RetailInvestor scenarios, while American-Atheist stays positive and Arab-Christianity remains near neutral.Bias magnitude increases substantially; under False labels, the groups show small, largely homogeneous biases with minimal directional differences.Source: G Detailed Results (Table 7 to 13) · p. 18
  • Identity-based bias is amplified under True labels in RetailInvestor scenarios and buffered in the more institutionalized CompanyOwner role.The reported pattern spans groups, with stronger and less stable identity effects in RetailInvestor and more uniformly positive, less dispersed deviations in CompanyOwner.Source: G Detailed Results (Table 7 to 13) · pp. 18–19
  • CompanyOwner scenarios show uniformly positive, less dispersed deviations across groups under True labels, attenuating religion–ethnicity differentiation.Source: G Detailed Results (Table 7 to 13) · p. 19
  • Tables 7–10 present results for MFMD-persona across English, Chinese, Greek, and Bengali, respectively.Source: Table 7 · p. 20, Table 8 · p. 21, Table 9 · p. 22
  • Tables 11–13 cover MFMD-region and MFMD-identity results, including a second part for MFMD-identity.Source: Table 11 · p. 24, Table 12 · p. 24, Table 13 · p. 25

H Specific Cases (Tables 14 to 16)

Loading 2601.05403v2…