Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Evaluating the Promise and Pitfalls of LLMs in Hiring Decisions

T0 review · 3 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 The paper claims that on roughly 10,000 real-world candidate-job pairs, a proprietary domain-specific hiring model outperforms state-of-the-art general-purpose LLMs on both accuracy (ROC AUC 0.85 vs 0.77) and demographic fairness (minimum…

desk verdict Real head-to-head data, but the headline accuracy edge looks like a label-feedback artifact. read the letter →

arxiv 2507.02087 v2 pith:F2XVMOQD submitted 2025-07-02 cs.LG cs.CLcs.CY

Eitan Anzenberg , Arunava Samajpati , Sivasankaran Chandrasekar , Varun Kacholia This is my paper · ORCID
classification cs.LGcs.CLcs.CY
keywords LLMhiringbiasimpactratiofour-fifthsruleresumescreeningalgorithmicfairnesscandidate-jobmatchingROCAUCintersectional
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 asks whether off-the-shelf large language models are safe and accurate enough to score resumes against job descriptions, and compares them with a bespoke supervised hiring model trained on real-world hiring data. The central claim is that the domain-specific model wins on both axes: it reaches a ROC AUC of 0.85 and a minimum race-wise impact ratio of 0.957, while the best LLM reaches 0.77 and 0.809, and every LLM drops below the four-fifths threshold when race and gender are examined together. The authors argue that this refutes the common assumption that fairness must be traded off against accuracy, and they caution that LLMs should not be deployed in hiring without bias audits and safeguards. The evidence comes from roughly 10,000 real candidate-job pairs with self-reported demographics, with all models given identical masked resumes.

What carries the argument

The argument is carried by a standardized evaluation protocol: each resume is parsed and stripped of personally identifiable information, every model receives the same masked resume and job description, scores are converted to binary select/no-select labels by median thresholding, and fairness is measured with the EEOC four-fifths rule using the impact ratio, defined as the smallest group scoring rate divided by the largest. This ratio, computed across gender, race, and race-gender intersectional subgroups, is the load-bearing fairness metric; the accuracy claim rests on ROC AUC, PR AUC, and F1 computed from the same binary labels. The proprietary Match Score model is the central object under comparison, and the paper attributes its combined accuracy-fairness advantage to supervised training on domain data with safeguards rather than to any post-hoc debiasing.

What would settle it

Find a subset of candidate-job pairs whose outcomes were decided before Match Score was deployed or without access to its output, then rerun the evaluation on that subset; if the ROC AUC gap between Match Score and the LLMs shrinks or disappears, the central claim would be weakened.

Watch

Extended reading notes

Core claim

The paper's central discovery is that a purpose-built candidate-job matching model (Match Score) is both more accurate and fairer than every tested general-purpose LLM on this real-world dataset. Match Score's ROC AUC is 0.85 versus 0.77 for the best LLM, its minimum impact ratio across race groups is 0.957 versus 0.809 or lower for all LLMs, and its intersectional impact ratio is 0.906 versus 0.773 or lower. On gender, the LLMs come closer to parity, but this does not protect them on race: one LLM reaches near-perfect gender parity while still falling to 0.774 on race. The paper interprets these results as evidence that accuracy and fairness can be jointly optimized, and that the disparity in LLM outcomes likely stems from pretraining biases rather than from the difficulty of the task.

Load-bearing premise

The ground-truth success labels are assumed to be independent of Match Score and the hiring platform from which the data was sampled—if the model shaped which candidates advanced, its 0.85 accuracy is partly measuring the model against its own output.

Editorial extensions

If this is right

  • A supervised model built for a specific hiring task can satisfy the four-fifths rule across race and intersectional groups without post-hoc debiasing, while every tested general-purpose LLM breaches the 0.80 intersectional threshold.
  • Organizations considering LLM-based screening should run bias audits before deployment; near-parity on one attribute does not imply fairness on others.
  • On this dataset, there is no observed tradeoff between accuracy and fairness: the most accurate model also has the most balanced scoring rates across demographic groups.
  • The paper's protocol—masked resumes, identical inputs, median thresholding, and impact-ratio reporting—can serve as a template for evaluating hiring models.

Reading between the lines

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

  • Beyond the paper: the reported 0.85 ROC AUC may be partly circular if Match Score's production outputs influenced which candidates received interviews, offers, or hires; a label-exogeneity test would settle this.
  • Beyond the paper: a replication using decisions made before Match Score existed, or labels generated by an independent panel, would show whether the accuracy gap is intrinsic to the models or an artifact of the dataset.
  • Beyond the paper: because only self-reported gender and EEOC race categories are included, the fairness conclusions may not extend to age, disability, veteran status, or non-binary gender, which are absent from the data.
  • Beyond the paper: the paper's per-1,000-inference cost figures imply that a deployment decision should weigh cost per accurate and fair decision, not just accuracy and fairness in separate columns.
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

3 major / 5 minor

Summary. The paper compares a proprietary supervised candidate-matching model (Match Score) with several closed- and open-weight LLMs on a dataset of roughly 10,000 real-world candidate–job pairs drawn from the authors' hiring platform. It reports that Match Score achieves higher accuracy (ROC AUC 0.85 vs. 0.77 for the best LLM) and better fairness outcomes (minimum race-wise impact ratio 0.957 vs. 0.809, and intersectional impact ratio 0.906 vs. 0.773), and uses these results to argue that a domain-specific supervised model can jointly optimize accuracy and demographic fairness, whereas off-the-shelf LLMs cannot. The paper also discusses possible sources of LLM bias and proposes a bias-audit methodology based on the EEOC four-fifths rule.

Significance. If the comparison were valid, the paper would make a practically important contribution: it provides a large real-world head-to-head benchmark of proprietary versus general-purpose models on a high-stakes task, with self-reported demographic attributes and both accuracy and fairness metrics. The paper also gives a clear prompting protocol and a transparent set of evaluation metrics, which are useful for replication. However, the central claim is undermined by a likely label-feedback problem: the outcome labels come from the same platform in which Match Score is deployed, and the paper does not establish that these labels were generated independently of Match Score's outputs. Given the large reported performance gaps (0.85 vs. 0.77 ROC AUC; race IR 0.957 vs. 0.809), this is not a minor caveat but a threat to the validity of the headline comparison. The paper also has unresolved issues with the per-model median thresholding and incomplete uncertainty reporting.

major comments (3)
  1. [Section 3.1; Section 3.2; Table 1] The central head-to-head comparison is not identified: the ground-truth labels are real-world outcomes (on-site interview, offer, hire) from the authors' platform, and Match Score is a deployed proprietary model on that platform. The paper never states that these outcomes were generated without exposure to Match Score scores. If recruiters or automated workflows used Match Score to shortlist candidates, then the label is partly a function of the very model being evaluated, and Match Score is rewarded for predicting outcomes it helped produce. The LLMs, evaluated zero-shot, could not have influenced the labels, so the comparison is asymmetric. Section 3.2's statement that Match Score was not trained on test data addresses contamination, not endogeneity; Section 6 does not mention this threat. The reported gaps (ROC AUC 0.85 vs. 0.77; race IR 0.957 vs. 0.809) are large enough that even partial feedback could account for them. Please provide evidence that labels were determined independently of Match Score (e.g., decisions made before model deployment or by humans without access to the score), or re-run the evaluation on labels that are exogenous to Match Score, and report a sensitivity analysis.
  2. [Section 3.3; Table 2] The binary outcome for every model is obtained by thresholding at that model's own median score. This sets each model's overall selection rate to approximately 50% (or to a rate determined by the tie structure), which makes the impact-ratio comparison across models ill-posed: a model with a narrow score distribution will have very different absolute selection rates from one with a wide distribution. The Table 2 scoring rates (e.g., 64.2% for Match Score females, 68.4% for GPT-4o females) are inconsistent with a literal median split and are not comparable across models because the thresholds differ. The fairness analysis should use a common, pre-specified operating point (or report the full score distribution), and the text should explain how ties at the median are handled.
  3. [Section 4.2; Table 2] The text states that 'the variance of scoring rates is large for the LLMs, therefore, disparate impact cannot be attributed to noise,' but no confidence intervals are reported for the group-specific scoring rates or impact ratios in Table 2, and Table 1 shows CIs only in the prose, not in the table. Without these intervals, and given the small sizes of some intersectional subgroups, the claim that the observed disparities are systematic is not supported. Please report CIs for all reported IRs and group SRs, and specify subgroup sample sizes.
minor comments (5)
  1. [Section 3.3] Please clarify whether 'above the median' means '≥ median' and whether the median is computed per model on the test set; the current wording is ambiguous and inconsistent with the scoring rates shown in Table 2.
  2. [Section 4.1; Table 1] The 95% confidence intervals mentioned in the text are not shown in Table 1; add them to the table or a supplement so that the reader can verify the stated differences.
  3. [Section 3.2] Match Score is described only as a proprietary in-house model; a high-level description of its inputs, training data, and the nature of its fairness safeguards would be needed to interpret the comparison and to assess the claim that it is 'well-designed' for fairness.
  4. [References] The citation for NYC Local Law 144 appears as 'of New York, 2023' in the text; this should be a proper citation to the City of New York, and the reference list entry should be formatted consistently.
  5. [Section 3.2] DeepSeek R1 is described as a retrieval-augmented transformer, but DeepSeek-R1 is a reasoning model; the cited GitHub repository appears to refer to a different project, so the description or reference should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the scorecard is computed directly from model outputs and fixed outcome labels, and the possible label-feedback concern is an external-validity limitation rather than a demonstrated circular derivation.

full rationale

The paper's claimed derivation is an empirical benchmark: each model produces a relevance score, the score is thresholded at its median, and the resulting binary predictions are compared against fixed ground-truth labels (on-site interview, offer, or hired) using standard accuracy and fairness metrics. These computations are transparent and not circular by construction. The main concern raised by the reader—that Match Score may have influenced the hiring outcomes used as labels—is a real threat to the causal interpretation of the accuracy gap, but the paper nowhere states or formalizes that labels are a function of Match Score outputs. Without an explicit statement or equation showing that the ground-truth label depends on the model being evaluated, this remains an external-validity and data-generation concern, not a circularity that can be exhibited from the text. The paper does contain self-referential elements (the proprietary model, the internal audit dataset), but none of them is used to substitute for a derivation step or to define the evaluation target in terms of the model itself. Therefore, under the rule that circularity must be demonstrated by a specific reduction and not inferred from plausibility, the correct finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper's central comparison relies on the independence of the outcome labels from Match Score, and on fairness metrics computed from self-reported demographics; these are unverified assumptions. The free parameters are thresholding and prompt choices that can materially change the reported scores.

free parameters (3)
  • median threshold per model = computed per model on the test set
    Scores are binarized at each model's median on the evaluation data (Section 3.3), which affects F1, scoring rates, and impact ratios; this operating point is chosen on the same data used for evaluation.
  • prompt template
    The standardized prompt (Figure 2) is hand-crafted; different prompt phrasings are known to change LLM outputs, and no prompt sensitivity analysis is provided.
  • intersectional group inclusion rule
    Intersectional groups are included 'where sample sizes permit reliable statistics' (Section 3.1); the exact sample-size cutoff is not stated, so the set of reported groups is a choice affecting the reported minimum impact ratios.
assumptions (4)
  • domain assumption Ground-truth binary outcome (on-site interview, offer, or hired) is a valid measure of candidate success and is independent of the models being evaluated.
    Section 3.1 defines the label; Section 6 discusses candidate self-selection but never addresses whether Match Score or the Eightfold platform influenced these outcomes, a necessary condition for an unbiased accuracy comparison.
  • domain assumption Masked resumes contain no protected attributes and LLMs do not infer them.
    Section 3.1 claims masking redacts PII; but the paper itself notes LLMs may pick up 'subtle language cues correlated with demographics' (Section 5), so the assumption that the LLM sees no demographic signal is not established.
  • domain assumption Self-reported gender and race/ethnicity categories are accurate and complete.
    Section 6 notes only self-reported data are available; measurement error in these attributes would affect impact-ratio estimates.
  • domain assumption The dataset is representative of real-world candidate-job pairs.
    Section 3.1 states a sample of ~10,000 pairs from the authors' platform covering 2023-2024; no sampling procedure or external validation is described.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evaluating the Promise and Pitfalls of LLMs in Hiring Decisions." pith.science (2026). https://pith.science/paper/F2XVMOQD

@misc{pith2026250702087,
  author       = {Pith},
  title        = {Pith review of: Evaluating the Promise and Pitfalls of LLMs in Hiring Decisions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F2XVMOQD}},
  note         = {Machine review of arXiv:2507.02087}
}
read the original abstract

The use of large language models (LLMs) in hiring promises to streamline candidate screening, but it also raises serious concerns regarding accuracy and algorithmic bias where sufficient safeguards are not in place. In this work, we benchmark several state-of-the-art foundational LLMs - including models from OpenAI, Anthropic, Google, Meta, and Deepseek, and compare them with our proprietary domain-specific hiring model (Match Score) for job candidate matching. We evaluate each model's predictive accuracy (ROC AUC, Precision-Recall AUC, F1-score) and fairness (impact ratio of cut-off analysis across declared gender, race, and intersectional subgroups). Our experiments on a dataset of roughly 10,000 real-world recent candidate-job pairs show that Match Score outperforms the general-purpose LLMs on accuracy (ROC AUC 0.85 vs 0.77) and achieves significantly more equitable outcomes across demographic groups. Notably, Match Score attains a minimum race-wise impact ratio of 0.957 (near-parity), versus 0.809 or lower for the best LLMs, (0.906 vs 0.773 for the intersectionals, respectively). We discuss why pretraining biases may cause LLMs with insufficient safeguards to propagate societal biases in hiring scenarios, whereas a bespoke supervised model can more effectively mitigate these biases. Our findings highlight the importance of domain-specific modeling and bias auditing when deploying AI in high-stakes domains such as hiring, and caution against relying on off-the-shelf LLMs for such tasks without extensive fairness safeguards. Furthermore, we show with empirical evidence that there shouldn't be a dichotomy between choosing accuracy and fairness in hiring: a well-designed algorithm can achieve both accuracy in hiring and fairness in outcomes.

Figures

Figures reproduced from arXiv: 2507.02087 by the authors.

Figure 1
Figure 1. Illustration of preprocessing: Top: Resume parsing masks the original resume (left) of personal information (right) and standardizes the resume format to be used for downstream models. Bottom: Raw text output from our resume parser for the same resume excerpt, including the sanitized list of extracted skills. Academic works have explored bias mitigation in hiring algorithms, from debiasing word embed￾dings in job ad… view at source ↗
Figure 2
Figure 2. Sample prompt we feed into our evaluator to score resume–job relevance. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Fairness Is Not Enough: Auditing Competence and Intersectional Bias in AI-powered Resume Screening

    cs.CY 2025-07 conditional novelty 6.0 of 10

    Models that appear demographically neutral in AI resume screening can actually be incompetent evaluators, a pattern the paper calls the Illusion of Neutrality.

Reference graph

Works this paper leans on

31 extracted references · 24 canonical work pages · cited by 1 Pith paper

  1. [1]

    Persistent anti-muslim bias in large language models

    Abubakar Abid, Maheen Farooqi, and James Zou. Persistent anti-muslim bias in large language models. Proceedings of the ACM Conference on Fairness, Accountability, and Transparency (FAccT), 2021

    work page 2021

  2. [2]

    Categorical Data Analysis

    Alan Agresti. Categorical Data Analysis. Wiley, 3 edition, 2013

    work page 2013

  3. [3]

    Claude 3.5 v2 : A research model for safe and creative reasoning

    Anthropic . Claude 3.5 v2 : A research model for safe and creative reasoning. https://www.anthropic.com/models/claude-3-5-v2, February 2025. Accessed May 2025

    work page 2025

  4. [4]

    On the dangers of stochastic parrots: Can language models be too big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, pages 610--623, 2021

    Emily Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. On the dangers of stochastic parrots: Can language models be too big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, pages 610--623, 2021

    work page 2021

  5. [5]

    Are Emily and Greg more employable than Lakisha and Jamal ? A field experiment on labor market discrimination

    Marianne Bertrand and Sendhil Mullainathan. Are Emily and Greg more employable than Lakisha and Jamal ? A field experiment on labor market discrimination. American Economic Review, 94 0 (4): 0 991--1013, 2004

    work page 2004

  6. [6]

    Putting fairness principles into practice: Challenges, metrics, and improvements

    Alex Beutel, Jilin Chen, Zhe Zhao, et al. Putting fairness principles into practice: Challenges, metrics, and improvements. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, 2019

    work page 2019

  7. [7]

    Man is to computer programmer as woman is to homemaker? debiasing word embeddings

    Tolga Bolukbasi, Kai-Wei Chang, James Zou, Venkatesh Saligrama, and Adam Kalai. Man is to computer programmer as woman is to homemaker? debiasing word embeddings. In Advances in Neural Information Processing Systems, pages 4349--4357, 2016

    work page 2016

  8. [8]

    Bias audit for New York City local law 144: Summary of bias audit results

    Shea Brown. Bias audit for New York City local law 144: Summary of bias audit results. BABL AI Inc. report, March 2025. URL https://eightfold.ai/wp-content/uploads/eightfold-summary-of-bias-audit-results.pdf

    work page 2025

Show all 31 references
  1. [9]

    Brown et al

    Tom B. Brown et al. Language models are few-shot learners. In Advances in Neural Information Processing Systems 33 (NeurIPS), 2020

    2020

  2. [10]

    Evaluating large language models trained on code

    Mark Chen et al. Evaluating large language models trained on code. arXiv preprint arXiv:2107.03374, 2021

    2021 arXiv

  3. [11]

    Amazon scraps secret AI recruiting tool that showed bias against women

    Jeffrey Dastin. Amazon scraps secret AI recruiting tool that showed bias against women. Reuters, October 2018

    2018

  4. [12]

    Gemini 1.5: Scaling up token capacity for large language models, 2024

    Google DeepMind. Gemini 1.5: Scaling up token capacity for large language models, 2024. https://deepmind.google/technologies/gemini/

    2024

  5. [13]

    Deepseek R1 : Retrieval‑augmented open‑weight language model

    Deepseek AI . Deepseek R1 : Retrieval‑augmented open‑weight language model. https://github.com/deepseek-ai/Deepseek-Retrieval-LLM, December 2024. Online; accessed 2 May 2025

    2024

  6. [14]

    Gaebler, Sharad Goel, Aziz Huq, and Prasanna Tambe

    Johann D. Gaebler, Sharad Goel, Aziz Huq, and Prasanna Tambe. Auditing the use of language models to guide hiring decisions. arXiv preprint arXiv:2404.03086, 2024. URL https://arxiv.org/abs/2404.03086

    2024 arXiv

  7. [15]

    Hanley and Barbara J

    James A. Hanley and Barbara J. McNeil. The meaning and use of the area under a receiver operating characteristic (roc) curve. Radiology, 143 0 (1): 0 29--36, 1982

    1982

  8. [16]

    99\ Online: https://www.jobscan.co/blog/99-percent-fortune-500-ats/, November 2019

    James Hu. 99\ Online: https://www.jobscan.co/blog/99-percent-fortune-500-ats/, November 2019

    2019

  9. [17]

    Documenting high-risk AI : A European regulatory perspective

    Isabelle Hupont, Mar \'i a M \'i chel, Biagio Di Stefano , Emilia G \'o mez , and Josep Soler Garrido . Documenting high-risk AI : A European regulatory perspective. Computer, 56 0 (5): 0 18--27, 2023

    2023

  10. [18]

    Obtaining confidence intervals for the risk ratio in cohort studies

    David Katz, Jennifer Baptista, Stanley Azen, and Melva Pike. Obtaining confidence intervals for the risk ratio in cohort studies. Biometrics, 34 0 (3): 0 469--474, 1978

    1978

  11. [19]

    Holistic evaluation of language models

    Percy Liang, Rishi Bommasani, et al. Holistic evaluation of language models. arXiv preprint arXiv:2211.09110, 2022

    2022 arXiv

  12. [20]

    A hiring law blazes a path for AI regulation

    Steve Lohr. A hiring law blazes a path for AI regulation. The New York Times, May 2023

    2023

  13. [21]

    The Llama 4 herd: The beginning of a new era of natively multimodal ai innovation

    Meta AI . The Llama 4 herd: The beginning of a new era of natively multimodal ai innovation. https://ai.meta.com/blog/llama-4-multimodal-intelligence/, 2025. Online; accessed 2 May 2025

    2025

  14. [22]

    Llama 3 : Open foundation models

    Meta AI . Llama 3 : Open foundation models. https://ai.meta.com/blog/meta-llama-3, April 2024. Online; accessed 2 May 2025

    2024

  15. [23]

    NYC local law 144: Automated employment decision tool bias audit law, 2023

    City of New York. NYC local law 144: Automated employment decision tool bias audit law, 2023. URL https://www.nyc.gov/assets/dca/downloads/pdf/about/LL144-AEDT-Rules.pdf. Passed April 2023

    2023

  16. [24]

    GPT -4 technical report, 2023

    OpenAI. GPT -4 technical report, 2023. https://openai.com/research/gpt-4

    2023

  17. [25]

    Mitigating bias in algorithmic hiring: Evaluating claims and practices

    Manish Raghavan, Solon Barocas, Jon Kleinberg, and Karen Levy. Mitigating bias in algorithmic hiring: Evaluating claims and practices. In Proceedings of the ACM Conference on Fairness, Accountability, and Transparency, pages 469--481, 2020

    2020

  18. [26]

    Investigating hiring bias in large language models

    Akshaj Kumar Veldanda, Fabian Grob, Shailja Thakur, Hammond Pearce, Benjamin Tan, Ramesh Karri, and Siddharth Garg. Investigating hiring bias in large language models. In R0-FoMo Poster, OpenReview, November 2023. URL https://openreview.net/forum?id=erl90pLIH0

    2023

  19. [27]

    Jobfair: A framework for benchmarking gender hiring bias in large language models, 2024

    Ze Wang, Zekun Wu, Xin Guan, Michael Thaler, Adriano Koshiyama, Qinyang Lu, Sachin Beepath, Ediz Ertekin Jr., and Maria Perez-Ortiz. Jobfair: A framework for benchmarking gender hiring bias in large language models, 2024. URL https://arxiv.org/abs/2406.15484v1

    2024 arXiv

  20. [28]

    Gender, race, and intersectional bias in resume screening via language model retrieval

    Kyra Wilson and Aylin Caliskan. Gender, race, and intersectional bias in resume screening via language model retrieval. arXiv preprint arXiv:2407.20371v2, July 2024

    2024 arXiv

  21. [29]

    Defending against neural fake news

    Rowan Zellers, Ari Holtzman, Hannah Rashkin, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Defending against neural fake news. In Advances in Neural Information Processing Systems 32 (NeurIPS), pages 9054--9065, 2019

    2019

  22. [30]

    Men also like shopping: Reducing gender bias amplification using corpus-level constraints

    Jieyu Zhao, Tianlu Wang, Mark Yatskar, Vicente Ordonez, and Kai-Wei Chang. Men also like shopping: Reducing gender bias amplification using corpus-level constraints. In EMNLP, 2017

    2017

  23. [31]

    Fragile prompting: LLM s can fail to follow simple instructions

    Shuyuan Zhou, Adrian Weller, Inioluwa Raji, et al. Fragile prompting: LLM s can fail to follow simple instructions. arXiv preprint arXiv:2306.05685, 2023

    2023 arXiv

Pith tools

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