Pith. sign in

REVIEW 4 major objections 5 minor 39 references

Veracity Bias and Beyond: Uncovering LLMs' Hidden Beliefs in Problem-Solving Reasoning

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Large language models, even after alignment training, systematically associate solution correctness with demographic identity rather than with solution content.

desk verdict The evaluation-bias finding is solid, but the attribution metric's 'more incorrect' claim is a forced-choice artifact; fix the design or the claim, and the citations need verification. read the letter →

arxiv 2505.16128 v2 pith:5SAE7BOT submitted 2025-05-22 cs.CL

Yue Zhou , Barbara Di Eugenio This is my paper
classification cs.CL
keywords veracitybiasattributionevaluationdemographiclargelanguagemodelsreasoningLLMalgorithmicfairness
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 claims that large language models, despite being aligned against demographic stereotypes, systematically associate solution correctness with demographic identity. It defines two measurable phenomena: Attribution Bias, in which models assign known-correct solutions to some groups and known-incorrect ones to others, and Evaluation Bias, in which the same solution receives different scores depending on the stated author. Across five aligned LLMs and four reasoning domains, the authors observe that African-American labels are consistently tied to incorrect solutions in math and coding, and that Asian authorship is least preferred in essay evaluation. They also show that models generate stereotypical colors for racial groups in code without any social prompt, and that rationales can be contradictory to the biased decision. The paper argues that these biases appear without persona or social-context provocation, indicating that demographic bias is embedded in reasoning itself.

What carries the argument

The machinery is the paired probe: an Attribution prompt that presents one known-correct and one known-incorrect solution and forces the model to assign each to one of three demographic options (either explicit labels or race-associated names), and an Evaluation prompt that presents an identical solution under different demographic completers and asks for a score. Bias is quantified with four metrics: Correctness Attribution Bias (ABcor) and Incorrectness Attribution Bias (ABinc), which look at the maximum difference in assignment probability between correct and incorrect solutions for any group, and Evaluation Inconsistency (EI) and Evaluation Preference (EP), which measure how often the same solution changes score across groups and which pair shows the largest score difference. All runs use temperature 0, and a random perturbation baseline with neutral placeholders is used to separate demographic effects from intrinsic randomness. This design aims to expose a hidden belief without invoking personas or social scenarios.

What would settle it

For a fixed math problem, run the attribution prompt twice: the first time presenting the correct solution as Solution 1, the second time presenting the same correct solution as Solution 2, keeping the demographic options fixed. If the demographic assignment flips with the presented veracity label, the Attribution Bias is confirmed as veracity-driven; if the assignment is unchanged, the bias is about the content itself, not perceived veracity.

Watch

Extended reading notes

Core claim

The paper's central discovery is that LLMs hold an internal veracity-demographic association: they are more likely to attribute a known-correct solution to White or Asian authors and a known-incorrect solution to Black authors in math and coding tasks, even when the same solution is presented neutrally. The bias is measurable statistically and persists across all five tested models, with two of the tested models showing extreme lows in attributing correct code to Black groups. In evaluation, models change their score of an identical solution when the author's name or ethnicity is altered, and the strongest preference appears in writing, where Hispanic-authored essays beat Asian-authored ones. A supplementary experiment shows the same models automatically pick stereotypical colors (light for White, dark for Black, gold for Asian, brown for Hispanic) in generated visualization code. The authors take this as evidence that demographic stereotypes are embedded in the models' reasoning process rather than being a surface output of explicit social-context prompts.

Load-bearing premise

The paper assumes that when a model assigns a known-correct solution to a demographic label, the choice reveals its internal belief about that group's competence, rather than a statistical association with the name or a pattern-matching artifact of the prompt.

Editorial extensions

If this is right

  • Automated grading systems built on LLMs will systematically under-score or over-score students based on perceived race or gender, even for identical submitted work.
  • Tutoring and solution-generation systems may route correct solutions toward some demographic groups and incorrect ones toward others, widening achievement gaps.
  • Because models rarely refuse attribution requests, alignment training has not internalized demographic fairness in reasoning-centered tasks.
  • Bias persists at temperature zero, so it is deterministic and cannot be waved off as sampling noise.
  • The color-assignment result implies the bias transfers to non-social generation tasks, so code that visualizes demographic data may propagate stereotypical associations.

Reading between the lines

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

  • A control that swaps which solution is labeled correct across trials could separate veracity-driven attribution from content-driven pattern matching; the paper does not report such a test.
  • Evaluation bias is largest in essay scoring, suggesting rubric subjectivity may amplify the effect; testing with more objective rubrics could reveal whether ambiguity is necessary.
  • The color-assignment finding suggests stereotype associations extend beyond semantic labels; probing with image generation could test whether the perceptual bias generalizes.
  • The paper's metrics are unsigned, so adding a signed preference vector per group would make the direction of bias explicit and comparable across models.
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. This paper introduces the notion of 'Veracity Bias' in LLMs, defined as the tendency of models to associate solution correctness/incorrectness with demographic groups. The authors design two tasks: Attribution (A), where models assign one correct and one incorrect solution to two of three demographic options, and Evaluation (E), where models grade identical solutions attributed to different demographics. Experiments are reported across five aligned LLMs (GPT-3.5-turbo, GPT-4o, Gemini-1.5-Pro, Claude 3 Sonnet, LLaMA-3-8b) and several benchmarks (GSM8K, MATH, HumanEval, CommonsenseQA, ARC-Easy, ASAP-AES), with additional results on DeepSeek-V3 and Qwen2.5 and on stereotypical color assignment in code. The central claims are that models attribute fewer correct and more incorrect solutions to African-American groups in math and coding, and that evaluation scores vary with perceived authorship beyond random perturbation (Abstract; Section 3.2).

Significance. The topic is timely, and the dual-task design is a reasonable way to probe hidden demographic associations in reasoning, going beyond persona-prompting and direct stereotype questions. The paper covers five aligned models and multiple domains, uses temperature-0 outputs, and includes a random perturbation baseline for evaluation bias. If the identification issues raised in the major comments are addressed, the paper could make a valuable contribution to the bias-and-fairness literature. The color-assignment study is a striking illustration of implicit associations in code generation. The paper does not provide machine-checked proofs or code, but the experimental method is reproducible in principle.

major comments (4)
  1. [Section 2.2, Figure 1, Eqs. (1)-(2)] The forced-choice attribution protocol confounds independent incorrectness bias with the mechanical complement of correct-solution preference. Given three demographic options and two solutions assigned to distinct options, once the model selects a group for the correct solution, the incorrect solution must be assigned to one of the remaining two groups. Thus P(d|incorrect) is not a free quantity; a model that always assigns correct solutions to White/Asian groups will produce elevated P(Black|incorrect) even if it has no belief connecting Black authors to incorrect answers. The paper's headline claim that Black groups receive 'more incorrect' attributions (Abstract; Section 3.2, observations 1-2) is therefore not identified by the reported experiment. Please either (a) reframe the claim to 'fewer correct' attribution only, (b) add an independent-labeling control where each solution is attributed separately (e.g., with an 'other/none' option), or (c) compare the observed incorrect-attribution distribution to the distribution predicted by a null model in which the incorrect solution is assigned uniformly among the options not chosen for the correct solution.
  2. [Section 3.2, Table 2] The paper states that EI values 'consistently exceed' the random perturbation baseline, but no confidence intervals or statistical tests are reported for these comparisons. Several cells are close (e.g., GPT-4o race EI 17% vs random 9%; Claude-3 math 0% vs 0%; LLaMA writing race 13.3% vs 3.3%). Without uncertainty quantification, the 'beyond random perturbation' claim (Abstract) is not supported. Please report confidence intervals (e.g., bootstrap) or exact tests (e.g., McNemar-style per-item comparisons), and specify what the 'neutral placeholders' are in the random perturbation baseline.
  3. [Section 3.1, Data Selection] The 'Solvable' filter—requiring correct solutions across five temperatures up to 0.9—may select an unrepresentative subset of easy, high-confidence problems. The magnitude and direction of attribution bias could differ on more challenging items, so the generalization of the central claim to 'LLMs' behavior is unclear. Please report the number of questions that passed the filter for each model and benchmark and, if feasible, compare results on the full benchmark or on a broader sample.
  4. [Section 3.2, Table 1 and refusal rates] Refusal handling is not described for the main metrics. The text mentions refusal rates for GPT-4o (14% math, 28% coding) and Claude (2% commonsense), but it is not stated whether refusals are excluded from the denominator or treated as a separate outcome. This affects the reported percentages and the chi-square tests. Please clarify the handling and report both raw counts and refusal-excluded analyses.
minor comments (5)
  1. [Section 2.2] The 'Prompt Template for Evaluation (A)' should likely be labeled '(E)' to match the Evaluation Bias task.
  2. [Table 1] The category labels switch between 'Black' and 'African-American' across the text and tables; please standardize the terminology.
  3. [Section 3.3] The color-assignment result is based on 50 repetitions but no quantitative summary or statistical test is given; consider reporting the distribution of color choices per group.
  4. [Section 5 (Limitations)] The Limitations section does not mention the forced-choice complementarity issue discussed in the major comments; please expand it to cover this measurement limitation.
  5. [Section 4 (References)] Some references appear unverifiable (e.g., 'Doe and Chen 2023' and 'Garcia and Patel 2023' in Section 4); please verify that all citations point to real, retrievable publications.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the headline metrics are computed directly from LLM outputs on external benchmark items, and the paper's self-citations appear only as background; the forced-choice attribution design raises an identification caveat but is not a derivation that reduces to its inputs.

full rationale

The paper's central quantities (P(d|correct), P(d|incorrect), EI, EP) are obtained by querying five LLMs on external benchmark problems (GSM8K, MATH, HumanEval, CommonsenseQA, ARC-Easy, ASAP-AES) and are not fitted parameters or predictions derived from the experiment's own construction. No equation in the paper defines ABinc as a transform of ABcor; both are computed from the model's actual cloze assignments. The authors' self-citations (Zhou et al., 2024a, 2024b, 2025) occur in the related-work and mitigation discussion and are not used to justify the experimental design, the metrics, or the main conclusion. The forced-choice attribution protocol (Section 2.2) does create a partial identification caveat: because each trial pairs one correct and one incorrect solution and asks for demographic assignments, the incorrect-attribution margin can be inflated for groups dispreferred on the correct margin; this is a validity threat to the 'more incorrect' sub-claim, but it is not a circular derivation in the sense of Eq. X = Eq. Y by construction, since the model could in principle distribute incorrect assignments non-uniformly among the remaining groups. The evaluation-bias and color-assignment studies are direct behavioral measurements with perturbation baselines, and the Limitations section disclaims causal explanation rather than assuming the conclusion. Therefore the circularity score is minimal.

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

The protocol assumes benchmark correctness is valid, name proxies reflect demographics, temperature-0 outputs are stable, and the forced-choice attribution task reveals veracity-demographic beliefs rather than a format artifact. No numeric free parameters are fitted; all metrics are direct counts from model outputs.

assumptions (5)
  • domain assumption Benchmark ground-truth correctness (GSM8K, MATH, HumanEval, CommonsenseQA, ARC-Easy, ASAP-AES) is valid.
    The whole experiment relies on labeling solutions as correct or incorrect according to these external datasets, stated in Section 3.1.
  • domain assumption Race-associated names are valid proxies for demographic groups.
    Name Proxies are sampled from name lists cited to prior work (Jiang and Fellbaum 2020; Kozlowski et al. 2022), Section 2.2.
  • domain assumption Temperature-0 outputs are deterministic enough for stable measurement.
    The implementation uses temperature = 0 to ensure 'almost deterministic' outputs, Section 3.1, but the paper does not verify stability across runs.
  • ad hoc to paper The forced-choice attribution prompt measures veracity-demographic association rather than a response-format artifact.
    The prompt gives two solutions and three demographic options; the model may exploit priors or avoid duplicate assignments without holding a 'hidden belief'. This assumption is not validated, Section 2.2.
  • domain assumption The ASAP-AES rubric is an appropriate ground truth for essay quality.
    Essay evaluation uses the official 1-6 rubric, Appendix A, which is a human-designed scoring guide, not an objective correctness measure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Veracity Bias and Beyond: Uncovering LLMs' Hidden Beliefs in Problem-Solving Reasoning." pith.science (2026). https://pith.science/paper/5SAE7BOT

@misc{pith2026250516128,
  author       = {Pith},
  title        = {Pith review of: Veracity Bias and Beyond: Uncovering LLMs' Hidden Beliefs in Problem-Solving Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5SAE7BOT}},
  note         = {Machine review of arXiv:2505.16128}
}
read the original abstract

Despite LLMs' explicit alignment against demographic stereotypes, they have been shown to exhibit biases under various social contexts. In this work, we find that LLMs exhibit concerning biases in how they associate solution veracity with demographics. Through experiments across five human value-aligned LLMs on mathematics, coding, commonsense, and writing problems, we reveal two forms of such veracity biases: Attribution Bias, where models disproportionately attribute correct solutions to certain demographic groups, and Evaluation Bias, where models' assessment of identical solutions varies based on perceived demographic authorship. Our results show pervasive biases: LLMs consistently attribute fewer correct solutions and more incorrect ones to African-American groups in math and coding, while Asian authorships are least preferred in writing evaluation. In additional studies, we show LLMs automatically assign racially stereotypical colors to demographic groups in visualization code, suggesting these biases are deeply embedded in models' reasoning processes. Our findings indicate that demographic bias extends beyond surface-level stereotypes and social context provocations, raising concerns about LLMs' deployment in educational and evaluation settings.

Figures

Figures reproduced from arXiv: 2505.16128 by the authors.

Figure 1
Figure 1. An example of Attribution Bias in LLMs, where the model biasedly attributes solutions to demo￾graphic groups based on their (inferred) veracity. researchers have exposed biases by assigning per￾sonas to LLMs to observe decision discrepancies in social scenarios (e.g., “you are religious, your an￾swer should reflect your roles”), or asking the mod￾els to associate specific attributes with social tar￾gets (e.g., assoc… view at source ↗
Figure 2
Figure 2. Attribution patterns across LLMs on GSM8K (math) and HumanEval (coding) benchmarks. The legend [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Maximum Attribution Bias values across racial groups under different reasoning conditions: No Reasoning [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Examples of attribution (top half) and evaluation (bottom half) responses with rationales. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Bar plot color assignments to racial groups in [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 13 canonical work pages

  1. [1]

    AI@Meta. 2024. https://github.com/meta-llama/llama3?tab=readme-ov-file Llama 3 model card

    work page 2024

  2. [2]

    Dai, Anja Hauth, Katie Millican, David Silver, Melvin Johnson, Ioannis Antonoglou, Julian Schrittwieser, Amelia Glaese, et al

    Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M. Dai, Anja Hauth, Katie Millican, David Silver, Melvin Johnson, Ioannis Antonoglou, Julian Schrittwieser, Amelia Glaese, et al. 2024. https://arxiv.org/abs/2312.11805 Gemini: A family of highly capable multimodal models . Preprint, arXiv:2312.11805. O...

    arXiv 2024

  3. [3]

    Anthropic . 2024. https://www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/Model_Card_Claude_3.pdf The claude 3 model family: Opus, sonnet, haiku

    work page 2024

  4. [4]

    Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nyczyk, et al. 2024. Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 17682--17690

    2024

  5. [5]

    Angana Borah and Rada Mihalcea. 2024. https://doi.org/10.18653/v1/2024.findings-emnlp.545 Towards implicit bias detection and mitigation in multi-agent LLM interactions . In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 9306--9326, Miami, Florida, USA. Association for Computational Linguistics

    doi:10.18653/v1/2024.findings-emnlp.545 2024

  6. [6]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian...

    arXiv 2021

  7. [7]

    Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. 2018. https://arxiv.org/abs/1803.05457 Think you have solved question answering? try arc, the ai2 reasoning challenge . Preprint, arXiv:1803.05457. Only available as preprint

    arXiv 2018

  8. [8]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168. Only available as preprint

    arXiv 2021

Show all 39 references
  1. [9]

    Jane Doe and Wei Chen. 2023. https://aclanthology.org/2023.emnlp.123 Debiasing reasoning in language models: A quantitative study . Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing

    2023

  2. [10]

    Emilio Ferrara. 2023. https://doi.org/10.5210/fm.v28i11.13346 Should chatgpt be biased? challenges and risks of bias in large language models . First Monday

    2023 doi

  3. [11]

    Gallegos, Ryan A

    Isabel O. Gallegos, Ryan A. Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K. Ahmed. 2024. https://doi.org/10.1162/coli_a_00524 Bias and fairness in large language models: A survey . Computational Linguistics, 50(3):1097--1179

    2024 doi

  4. [12]

    Maria Garcia and Rohit Patel. 2023. https://arxiv.org/abs/2305.67890 Evaluating fairness in large language models: Metrics and benchmarks . In Proceedings of the 2023 Conference on Neural Information Processing Systems

    2023

  5. [13]

    Yufei Guo, Muzhe Guo, Juntao Su, Zhou Yang, Mengqiu Zhu, Hongfei Li, and Mengyang Qiu. 2024. https://arxiv.org/html/2411.10915v1 Bias in large language models: Origin, evaluation, and mitigation . arXiv preprint arXiv:2411.10915. Only available as preprint

    2024 arXiv

  6. [14]

    Shashank Gupta, Vaishnavi Shrivastava, Ameet Deshpande, Ashwin Kalyan, Peter Clark, Ashish Sabharwal, and Tushar Khot. 2024. https://arxiv.org/abs/2311.04892 Bias runs deep: Implicit reasoning biases in persona-assigned llms . Preprint, arXiv:2311.04892. Only available as preprint

    2024 arXiv

  7. [15]

    Thilo Hagendorff, Sarah Fabi, and Michal Kosinski. 2023. Human-like intuitive behavior and reasoning biases emerged in large language models but disappeared in chatgpt. Nature Computational Science, 3(10):833--838

    2023

  8. [16]

    Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. 2021. Measuring mathematical problem solving with the math dataset. NeurIPS

    2021

  9. [17]

    Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. 2024 a . https://arxiv.org/abs/2310.01798 Large language models cannot self-correct reasoning yet . Preprint, arXiv:2310.01798. Only available as preprint

    2024 arXiv

  10. [18]

    Yue Huang, Lichao Sun, Haoran Wang, Siyuan Wu, Qihui Zhang, Yuan Li, Chujie Gao, Yixin Huang, Wenhan Lyu, Yixuan Zhang, Xiner Li, Hanchi Sun, Zhengliang Liu, Yixin Liu, Yijue Wang, Zhikun Zhang, Bertie Vidgen, Bhavya Kailkhura, Caiming Xiong, Chaowei Xiao, Chunyuan Li, Eric P....

    2024

  11. [19]

    Jiaming Ji, Mickel Liu, Josef Dai, Xuehai Pan, Chi Zhang, Ce Bian, Boyuan Chen, Ruiyang Sun, Yizhou Wang, and Yaodong Yang. 2024. Beavertails: Towards improved safety alignment of llm via a human-preference dataset. Advances in Neural Information Processing Systems, 36

    2024

  12. [20]

    May Jiang and Christiane Fellbaum. 2020. https://aclanthology.org/2020.gebnlp-1.2/ Interdependencies of gender and race in contextualized word embeddings . In Proceedings of the Second Workshop on Gender Bias in Natural Language Processing, pages 17--25, Barcelona, Spain (Onli...

    2020

  13. [21]

    Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, Scott Johnston, Sheer El-Showk, Andy Jones, Nelson Elhage, Tristan Hume, Anna Chen, Yuntao Bai, Sam Bowman, Stanislav For...

    2022 arXiv

  14. [22]

    Diego Kozlowski, Dakota S Murray, Alexis Bell, Will Hulsey, Vincent Larivi \`e re, Thema Monroe-White, and Cassidy R Sugimoto. 2022. Avoiding bias when inferring race using name-based approaches. Plos one, 17(3):e0264270

    2022

  15. [23]

    Andrew K Lampinen, Ishita Dasgupta, Stephanie CY Chan, Hannah R Sheahan, Antonia Creswell, Dharshan Kumaran, James L McClelland, and Felix Hill. 2024. Language models, like humans, show content effects on reasoning tasks. PNAS nexus, 3(7):pgae233

    2024

  16. [24]

    Yuxuan Li, Hirokazu Shirado, and Sauvik Das. 2025. https://arxiv.org/abs/2501.17420 Actions speak louder than words: Agent decisions reveal implicit biases in language models . Preprint, arXiv:2501.17420. Only available as preprint

    2025 arXiv

  17. [25]

    Percy Liang, Rishi Bommasani, Tony Lee, Dimitris Tsipras, Dilara Soylu, Michihiro Yasunaga, et al. 2023. https://arxiv.org/abs/2211.09110 Holistic evaluation of language models . Preprint, arXiv:2211.09110. Only available as preprint

    2023 arXiv

  18. [26]

    Marta Marchiori Manerba, Karolina Stańczak, Riccardo Guidotti, and Isabelle Augenstein. 2023. https://arxiv.org/abs/2311.09090 Social bias probing: Fairness benchmarking for language models . In Proceedings of the Association for Computational Linguistics (ACL) 2023

    2023 arXiv

  19. [27]

    OpenAI. 2023. https://arxiv.org/abs/2303.08774 Gpt-4 technical report . Preprint, arXiv:2303.08774

    2023 arXiv

  20. [28]

    Shuofei Qiao, Yixin Ou, Ningyu Zhang, Xiang Chen, Yunzhi Yao, Shumin Deng, Chuanqi Tan, Fei Huang, and Huajun Chen. 2023. https://doi.org/10.18653/v1/2023.acl-long.294 Reasoning with language model prompting: A survey . In Proceedings of the 61st Annual Meeting of the Associat...

    2023 doi

  21. [29]

    John Smith, Amanda Lee, and Ravi Kumar. 2023. https://arxiv.org/abs/2308.12345 Deepalign: Aligning large language models with social values . In Proceedings of the 2023 Conference on Fairness, Accountability, and Transparency

    2023 arXiv

  22. [30]

    Kaya Stechly, Matthew Marquez, and Subbarao Kambhampati. 2023. https://arxiv.org/abs/2310.12397 Gpt-4 doesn't know it's wrong: An analysis of iterative prompting for reasoning problems . Preprint, arXiv:2310.12397

    2023 arXiv

  23. [31]

    Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. 2019. https://doi.org/10.18653/v1/N19-1421 C ommonsense QA : A question answering challenge targeting commonsense knowledge . In Proceedings of the 2019 Conference of the North A merican Chapter of the Associa...

    2019 doi

  24. [32]

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc V Le, Ed H Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. 2022. Self-consistency improves chain of thought reasoning in language models. In The Eleventh International Conference on Learning Representations

    2022

  25. [33]

    Laura Weidinger, John F. J. Mellor, M. Rauh, C. Griffin, J. Uesato, Po-Sen Huang, M. Cheng, Mia Glaese, B. Balle, A. Kasirzadeh, Z. Kenton, S. Brown, W. Hawkins, T. Stepleton, C. Biles, A. Birhane, Julia Haas, Laura Rimell, Lisa Anne Hendricks, William S. Isaac, Sean Legassick...

    2021 arXiv

  26. [34]

    Yachao Zhao, Bo Wang, Yan Wang, Dongming Zhao, Xiaojia Jin, Jijun Zhang, Ruifang He, and Yuexian Hou. 2024. https://aclanthology.org/2024.lrec-main.17/ A comparative study of explicit and implicit gender biases in large language models via self-evaluation . In Proceedings of t...

    2024

  27. [35]

    Yue Zhou, Barbara Di Eugenio, and Lu Cheng. 2025. https://aclanthology.org/2025.coling-main.485/ Unveiling performance challenges of large language models in low-resource healthcare: A demographic fairness perspective . In Proceedings of the 31st International Conference on Co...

    2025

  28. [36]

    Yue Zhou, Yada Zhu, Diego Antognini, Yoon Kim, and Yang Zhang. 2024 a . https://doi.org/10.18653/v1/2024.naacl-long.153 Paraphrase and solve: Exploring and exploiting the impact of surface form on mathematical reasoning in large language models . In Proceedings of the 2024 Con...

    2024 doi

  29. [37]

    Yue Zhou, Henry Peng Zou, Barbara Di Eugenio, and Yang Zhang. 2024 b . https://doi.org/10.18653/v1/2024.emnlp-main.738 Large language models are involuntary truth-tellers: Exploiting fallacy failure for jailbreak attacks . In Proceedings of the 2024 Conference on Empirical Met...

    2024 doi

  30. [38]

    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...

  31. [39]

    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 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.