Pith. sign in

REVIEW 4 major objections 5 minor 40 references

LLMs on Trial: Evaluating Judicial Fairness for Large Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read No tested LLM passes a judicial-fairness test on Chinese criminal sentencing tasks, according to a 16-model evaluation.

desk verdict A serious, large-scale audit of legal-LLM fairness whose headline overstates the effect sizes; worth refereeing after effect-size and manipulation checks. read the letter →

arxiv 2507.10852 v2 pith:LE4UJQK4 submitted 2025-07-14 cs.CL

Yiran Hu , Zongyue Xue , Haitao Li , Siyuan Zheng , Qingjing Chen , Shaochun Wang , Xihan Zhang , Ning Zheng
show 5 more authors
Yun Liu Qingyao Ai Yiqun Liu Charles L.A. Clarke Weixing Shen
This is my paper
classification cs.CL
keywords LLMfairnessjudicialcounterfactualpromptingsentencingpredictionChinesecriminallawbiasevaluationimbalancedinaccuracy
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that large language models cannot currently be trusted to judge criminal cases fairly, at least in Chinese sentencing. It builds a fairness framework separating substance from procedure and demographic from non-demographic factors, then tests 16 LLMs on JudiFair, a dataset of 177,100 counterfactual case facts created by swapping 65 legally irrelevant labels one at a time. Using three metrics—inconsistency, bias, and imbalanced inaccuracy—the authors find that every model changes its sentence when only the swapped label changes, most models show statistically systematic bias, and prediction errors are unevenly distributed across groups. If the result holds, general-domain fairness scores tell us little about whether an LLM actually treats people equally in high-stakes legal settings.

What carries the argument

The evaluation is built on counterfactual prompting with trigger sentences: for each of 65 labels and its 161 values, the authors locate the sentence in a real judicial document that expresses that attribute, replace it with an alternative value, and tell the model, when needed, that this sentence "shall prevail" over any conflicting description. Each case-label-value combination becomes one prompt, so each model sees one alternative at a time. The three metrics are weighted inconsistency, regression-based bias with document fixed effects on $\ln(\text{Sentence}+1)$ and cluster-robust errors, and imbalanced inaccuracy from the same fixed-effect regression applied to absolute prediction error; a Bernoulli test over the label values separates systematic patterns from random noise.

What would settle it

Remove or invert the "this one shall prevail" instruction, or track the model's reasoning, and rerun the same 16 models on the same counterfactual cases; if most measured bias and inconsistency persists under the weakened instruction, the effects are not caused by the manipulated label in the way the paper assumes.

Watch

Extended reading notes

Core claim

The paper's claim is that judicial unfairness in LLMs is pervasive, systematic, and measurable. When a single non-legal attribute is changed while all other case facts stay fixed, the models change their output on a substantial share of documents; fixed-effect regressions find many label values with statistically significant effects on predicted sentence length; and the same style of regression on absolute prediction error shows imbalanced inaccuracy across groups. The authors report that demographic attributes trigger the strongest biases, that procedure factors such as court level and compulsory measures are more biased than substance factors, and that the direction of LLM bias tends to mirror bias documented in human sentencing—for example, female defendants treated more leniently and rural-hukou defendants punished more harshly. They also report two correlations: more random outputs mask bias but increase inconsistency, and models with lower prediction error tend to show more bias, suggesting LLMs learn the systematic skews of the real-world data.

Load-bearing premise

The load-bearing premise is that the model actually obeys the added trigger sentence and reinterprets only that one attribute; if the model instead weights the rest of the case or hedges when the trigger conflicts, the measured inconsistency and bias are not cleanly attributable to the swapped label.

Editorial extensions

If this is right

  • No model in the sample can be treated as judicially fair on Chinese criminal sentencing tasks without mitigation.
  • Fairness scores from general-domain benchmarks cannot substitute for judicial-specific, multi-label audits.
  • Legally irrelevant procedure factors such as court level and compulsory measures shift LLM sentences, so evaluations that ignore procedure understate unfairness.
  • The positive correlation between prediction accuracy and bias means sentencing-prediction benchmarks may reward models that internalize human sentencing disparities.
  • Raising temperature masks bias and imbalanced inaccuracy but worsens inconsistency, so it is not a fairness fix.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If labels that never appear in Chinese judicial records, such as sexual orientation or judge religion, still drive bias, the bias likely comes from pretraining stereotypes rather than legal corpora; a direct test would probe the same models on neutral non-legal tasks with the same attribute swaps.
  • The accuracy-bias correlation suggests a concrete extension: hold predictive accuracy fixed while fine-tuning with a fairness penalty on the 65 labels, and see whether the bias count drops without raising mean absolute error.
  • The framework should transfer to other jurisdictions, but the label set is Chinese-law-specific; embedding the same three metrics in multilingual legal corpora would show whether procedure-factor bias is a general LLM phenomenon or an artifact of Chinese case formats.
  • A practical consequence left implicit: any legal-technology product using these models for sentencing assistance should report label-level fairness metrics, not just overall accuracy, before deployment.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper develops a framework for auditing judicial fairness of large language models in Chinese criminal sentencing. It constructs JudiFair, a dataset of 177,100 counterfactually generated case facts derived from 1,100 judicial documents, covering 65 labels and 161 label values, and evaluates 16 LLMs with three metrics: inconsistency, bias, and imbalanced inaccuracy. The statistical analysis uses document fixed effects, cluster-robust standard errors, and Bernoulli tests to claim that LLMs exhibit pervasive and severe judicial unfairness, with stronger bias on demographic labels and a positive correlation between accuracy and bias. A public toolkit and dataset are promised for future evaluations.

Significance. The paper makes a substantial empirical contribution in scale and scope: 65 legally motivated labels, 177,100 counterfactual case facts, 16 models, and a publicly released toolkit are valuable resources for LLM fairness auditing. The use of document fixed effects to compare counterfactual variants of the same case is a sound design choice, and the multiple robustness checks in Appendix E.4 strengthen confidence in the regression results. However, the central causal claim depends on an unvalidated assumption about how LLMs process conflicting trigger sentences, and the headline conclusion of 'severe' unfairness is not connected to any effect-size threshold. These issues are fixable but require additional analysis, so the result as currently stated is stronger than the evidence supports.

major comments (4)
  1. [Appendix C.1, D.3; Eq. (2), Eq. (4), Eq. (1)] The causal interpretation of all three metrics depends on the assumption that inserting or replacing a trigger sentence changes only the manipulated label in the model's internal representation. The paper provides no direct validation of this assumption. For most labels (e.g., defendant gender, judge age, defender ethnicity), the trigger sentence can contradict facts elsewhere in the original judicial document, and the override instruction 'this one shall prevail' appears explicitly only for crime location, crime date, and crime time in Tables A9 and A10. If the model instead treats the contradiction as a signal to adjust other beliefs, or simply produces different outputs because the prompt is internally inconsistent, then the coefficients in Eqs. (2) and (4) and the inconsistency proportions in Eq. (1) conflate label effects with instruction-following and contradiction artifacts. Please add a manipulation check, such as repeated identical prompts at temperature 0, paraphrased triggers with the same label value, or a control condition inserting a contradictory but legally irrelevant sentence, and report the extent to which outcomes change under those controls.
  2. [Table A17, Table A21, Eq. (2); abstract and Section 6.2] The headline 'severe LLM judicial unfairness' is a magnitude claim, but the bias metric counts any label value with p <= 0.1 regardless of effect size. Many of the significant coefficients are very small: for example, Glm4's defendant-gender female coefficient is -0.028 (about 2.8% sentence length) with p = 0.012, and its defendant household registration Not Local coefficient is 0.010 with p = 0.028. A statistically detectable 1-3% difference may be important, but it does not by itself establish 'severe' unfairness. Please report effect sizes systematically (e.g., percentage sentence differences, standardized coefficients, or the share of labels exceeding a pre-specified magnitude threshold) and calibrate the abstract and conclusion claims to those magnitudes.
  3. [Eq. (3), Sections 5.2.2 and 5.2.3] The Bernoulli test treats each label-value test as an independent Bernoulli trial and then aggregates across label values and across all LLMs. Within a model, the tests share the same judicial documents and the same model outputs, so they are not independent; across models, the prompts and datasets are also not independent. This can substantially overstate the evidence for 'systematic' bias. Please report the full distribution of p-values, use an appropriate multiple-testing correction, or otherwise justify the independence assumption. The text also says the tests cover '96 label values across 65 labels' while Tables A17 and A18 report different denominators for bias and imbalanced inaccuracy; the exact N used in each Bernoulli test should be stated.
  4. [Section 5.2.1, Eq. (1)] The inconsistency metric counts any change in output when a label value changes, but it does not distinguish a meaningful change in the sentencing decision from a formatting artifact or a response to a contradictory prompt. The prompt evaluation in Appendix C.2 checks format adherence on 420 queries, but it does not establish that temperature-0 repeated identical queries produce identical parsed sentences for the full set of 16 models. Given that the inconsistency numbers in Table A17 are a central part of the fairness claim, please report identical-input stability at temperature 0 for each model and, if possible, disaggregate inconsistency by whether the changed output reflects a numerically different sentence versus a parse/format change.
minor comments (5)
  1. [Eq. (3)] The exponent of (1 - tau) in Eq. (3) is printed as L - l, but it should presumably be N - l; as written, the formula is dimensionally inconsistent with the stated binomial test.
  2. [Table A16] The trigger sentence for 'Recusal applied' is identical to the trigger sentence for 'Trial duration' ('The case was concluded shortly. / The case was concluded after a prolonged duration.'), which appears to be a copy-paste error and should be corrected because it affects the validity of that label.
  3. [Tables A17 and A18] Table A17 lists 15 models and Table A18 lists 13 models, but the text says the experiment includes 16 LLMs and reports results for '15 models with a temperature of 0'. The sample sizes for each temperature setting should be clarified, and the missing rows should be reconciled.
  4. [Tables A2-A16] There are several duplicated rows and typographical errors: 'poltical' in Table A2, 'Prosecurate' throughout, duplicated 'Defender age' and 'Prosecurate age' rows in Tables A10-A12, duplicated 'Pretrial conference' rows in Table A4, and duplicated 'Open trial' rows in Table A15. These should be cleaned up.
  5. [Section 6.2] The phrase 'restuls' should read 'results', and the notation alternates between 'consistency' and 'inconsistency' in the main text and Appendix G.1; please standardize the terminology.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the headline metrics are measured directly from LLM outputs on counterfactual prompts, and the only self-citation (to LEEC) supports data provenance rather than the conclusion.

full rationale

The paper's central claims are empirical measurements rather than derivations that reduce to their inputs. Inconsistency (Equation 1), bias (Equation 2), and imbalanced inaccuracy (Equation 4) are defined directly on the LLM's sentencing outputs before and after a single counterfactual label change; no parameter is fitted to the target quantities and then renamed as a prediction. The outcome variable is the model's own sentence, and the comparison is made against real human sentences, so the unfairness findings are not forced by construction. The only salient self-reference is the use of LEEC (Xue et al., 2024), co-authored by current authors, for the underlying judicial documents and 13 of the 65 labels; the paper explicitly states that the label system 'expands upon the LEEC dataset' and that LEEC's role is to provide reliable case data. This is data provenance, not a circular derivation: the conclusion that LLMs are judicially unfair is not assumed by the dataset construction, and the experiments are held-out counterfactual predictions evaluated against external human judgments. Possible validity concerns, such as whether the trigger sentences actually isolate the manipulated attribute in the LLM's internal reasoning, are empirical assumptions about the measurement instrument, not circularity; they would be correctness risks at most.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claims do not introduce new physical or conceptual entities; the metrics are operational definitions of existing fairness notions. The main borrowed components are the LEEC dataset (with author overlap) and the human sentencing gold standard.

free parameters (2)
  • Significance threshold tau for per-label tests = 0.1 and 0.05
    Chosen conventionally for Bernoulli tests in Section 5.2.2; not fitted, but any threshold affects reported Bias No. and unfair inaccuracy counts.
  • Age counterfactual exclusion window = 10 years above or below original age
    Data-generation rule in Tables A6, A7, A10, A12, and A13 to avoid implausible ages; affects the spacing of age values and therefore measured age effects.
assumptions (5)
  • standard math The fixed-effect linear regression model (Eq. 2) adequately maps the counterfactual dataset to label effects; linearity and no interference across observations.
    Invoked in Section 5.2.2; clustered robust standard errors mitigate some violations, but the causal interpretation rests on this model.
  • standard math Bernoulli trials across labels are independent and identically distributed under the null.
    Needed for Eq. 3 in Section 5.2.2; labels are correlated within a model because they share documents and model internals, so the Bernoulli p-values are approximate.
  • domain assumption Labels and trigger sentences chosen by the legal experts are genuinely extra-legal in the cases kept; the exclusion rules remove all cases where the factor is legally relevant.
    Stated in Section 4.2.2 and Appendix D.3; if exclusions are incomplete, significant coefficients may reflect legally required differences, not unfairness.
  • domain assumption Actual human sentencing outcomes from LEEC are an appropriate gold standard for computing imbalanced inaccuracy.
    Section 5.1 metric 3; if human sentences encode their own biases (the paper itself says LLM biases mirror real-world judicial biases in Section 6.2), then deviations from them are not necessarily unfair.
  • ad hoc to paper LLMs follow the instruction that trigger sentences override conflicting facts in the case text.
    Appendix C.1 instructs models to prioritize trigger sentences; all causal attribution in inconsistency and bias relies on this being effective.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LLMs on Trial: Evaluating Judicial Fairness for Large Language Models." pith.science (2026). https://pith.science/paper/LE4UJQK4

@misc{pith2026250710852,
  author       = {Pith},
  title        = {Pith review of: LLMs on Trial: Evaluating Judicial Fairness for Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LE4UJQK4}},
  note         = {Machine review of arXiv:2507.10852}
}
read the original abstract

Large Language Models (LLMs) are increasingly used in high-stakes fields where their decisions impact rights and equity. However, LLMs' judicial fairness and implications for social justice remain underexplored. When LLMs act as judges, the ability to fairly resolve judicial issues is a prerequisite to ensure their trustworthiness. Based on theories of judicial fairness, we construct a comprehensive framework to measure LLM fairness, leading to a selection of 65 labels and 161 corresponding values. Applying this framework to the judicial system, we compile an extensive dataset, JudiFair, comprising 177,100 unique case facts. To achieve robust statistical inference, we develop three evaluation metrics, inconsistency, bias, and imbalanced inaccuracy, and introduce a method to assess the overall fairness of multiple LLMs across various labels. Through experiments with 16 LLMs, we uncover pervasive inconsistency, bias, and imbalanced inaccuracy across models, underscoring severe LLM judicial unfairness. Particularly, LLMs display notably more pronounced biases on demographic labels, with slightly less bias on substance labels compared to procedure ones. Interestingly, increased inconsistency correlates with reduced biases, but more accurate predictions exacerbate biases. While we find that adjusting the temperature parameter can influence LLM fairness, model size, release date, and country of origin do not exhibit significant effects on judicial fairness. Accordingly, we introduce a publicly available toolkit containing all datasets and code, designed to support future research in evaluating and improving LLM fairness.

Figures

Figures reproduced from arXiv: 2507.10852 by the authors.

Figure 1
Figure 1. Classification of LLM fairness. Judicial fairness poses unique challenges for current LLMs. As [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Framework of LLM judicial fairness. 3 [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparison of imbalanced inaccuracy and bias across scenarios. In Scenario 1, LLMs [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Evaluation framework of LLM judicial fairness. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 22 canonical work pages

  1. [1]

    On that date,The Supreme People’s Court Pro- visions on People’s Courts Release of Judgments on the Internetcame into effect, mandating the public release of most adjudications

    60 Working Paper E.4.4 REGRESSIONSEXCLUDINGCASESFILED BEFORE2014 We exclude cases filed before January 1, 2014, to mitigate potential selection bias stemming from non-systematic disclosure of judicial documents. On that date,The Supreme People’s Court Pro- visions on People’s Courts Release of Judgments on the Internetcame into effect, mandating the publi...

    work page 2014

  2. [3]

    Can few-shot work in long- context? recycling the context to generate demonstrations.arXiv preprint arXiv:2406.13632,

    Arie Cattan, Alon Jacovi, Alex Fabrikant, Jonathan Herzig, Roee Aharoni, Hannah Rashkin, Dror Marcus, Avinatan Hassidim, Yossi Matias, Idan Szpektor, et al. Can few-shot work in long- context? recycling the context to generate demonstrations.arXiv preprint arXiv:2406.13632,

    arXiv

  3. [4]

    Linear models with high-dimensional fixed effects: An efficient and feasible estima- tor.Unpublished manuscript, http://scorreia

    Sergio Correia. Linear models with high-dimensional fixed effects: An efficient and feasible estima- tor.Unpublished manuscript, http://scorreia. com/research/hdfe. pdf (last accessed 25 October 2019), 4(2),

    work page 2019

  4. [5]

    Jeffrey Dastin

    URL https://www.amazon.science/ publications/fairness-measures-for-machine-learning-in-finance. Jeffrey Dastin. Amazon scraps secret ai recruiting tool that showed bias against women. reuters (2018),

    work page 2018

  5. [8]

    Fairness-Aware Multi-Group Target Detection in Online Discussion

    URL https://arxiv.org/abs/2407.11933. Version 2, submitted June

    work page Pith review arXiv

  6. [10]

    Better zero-shot reasoning with role-play prompting.arXiv preprint arXiv:2308.07702,

    Aobo Kong, Shiwan Zhao, Hao Chen, Qicheng Li, Yong Qin, Ruiqi Sun, Xin Zhou, Enzhi Wang, and Xiaohang Dong. Better zero-shot reasoning with role-play prompting.arXiv preprint arXiv:2308.07702,

    arXiv

  7. [11]

    Decoding biases: Automated methods and llm judges for gender bias detection in language models.arXiv preprint arXiv:2408.03907,

    Shachi H Kumar, Saurav Sahay, Sahisnu Mazumder, Eda Okur, Ramesh Manuvinakurike, Nicole Beckage, Hsuan Su, Hung-yi Lee, and Lama Nachman. Decoding biases: Automated methods and llm judges for gender bias detection in language models.arXiv preprint arXiv:2408.03907,

    arXiv

  8. [12]

    Open Models, Closed Minds? On Agents Capabilities in Mimicking Human Personalities through Open Large Language Models

    Lucio La Cava and Andrea Tagarelli. Open models, closed minds? on agents capabilities in mimick- ing human personalities through open large language models.arXiv preprint arXiv:2401.07115,

    work page Pith review arXiv

Show all 40 references
  1. [13]

    Muser: A multi-view similar case retrieval dataset

    13 Working Paper Qingquan Li, Yiran Hu, Feng Yao, Chaojun Xiao, Zhiyuan Liu, Maosong Sun, and Weixing Shen. Muser: A multi-view similar case retrieval dataset. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management, pp. 5336–5340, 2023a....

    arXiv

  2. [14]

    Victim age and capital sentencing outcomes in north carolina (1977–2009).Criminal justice studies, 31(1):62–79,

    Christopher J Marier, John K Cochran, M Dwayne Smith, Sondra J Fogel, and Beth Bjerregaard. Victim age and capital sentencing outcomes in north carolina (1977–2009).Criminal justice studies, 31(1):62–79,

    1977

  3. [17]

    A compara- tive study of prompting strategies for legal text classification

    Ali Hakimi Parizi, Yuyang Liu, Prudhvi Nokku, Sina Gholamian, and David Emerson. A compara- tive study of prompting strategies for legal text classification. InProceedings of the Natural Legal Language Processing Workshop 2023, pp. 258–265,

    2023

  4. [19]

    doi: 10.18653/v1/2022.findings-acl.165

    Association for Computational Linguistics. doi: 10.18653/v1/2022.findings-acl.165. URL https://aclanthology.org/2022.findings-acl.165. Andrea Pinto, Tomer Galanti, and Randall Balestriero. The fair language model paradox.arXiv preprint arXiv:2410.11985,

    2022 arXiv

  5. [21]

    A comprehensive survey of bias in llms: Current landscape and future directions.arXiv preprint arXiv:2409.16430,

    Rajesh Ranjan, Shailja Gupta, and Surya Narayan Singh. A comprehensive survey of bias in llms: Current landscape and future directions.arXiv preprint arXiv:2409.16430,

    arXiv

  6. [23]

    The power of prompts: Evaluating and mitigating gender bias in mt with llms.arXiv preprint arXiv:2407.18786,

    14 Working Paper Aleix Sant, Carlos Escolano, Audrey Mash, Francesca De Luca Fornaciari, and Maite Melero. The power of prompts: Evaluating and mitigating gender bias in mt with llms.arXiv preprint arXiv:2407.18786,

    arXiv

  7. [25]

    Cail2018: A large-scale legal dataset for judg- ment prediction.arXiv preprint arXiv:1807.02478,

    Chaojun Xiao, Haoxi Zhong, Zhipeng Guo, Cunchao Tu, Zhiyuan Liu, Maosong Sun, Yansong Feng, Xianpei Han, Zhen Hu, Heng Wang, et al. Cail2018: A large-scale legal dataset for judg- ment prediction.arXiv preprint arXiv:1807.02478,

    arXiv

  8. [26]

    Walking in others’ shoes: How perspective-taking guides large language models in reducing toxicity and bias.arXiv preprint arXiv:2407.15366,

    Rongwu Xu, Zi’an Zhou, Tianwei Zhang, Zehan Qi, Su Yao, Ke Xu, Wei Xu, and Han Qiu. Walking in others’ shoes: How perspective-taking guides large language models in reducing toxicity and bias.arXiv preprint arXiv:2407.15366,

    arXiv

  9. [27]

    Leven: A large-scale chinese legal event detection dataset

    Feng Yao, Chaojun Xiao, Xiaozhi Wang, Zhiyuan Liu, Lei Hou, Cunchao Tu, Juanzi Li, Yun Liu, Weixing Shen, and Maosong Sun. Leven: A large-scale chinese legal event detection dataset. arXiv preprint arXiv:2203.08556,

    arXiv

  10. [28]

    Correcting negative bias in large language models through negative attention score alignment.arXiv preprint arXiv:2408.00137,

    Sangwon Yu, Jongyoon Song, Bongkyu Hwang, Hoyoung Kang, Sooah Cho, Junhwa Choi, Seongho Joe, Taehee Lee, Youngjune L Gwon, and Sungroh Yoon. Correcting negative bias in large language models through negative attention score alignment.arXiv preprint arXiv:2408.00137,

    arXiv

  11. [29]

    Evaluation ethics of llms in legal domain.arXiv preprint arXiv:2403.11152, 2024a

    Ruizhe Zhang, Haitao Li, Yueyue Wu, Qingyao Ai, Yiqun Liu, Min Zhang, and Shaoping Ma. Evaluation ethics of llms in legal domain.arXiv preprint arXiv:2403.11152, 2024a. Yifan Zhang. Meta prompting for agi systems.arXiv preprint arXiv:2311.11482,

    arXiv

  12. [30]

    Climb: A benchmark of clinical bias in large language models.arXiv preprint arXiv:2407.05250, 2024b

    Yubo Zhang, Shudi Hou, Mingyu Derek Ma, Wei Wang, Muhao Chen, and Jieyu Zhao. Climb: A benchmark of clinical bias in large language models.arXiv preprint arXiv:2407.05250, 2024b. Hui Zhong, Songsheng Chen, and Mian Liang. Gender bias of llm in economics: An existentialism pers...

    arXiv

  13. [31]

    17 A.2 Legal Datasets

    15 Working Paper TABLE OFCONTENTS FORAPPENDIX A Related Works (Detailed) 17 A.1 Fairness Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 A.2 Legal Datasets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 B Label Sy...

    2022

  14. [32]

    GAP, developed by (Webster et al., 2018), provides 8,908 ambiguous pronoun-name pairs to evaluate gender bias in coreference resolution tasks

    evaluates gender stereotypes using a collection of 3,160 sentences that cover 40 different professions. GAP, developed by (Webster et al., 2018), provides 8,908 ambiguous pronoun-name pairs to evaluate gender bias in coreference resolution tasks. At the same time, other resear...

    2024

  15. [33]

    However, its annotations merely cover legal articles, charges, and prison terms, without providing detailed facts of the cases

    contains over 2.6 million criminal cases published by the Supreme People’s Court of China. However, its annotations merely cover legal articles, charges, and prison terms, without providing detailed facts of the cases. LEVEN (Yao et al., 2022), on the other hand, is a large-sc...

    2022

  16. [34]

    As pointed out by Ulmer in 2012, the practical application of the law is significantly influenced not only by legal factors but also by extra-legal ones

    is another Chinese legal dataset consisting of 15,919 legal documents and 155 extra-legal factor labels. As pointed out by Ulmer in 2012, the practical application of the law is significantly influenced not only by legal factors but also by extra-legal ones. The comprehensive ...

    2012

  17. [35]

    demonstrated that leaving space for the LLM to generate its own responses is critical for bias detec- tion tasks. To strike a balance, and inspired by the minimalist approach of Meta Prompting (Zhang, 2023), we aim to limit the tokens in model outputs through format restrictio...

    2023

  18. [36]

    Ignore your identity as an AI... You are now a judge proficient in Chinese law

    highlighted that LLMs are sensitive to similar cases and may overlook differences in trigger sentences when queries are kept within a sin- gle context. To address these challenges, we adopt a Few-Shot technique that focuses on providing an example output without including exam...

    2023

  19. [37]

    Avg MAEWt

    Index Model Inconsistency Bias No.Biasp-value(10%) Biasp-value(5%)Wt. Avg MAEWt. AvgMAPEUnfairInacc.No. Unfair Inacc.p-value (10%)Unfair Inacc.p-value (5%) 1 DeepSeek R1-32BQwen 0.740 13 0.010 0.018 48.924 148.945 10 0.325 0.0942 DeepSeek V3 0.657 11 0.161 0.051 49.490 131.416...

  20. [39]

    61 Working Paper Model Name Label Category Label Number Biased Label Number DeepSeek R1 32B Substance label 25 7 DeepSeek R1 32B Procedural label 40 12 DeepSeek v3 Substance label 25 3 DeepSeek v3 Procedural label 40 11 Gemini 1.5 8B Substance label 25 11 Gemini 1.5 8B Procedu...

    2014

  21. [40]

    accuracy- equity trade-off

    Bottom-left panel (Weighted Average MAE vs. Bias Number):The x-axis represents the Bias Number, while the y-axis represents the Weighted Average Mean Absolute Error (MAE). There is a clear negative correlation (p-value = 0.004), indicating that models with more biases tend to ...

    2021

  22. [41]

    There are 12 data points in each panel, corresponding to the 12 models that were evaluated under both temperature settings

    and key fairness-related metrics: inconsistency, bias number, and unfair inaccuracy number. There are 12 data points in each panel, corresponding to the 12 models that were evaluated under both temperature settings. The correspondingp-value for each regression is annotated wit...

    2025

  23. [1971]

    Gender bias in coreference resolution.arXiv preprint arXiv:1804.09301,

    Rachel Rudinger, Jason Naradowsky, Brian Leonard, and Benjamin Van Durme. Gender bias in coreference resolution.arXiv preprint arXiv:1804.09301,

    arXiv

  24. [1991]

    Legal prompt engineering for multilingual legal judgement prediction.arXiv preprint arXiv:2212.02199,

    Dietrich Trautmann, Alina Petrova, and Frank Schilder. Legal prompt engineering for multilingual legal judgement prediction.arXiv preprint arXiv:2212.02199,

    arXiv

  25. [2006]

    Examining gender and race bias in two hundred sentiment analysis systems.arXiv preprint arXiv:1805.04508,

    Svetlana Kiritchenko and Saif M Mohammad. Examining gender and race bias in two hundred sentiment analysis systems.arXiv preprint arXiv:1805.04508,

    arXiv

  26. [2007]

    Who is gpt-3? an exploration of personal- ity, values and demographics.arXiv preprint arXiv:2209.14338,

    Maril`u Miotto, Nicola Rossberg, and Bennett Kleinberg. Who is gpt-3? an exploration of personal- ity, values and demographics.arXiv preprint arXiv:2209.14338,

    arXiv

  27. [2010]

    Perturbation augmentation for fairer nlp.arXiv preprint arXiv:2205.12586,

    Rebecca Qian, Candace Ross, Jude Fernandes, Eric Smith, Douwe Kiela, and Adina Williams. Perturbation augmentation for fairer nlp.arXiv preprint arXiv:2205.12586,

    arXiv

  28. [2018]

    Questioning biases in case judgment summaries: Legal datasets or large language models?arXiv preprint arXiv:2312.00554,

    Aniket Deroy and Subhankar Maity. Questioning biases in case judgment summaries: Legal datasets or large language models?arXiv preprint arXiv:2312.00554,

    arXiv

  29. [2021]

    Llm stability: A detailed analysis with some surprises.arXiv preprint arXiv:2408.04667,

    Berk Atil, Alexa Chittams, Liseng Fu, Ferhan Ture, Lixinyu Xu, and Breck Baldwin. Llm stability: A detailed analysis with some surprises.arXiv preprint arXiv:2408.04667,

    arXiv

  30. [2022]

    Reasoning beyond bias: A study on counterfactual prompting and chain of thought reasoning.arXiv preprint arXiv:2408.08651,

    Kyle Moore, Jesse Roberts, Thao Pham, and Douglas Fisher. Reasoning beyond bias: A study on counterfactual prompting and chain of thought reasoning.arXiv preprint arXiv:2408.08651,

    arXiv

  31. [2023]

    BBQ: A hand-built bias benchmark for ques- tion answering

    Alicia Parrish, Angelica Chen, Nikita Nangia, Vishakh Padmakumar, Jason Phang, Jana Thomp- son, Phu Mon Htut, and Samuel Bowman. BBQ: A hand-built bias benchmark for ques- tion answering. InFindings of the Association for Computational Linguistics: ACL 2022, pp. 2086–2105, Dub...

    2022

  32. [2024]

    Measuring political bias in large language models: What is said and how it is said.arXiv preprint arXiv:2403.18932,

    Yejin Bang, Delong Chen, Nayeon Lee, and Pascale Fung. Measuring political bias in large language models: What is said and how it is said.arXiv preprint arXiv:2403.18932,

    arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.