Pith. sign in

REVIEW 4 major objections 5 minor 40 references

The way a loan application is written out as text — JSON, sentence fragments, or fluent prose — materially changes both how accurately LLMs approve loans and how evenly they treat male and female applicants, and in zero-shot mode none of th

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Serialization format and in-context examples change both accuracy and gender fairness of LLM loan approvals, with finance-tuned models often showing larger disparities.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection The benchmark is useful, but the fairness conclusions don't hold until degenerate predictions are separated from real disparities. the 4 major comments →

arxiv 2508.21512 v1 pith:PHTSF7BB submitted 2025-08-29 cs.LG cs.CLcs.CY

Accept or Deny? Evaluating LLM Fairness and Performance in Loan Approval across Table-to-Text Serialization Approaches

Israel Abebe Azime , Deborah D. Kanubala , Tejumade Afonja , Mario Fritz , Isabel Valera , Dietrich Klakow , Philipp Slusallek This is my paper
classification cs.LG cs.CLcs.CY
keywords LLM fairnessloan approvaltable-to-text serializationin-context learningequality of opportunitystatistical paritygender biastabular data
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that the text format used to feed loan applications to large language models is a first-order determinant of both accuracy and gender fairness. Evaluating ten open models on loan data from Ghana, Germany, and the United States, it finds that formats such as GReaT and LIFT can raise F1 scores while simultaneously widening equality-of-opportunity gaps, and that zero-shot LLMs consistently underperform a simple logistic regression baseline, often defaulting to uniform approval or denial. In-context learning improves F1 by 4.9–59.6% but its fairness effect is unstable across datasets, formats, and shot counts. If these claims hold, lenders must treat data serialization and example selection as fairness-relevant design choices, and cannot yet rely on LLMs as primary loan decision-makers.

Core claim

On the paper's own terms, the central claim is that serialization — converting each tabular loan application into text — is a first-order factor in both performance and fairness of LLM-based loan approval. Across seven formats and ten open models, the natural-language formats GReaT and LIFT delivered the highest F1 scores in several settings but also the largest equality-of-opportunity gaps between gender groups. In zero-shot prompting every LLM scored below a logistic regression benchmark, several effectively approving or denying everyone; adding 2–8 gender-balanced in-context examples improved F1 by 4.9–59.6% yet did not uniformly reduce bias, pushing EO differences above 0.10 in some conf

What carries the argument

The machine doing the work is the serialization pipeline plus the fairness metrics it is measured against. Each applicant row is rendered in seven formats — JSON and List as bare key-value pairs, Text, HTML, and LaTeX as structured or marked-up strings, GReaT as 'feature is value' sentences, and LIFT as fluent prose ('A 32-year-old female is applying for a loan for 48 months...'). The same ten models read these strings zero-shot or with 2–8 gender-balanced in-context examples. Fairness is scored by equality of opportunity (true-positive rate equal across gender) and statistical parity (approval rate equal across gender), and the paper's headline is that these two differences move as much wit

Load-bearing premise

The fairness comparisons assume that gender-specific approval and true-positive rates estimated from roughly 120 test applicants in the smallest dataset are stable enough to compare; the paper reports them without confidence intervals, so some of the disparities it highlights could be sampling noise rather than systematic bias.

What would settle it

Recompute the equality-of-opportunity and statistical-parity values from the paper's own test splits with bootstrap confidence intervals (e.g., 1,000 resamples per model–format–dataset cell). If the 95% interval around FinMA-7B-full's 0.16 EO gap on Germany or the >0.10 gaps at n=8 on Ghana includes zero, the claim that serialization and financial fine-tuning systematically drive gender bias loses support; if the intervals exclude zero, the claim holds.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Practitioners should treat serialization choice as a decision variable with fairness consequences, not a formatting detail: the same model can flip from fair to biased purely from how the table is written out.
  • Zero-shot LLM deployment for loan approval is not yet justified: every tested LLM underperformed logistic regression, and several effectively defaulted to uniform approval or denial.
  • Financial fine-tuning does not automatically confer accuracy or fairness; the finance-tuned FinMA-7B-full model produced the largest gender disparities on Ghana and Germany.
  • In-context learning is a performance lever but not a fairness guarantee: adding examples lifted F1 by up to 59.6% yet sometimes pushed equality-of-opportunity differences above 0.10.
  • Robust evaluation of LLM lending tools requires multiple serializations and regionally diverse data, since format–dataset interactions can matter more than which model is used.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The authors leave implicit that the smallest test splits make several headline disparities statistically fragile: Ghana's 614 rows yield roughly 123 test rows, so a per-gender equality-of-opportunity gap like FinMA-7B-full's 0.16 could easily sit within sampling error; a bootstrap analysis would test whether the format–fairness link is systematic.
  • A testable extension the paper does not run: prompt the same models with deliberately imbalanced in-context examples to separate the effect of example demographics from the effect of serialization format on fairness.
  • If format–region interactions dominate model identity, a practical next step is serialization-robust inference — requiring a model's prediction to stay fixed across all seven formats and flagging formats where it flips — as a fairness guardrail in deployment.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper benchmarks ten open-source LLMs on loan-approval classification using three geographically distinct datasets (Ghana, Germany, U.S.), seven table-to-text serialization formats, and both zero-shot and few-shot in-context learning. Performance is measured by weighted F1 and fairness by equality of opportunity (EO) and statistical parity (SP) differences across gender. The central claims are that serialization format significantly affects both performance and fairness, that formats such as GReaT and LIFT improve F1 while worsening fairness disparities, that zero-shot LLMs generally underperform a logistic regression baseline and often default to uniform predictions, and that in-context learning improves F1 but has context-dependent fairness effects. The paper also discusses model size, instruction tuning, and prompt sensitivity, with supplementary token-attribution experiments.

Significance. If supported, the paper would provide a useful broad empirical resource: ten models, seven serializations, three datasets, and standard fairness metrics, with resources released on Hugging Face and a reproducible lm-eval harness. The descriptive observation that serialization and ICL are first-order variables for LLM loan approval is practically important, as is the recommendation to keep classical baselines as decision support. However, the headline fairness conclusions are currently stronger than the evidence: degeneracy of predictions, small per-group test samples, and the absence of uncertainty quantification all affect the load-bearing claims about 'amplifying' bias and about specific serialization formats 'exacerbating fairness disparities.' The paper is best treated as a valuable descriptive benchmark that needs targeted statistical and conceptual fixes before its fairness conclusions can be accepted.

major comments (4)
  1. [§3.5, Table 4, Fig. 5, Discussion] Fairness comparisons are confounded by degenerate predictions. A constant classifier has EO=SP=0 by definition, and the paper itself reports zero/one baselines with 0.00 on both metrics and later states that zero-shot LLMs 'frequently default to uniform approval or denial.' Many zero entries for LLMs in Table 4 and Fig. 5 may therefore reflect constant predictions, not fairness. Comparing FinMA-7B-full's nonzero EO/SP against such entries conflates predictive variability with bias. The claims that FinMA-7B-full 'amplifies historical gender bias' and that certain serializations 'exacerbate fairness disparities' are not established unless the analysis conditions on non-degeneracy (e.g., reports prediction entropy or positive rate, or restricts to models with non-constant outputs).
  2. [§3.5, Table 4, Fig. 5, §4.5] No confidence intervals, bootstrap estimates, or significance tests are reported for the fairness metrics. For the Ghana dataset the test set is about 20% of 614 rows (~123), so after splitting by gender, EO differences such as FinMA-7B-full's 0.16 on Germany or the various Ghana values are estimated from tens of applicants per group. The statement that serialization 'significantly affects' fairness, and the stronger assertion that financial models 'amplify existing historical gender bias,' require uncertainty quantification or at minimum explicit acknowledgment that the observed differences may be sampling noise. The same issue affects the few-shot fairness comparisons in Fig. 7.
  3. [Abstract, §4.4, Figs. 3-4] The abstract's numerical claim that ICL improves performance by 4.9-59.6% is not tied to a table or equation in the main text, and the supporting figures (Figs. 3-4) do not show per-condition error bars or confidence intervals. Since the size of the improvement varies widely across datasets and formats, this headline range needs an explicit definition of what is being compared, which model/format conditions are included, and how the variability across prompts or splits is accounted for. Without that, the number is not verifiable from the paper as written.
  4. [Discussion vs. Figure 2, §4.1-4.2] The summary statement that 'in zero-shot scenarios, all LLMs perform worse than a Logistic Regression model baseline' is contradicted by the paper's own Figure 2 caption and Section 4.2, which name GReaT on Ghana, List/Text on Germany, and Gemma-2-27b-it on the U.S. as showing 'modest improvements' over the logistic regression baseline. If the claim is intended only for the JSON serialization, it needs to say so explicitly; as written, the discrepancy between 'all' and 'most' undermines a supporting conclusion.
minor comments (5)
  1. [§4.10, Fig. 7] The text says 'In the German dataset... Few-shot examples (e.g., n=8) can lead to significant fairness disparities... reaching differences of over 0.10 for some serialization methods in the Ghana datasets.' The mixed German/Ghana reference is confusing and should be corrected.
  2. [Table 3, §4.5] The sentence that FinMA-7B-full is 'the only one fine-tuned for finance' contradicts Table 3, which lists both FinMA-7B-full and FinMA-7B-NLP as fine-tuned on financial data.
  3. [Throughout, Table 8] Inconsistent capitalizations appear: 'GReat'/'great' vs. 'GReaT', 'LlaMA-3-70B' vs. 'LLaMA-3-70B', and 'Latex' vs. 'LaTeX'. There is also a typo 'purpuse' in Table 8.
  4. [Appendix H] The token-attribution results are described as inconclusive and are relegated to the appendix; the main text should either briefly state this or remove references to explaining model behavior in the introduction. The appendix currently contains many large figures with limited interpretation.
  5. [§B] The evaluation setup mentions three prompts per dataset but does not describe how generations were parsed or what 'log-likelihood prediction' means for multi-token outputs. A short description of the parsing/filtering of predictions would improve reproducibility.

Circularity Check

0 steps flagged

No circularity: empirical benchmark against external ground-truth labels; no quantity is defined in terms of the target result.

full rationale

This paper is an empirical evaluation. The reported quantities (weighted F1, equality of opportunity, statistical parity) are measured from model predictions on held-out test splits against ground-truth labels. No parameter is fitted to the reported metrics and then presented as a prediction: the Logistic Regression baseline is trained with default scikit-learn parameters on the training split and used only as an external reference point, not as part of the LLM claims. In-context examples are selected by gender balance ('top-n examples balanced by gender'), not by optimizing F1 or fairness, so the fairness and performance outcomes are not forced by construction. The serialization formats are pre-existing external methods (JSON, GReaT, LIFT, Text, HTML, Latex) with citations to prior work; the paper does not define its dependent variables in terms of its independent variables. The zero/one baselines having zero EO/SP is a mathematical property of constant classifiers, but the paper reports these as baselines rather than deriving its central claim from that property. The skeptic's concern about degenerate predictions is a validity threat to fairness comparisons, not a circularity: it does not show that any claimed result reduces to its input by definition or by self-citation. Self-citations (Kanubala et al. 2024; Majumdar et al. 2025) are contextual and not load-bearing for the empirical conclusions. The derivation chain is therefore self-contained against external benchmarks and no circular step is present.

Axiom & Free-Parameter Ledger

0 free parameters · 5 axioms · 0 invented entities

No free parameters are fitted to data; the study is an empirical evaluation. Shot counts (2, 4, 6, 8), prompts, and serialization templates are design choices, not fitted parameters. The logistic regression baseline uses default scikit-learn parameters as a reference point.

axioms (5)
  • domain assumption Gender is a binary attribute (male/female) and is the sole sensitive attribute evaluated
    Section 3.5 defines A as belonging to {male, female}; real populations include non-binary genders and multiple sensitive attributes, which may affect fairness estimates.
  • domain assumption Weighted F1, EO, and SP are appropriate evaluation measures for loan approval decisions
    Section 3.5 chooses these metrics; they emphasize either overall accuracy or binary group fairness but ignore other fairness notions and cost asymmetries.
  • domain assumption The three datasets are comparable proxies for loan approval in their respective regions
    Section 3.2 states regions were arbitrarily chosen and datasets come from different sources with different features and label definitions (Yes/No, Good/Bad, Loan_Status), so cross-region comparisons may be confounded.
  • domain assumption LLM log-likelihood scoring over the possible label tokens adequately represents the model's decision
    Appendix B uses the lm-eval harness with log-likelihood prediction for possible class outputs; this assumes token likelihood corresponds to decision quality and ignores decoding variability and ties.
  • domain assumption The three hand-written prompts per dataset adequately span prompt sensitivity
    Appendix F provides three prompts; results average over them, but broader prompt families may change conclusions.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Accept or Deny? Evaluating LLM Fairness and Performance in Loan Approval across Table-to-Text Serialization Approaches." pith.science (2026). https://pith.science/paper/PHTSF7BB

@misc{pith2026250821512,
  author       = {Pith},
  title        = {Pith review of: Accept or Deny? Evaluating LLM Fairness and Performance in Loan Approval across Table-to-Text Serialization Approaches},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PHTSF7BB}},
  note         = {Machine review of arXiv:2508.21512}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large Language Models (LLMs) are increasingly employed in high-stakes decision-making tasks, such as loan approvals. While their applications expand across domains, LLMs struggle to process tabular data, ensuring fairness and delivering reliable predictions. In this work, we assess the performance and fairness of LLMs on serialized loan approval datasets from three geographically distinct regions: Ghana, Germany, and the United States. Our evaluation focuses on the model's zero-shot and in-context learning (ICL) capabilities. Our results reveal that the choice of serialization (Serialization refers to the process of converting tabular data into text formats suitable for processing by LLMs.) format significantly affects both performance and fairness in LLMs, with certain formats such as GReat and LIFT yielding higher F1 scores but exacerbating fairness disparities. Notably, while ICL improved model performance by 4.9-59.6% relative to zero-shot baselines, its effect on fairness varied considerably across datasets. Our work underscores the importance of effective tabular data representation methods and fairness-aware models to improve the reliability of LLMs in financial decision-making.

Figures

Figures reproduced from arXiv: 2508.21512 by Deborah D. Kanubala, Dietrich Klakow, Isabel Valera, Israel Abebe Azime, Mario Fritz, Philipp Slusallek, Tejumade Afonja.

Figure 1
Figure 1. Figure 1: Overview of our approach. We first utilize different serialization approaches to acquire our serialized data, and we investigate the LLMs’ performance and fairness by applying zero- and few-shot learning to the datasets. how do these factors vary across different serializa￾tion methods and geographical regions? To address the research questions outlined above, this work makes the following contributions2 :… view at source ↗
Figure 2
Figure 2. Figure 2: Zero-shot weighted average F1 score performance of LLMs on loan approval tasks. Evaluated across three prompts (variation shown by error bars) and multiple table-to-text serialization methods. The Logistic Regression model baseline (green dashed line) uses default JSON serialization with variables as individual features. Most LLMs underperform relative to this baseline, with only GReaT on Ghana, List/Text … view at source ↗
Figure 3
Figure 3. Figure 3: Average weighted F1 score trends across serialization formats for few-shot examples, showing higher gains in U.S. data across formats, while Germany lags consistently despite increasing shot numbers. the other two datasets. We also conducted experi￾ments on model token attribution, which is detailed in the Appendix H. 4.1 Do LLMs Perform Better Than Baseline or Benchmark Models on the Default Serialization… view at source ↗
Figure 4
Figure 4. Figure 4: Few-shot weighted F1 trends. Adding a small number of in-context examples improves performance, while differences among serialization formats remain modest across datasets. ­ Our results show that increasing the natural￾ness of input formatting does not consistently enhance model performance. 4.4 Does Using Few-Shot Examples Improve the Decision-Making Abilities of LLMs? Given LLMs’ subpar performance in t… view at source ↗
Figure 5
Figure 5. Figure 5: Mean difference in EO for different serialization methods and models. Finance-based models show higher gender-based disparity for certain serializations, while the results are highly region and format-dependent. 0.10 for certain serialization methods in the Ghana dataset (see [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: KDE plot comparing age and loan amount distributions across datasets, highlighting inherent socio [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Equality of Opportunity Difference for Few-Shot Learning Across Serialization Methods and Datasets. In-context learning (ICL) does not consistently reduce bias; in some cases, models exhibit significantly unfair behavior, particularly in certain shot configurations. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Statistical Parity Difference Across Serialization Methods and Models This figure illustrates the statistical parity difference for various serialization methods and models. We observe that financial models exhibit notably high bias, particularly for the Ghana and Germany datasets. F Prompt Examples In [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Average F1 Score for Few-Shot Learning Across Different Serialization Methods This figure presents the average F1 scores across various serialization methods for each dataset. We observe that the same models exhibit similar performance trends within each dataset, regardless of format. While the text format of the Ghana dataset may not share characteristics with the text format of the Germany dataset, Ghana… view at source ↗
Figure 10
Figure 10. Figure 10: Attribution scores of Ghana data for example 1. Positive attribution scores are indicated in green, while [PITH_FULL_IMAGE:figures/full_fig_p021_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Attribution scores of Ghana data for example 2. Positive attribution scores are indicated in green, while [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: This figure displays the attribution scores for Example 1 of the Germany dataset. Positive attribution [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: This figure displays the attribution scores for Example 2 of the Germany dataset. Positive attribution [PITH_FULL_IMAGE:figures/full_fig_p024_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: This figure displays the attribution scores for Example 1 of the US dataset. Positive attribution scores [PITH_FULL_IMAGE:figures/full_fig_p025_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: This figure displays the attribution scores for Example 2 of the US dataset. Positive attribution scores [PITH_FULL_IMAGE:figures/full_fig_p026_15.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 17 canonical work pages · 2 internal anchors

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Abubakar Abid, Maheen Farooqi, and James Zou. 2021. Persistent anti-muslim bias in large language models. In Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society, pages 298--306

    work page 2021

  4. [4]

    Rishabh Agarwal, Avi Singh, Lei Zhang, Bernd Bohnet, Luis Rosias, Stephanie Chan, Biao Zhang, Ankesh Anand, Zaheer Abbas, Azade Nova, et al. 2025. Many-shot in-context learning. Advances in Neural Information Processing Systems, 37:76930--76966

    work page 2025

  5. [5]

    Zhang, Bernd Bohnet, Luis Rosias, Stephanie Chan, Biao Zhang, Ankesh Anand, Zaheer Abbas, Azade Nova, John D

    Rishabh Agarwal, Avi Singh, Lei M. Zhang, Bernd Bohnet, Luis Rosias, Stephanie Chan, Biao Zhang, Ankesh Anand, Zaheer Abbas, Azade Nova, John D. Co-Reyes, Eric Chu, Feryal Behbahani, Aleksandra Faust, and Hugo Larochelle. 2024. Many-shot in-context learning. arXiv preprint arXiv:2404.11018

    Pith/arXiv arXiv 2024

  6. [6]

    Carlos Alejandro Aguirre, Kuleen Sasse, Isabel Alyssa Cachola, and Mark Dredze. 2024. https://doi.org/10.18653/v1/2024.nlp4pi-1.4 Selecting shots for demographic fairness in few-shot learning with large language models . In Proceedings of the Third Workshop on NLP for Positive Impact, Miami, Florida, USA. Association for Computational Linguistics

    work page doi:10.18653/v1/2024.nlp4pi-1.4 2024

  7. [7]

    Microsoft Research AI4Science and Microsoft Azure Quantum. 2023. The impact of large language models on scientific discovery: a preliminary study using gpt-4. arXiv preprint arXiv:2311.07361

    Pith/arXiv arXiv 2023

  8. [8]

    Tolga Bolukbasi, Kai-Wei Chang, James Y Zou, Venkatesh Saligrama, and Adam T Kalai. 2016. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. Advances in neural information processing systems, 29

    2016

  9. [9]

    Vadim Borisov, Kathrin Se ler, Tobias Leemann, Martin Pawelczyk, and Gjergji Kasneci. 2022. Language models are realistic tabular data generators. arXiv preprint arXiv:2210.06280

    Pith/arXiv arXiv 2022

  10. [10]

    Han Ding, Yinheng Li, Junhao Wang, and Hang Chen. 2024. Large language model agent in financial trading: A survey. arXiv preprint arXiv:2408.06361

    arXiv 2024

  11. [11]

    Tuan Dinh, Yuchen Zeng, Ruisu Zhang, Ziqian Lin, Michael Gira, Shashank Rajput, Jy-yong Sohn, Dimitris Papailiopoulos, and Kangwook Lee. 2022. Lift: Language-interfaced fine-tuning for non-language machine learning tasks. Advances in Neural Information Processing Systems, 35:11763--11784

    work page 2022

  12. [12]

    Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Reingold, and Richard Zemel. 2012. Fairness through awareness. In Proceedings of the 3rd innovations in theoretical computer science conference, pages 214--226

    2012

  13. [13]

    Duanyu Feng, Yongfu Dai, Jimin Huang, Yifang Zhang, Qianqian Xie, Weiguang Han, Zhengyu Chen, Alejandro Lopez-Lira, and Hao Wang. 2023. Empowering many, biasing a few: Generalist credit scoring through large language models. arXiv preprint arXiv:2310.00566

    Pith/arXiv arXiv 2023

  14. [14]

    Leo Gao, Jonathan Tow, Baber Abbasi, Stella Biderman, Sid Black, Anthony DiPofi, Charles Foster, Laurence Golding, Jeffrey Hsu, Alain Le Noac'h, Haonan Li, Kyle McDonell, Niklas Muennighoff, Chris Ociepa, Jason Phang, Laria Reynolds, Hailey Schoelkopf, Aviya Skowron, Lintang Sutawika, Eric Tang, Anish Thite, Ben Wang, Kevin Wang, and Andy Zou. 2024. https...

    doi:10.5281/zenodo.12608602 2024

  15. [15]

    Nikhil Garg, Londa Schiebinger, Dan Jurafsky, and James Zou. 2018. Word embeddings quantify 100 years of gender and ethnic stereotypes. Proceedings of the National Academy of Sciences, 115(16):E3635--E3644

    2018

  16. [16]

    FM Haque and Md Mahedi Hassan. 2024. Bank loan prediction using machine learning techniques. arXiv preprint arXiv:2410.08886

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  17. [17]

    Moritz Hardt, Eric Price, and Nati Srebro. 2016. Equality of opportunity in supervised learning. Advances in neural information processing systems, 29

    2016

  18. [18]

    Stefan Hegselmann, Alejandro Buendia, Hunter Lang, Monica Agrawal, Xiaoyi Jiang, and David Sontag. 2023. Tabllm: Few-shot classification of tabular data with large language models. In International Conference on Artificial Intelligence and Statistics, pages 5549--5581. PMLR

    work page 2023

  19. [19]

    Noah Hollmann, Samuel M \"u ller, Katharina Eggensperger, and Frank Hutter. 2022. Tabpfn: A transformer that solves small tabular classification problems in a second. arXiv preprint arXiv:2207.01848

    Pith/arXiv arXiv 2022

  20. [20]

    Sukriti Jaitly, Tanay Shah, Ashish Shugani, and Razik Singh Grewal. 2023. Towards better serialization of tabular data for few-shot classification. arXiv preprint arXiv:2312.12464

    Pith/arXiv arXiv 2023

  21. [21]

    Loan approval prediction dataset

    Kaggle . Loan approval prediction dataset. https://www.kaggle.com/altruistdelhite04/loan-prediction-problem-dataset. Accessed: 2024-07-19

    work page 2024

  22. [22]

    Deborah D Kanubala, Isabel Valera, and Kavya Gupta. 2024. Fairness beyond binary decisions: A case study on german credit. European Workshop on Algorithmic Fairness

    work page 2024

  23. [23]

    Siavash Kazemian, Cosmin Munteanu, and Gerald Penn. 2022. https://doi.org/10.18653/v1/2022.finnlp-1.10 A taxonomical NLP blueprint to support financial decision making through information-centred interactions . In Proceedings of the Fourth Workshop on Financial Technology and Natural Language Processing (FinNLP), pages 89--98, Abu Dhabi, United Arab Emira...

    work page doi:10.18653/v1/2022.finnlp-1.10 2022

  24. [24]

    Narine Kokhlikyan, Vivek Miglani, Miguel Martin, Edward Wang, Bilal Alsallakh, Jonathan Reynolds, Alexander Melnikov, Natalia Kliushkina, Carlos Araya, Siqi Yan, et al. 2020. Captum: A unified and generic model interpretability library for pytorch. arXiv preprint arXiv:2009.07896

    Pith/arXiv arXiv 2020

  25. [25]

    Yaxuan Kong, Yuqi Nie, Xiaowen Dong, John M Mulvey, H Vincent Poor, Qingsong Wen, and Stefan Zohren. 2024. Large language models for financial and investment management: Applications and benchmarks. Journal of Portfolio Management, 51(2)

    work page 2024

  26. [26]

    Hadas Kotek, Rikker Dockum, and David Sun. 2023. Gender bias and stereotypes in large language models. In Proceedings of the ACM collective intelligence conference, pages 12--24

    2023

  27. [27]

    Nikita Kozodoi, Johannes Jacob, and Stefan Lessmann. 2022. Fairness in credit scoring: Assessment, implementation and profit implications. European Journal of Operational Research, 297(3):1083--1094

    work page 2022

  28. [28]

    Ayan Majumdar, Deborah D Kanubala, Kavya Gupta, and Isabel Valera. 2025. A causal framework to measure and mitigate non-binary treatment discrimination. arXiv preprint arXiv:2503.22454

    work page arXiv 2025

  29. [29]

    Meta. 2024. I ntroducing M eta L lama 3: T he most capable openly available L L M to date --- ai.meta.com. https://ai.meta.com/blog/meta-llama-3/. [Accessed 01-06-2024]

    work page 2024

  30. [30]

    Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Nick Barnes, and Ajmal Mian. 2023. A comprehensive overview of large language models. arXiv preprint arXiv:2307.06435

    Pith/arXiv arXiv 2023

  31. [31]

    Roberto Navigli, Simone Conia, and Bj \"o rn Ross. 2023. Biases in large language models: origins, inventory, and discussion. ACM Journal of Data and Information Quality, 15(2):1--21

    work page 2023

  32. [32]

    Frank Gyimah Sackey and Peter Nkrumah Amponsah. 2018. Gender discrimination in commercial banks’ credit markets in ghana: a decomposition and counterfactual analysis. African Journal of Business and Economic Research, 13(2):121--140

    work page 2018

  33. [33]

    Ananya Singha, Jos \'e Cambronero, Sumit Gulwani, Vu Le, and Chris Parnin. 2023. Tabular representation, noisy operators, and impacts on table structure understanding tasks in llms. arXiv preprint arXiv:2310.10358

    Pith/arXiv arXiv 2023

  34. [34]

    Statlog (german credit data)

    Statlog. Statlog (german credit data). https://archive.ics.uci.edu/ml/datasets/Statlog+ Accessed: 2024-07-19

    work page 2024

  35. [35]

    Yuan Sui, Mengyu Zhou, Mingjie Zhou, Shi Han, and Dongmei Zhang. 2024. Table meets llm: Can large language models understand structured table data? a benchmark and empirical study. In Proceedings of the 17th ACM International Conference on Web Search and Data Mining, pages 645--654

    2024

  36. [36]

    Tony Sun, Andrew Gaut, Shirlyn Tang, Yuxin Huang, Mai ElSherief, Jieyu Zhao, Diba Mirza, Elizabeth Belding, Kai-Wei Chang, and William Yang Wang. 2019. Mitigating gender bias in natural language processing: Literature review. arXiv preprint arXiv:1906.08976

    Pith/arXiv arXiv 2019

  37. [37]

    Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, Pouya Tafti, Léonard Hussenot, Pier Giuseppe Sessa, Aakanksha Chowdhery, Adam Roberts, Aditya Barua, Alex Botev, Alex Castro-Ros, Ambrose Slone, Amélie Héliou, Andrea Tacchetti, Anna Bulanova, Anto...

    Pith/arXiv arXiv 2024

  38. [38]

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. 2023. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288

    Pith/arXiv arXiv 2023

  39. [39]

    Qianqian Xie, Weiguang Han, Xiao Zhang, Yanzhao Lai, Min Peng, Alejandro Lopez-Lira, and Jimin Huang. 2023. https://arxiv.org/abs/2306.05443 Pixiu: A large language model, instruction data and evaluation benchmark for finance . Preprint, arXiv:2306.05443

    Pith/arXiv arXiv 2023

  40. [40]

    Miaoran Zhang, Vagrant Gautam, Mingyang Wang, Jesujoba O Alabi, Xiaoyu Shen, Dietrich Klakow, and Marius Mosbach. 2024. The impact of demonstrations on multilingual in-context learning: A multidimensional analysis. arXiv preprint arXiv:2402.12976

    work page internal anchor Pith review Pith/arXiv arXiv 2024

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.