Pith. sign in

REVIEW 5 major objections 6 minor 61 references

This paper claims that a prompt-level guardrail can reduce harmful LLM outputs to vulnerable populations by 25–30%, with formal convergence and safety bounds.

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

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

T0 review

2026-08-04 19:59 UTC pith:IURNPPG6

load-bearing objection A useful architecture blueprint, but the 'formal proofs' are asserted, not shown—don't take the 25-30% figure at face value. the 5 major comments →

arxiv 2509.08910 v1 pith:IURNPPG6 submitted 2025-09-10 cs.CV cs.AI

PromptGuard: An Orchestrated Prompting Framework for Principled Synthetic Text Generation for Vulnerable Populations using LLMs with Enhanced Safety, Fairness, and Controllability

Tung Vu , Lam Nguyen , Quynh Dao This is my paper
classification cs.CV cs.AI
keywords large language modelssynthetic text generationvulnerable populationsprompt guardrailscontrastive promptingharm preventionmulti-objective optimizationinformation-theoretic bounds
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

PromptGuard's central claim is that harmful text generation for vulnerable groups can be prevented at the prompt level rather than filtered after the fact. Its VulnGuard component builds a population-specific protective barrier from community-sourced contrastive examples, ethical chain-of-thought reasoning, and adaptive role instructions, and frames the whole generation step as a constrained multi-objective optimization. The paper argues that this formulation yields formal guarantees: convergence to a harm-bounded state, an information-theoretic safety bound, Pareto optimality across harm, utility, efficiency, and cost, and an analytical harm reduction of 25–30% over baseline prompting. If right, it would mean a purely prompt-based, no-retraining route to safer synthetic data for groups such as LGBTQ+ individuals and single parents.

Core claim

VulnGuard Prompt is the paper's central object: a hybrid prompt template that inserts harmful-versus-safe example pairs drawn from curated community datasets directly into the context, then directs the model to reason ethically before producing an answer. The framework defines generation as the minimization of a composite loss L_total = αL_harm + βL_utility + γL_coherence, with harm loss computed against the safety patterns in the dataset. Three theorems are offered: convergence to an ε-optimal harm-reduced state in O(log(1/ε)) iterations when the dataset has enough examples per harm category; a safety bound P(harm) ≤ exp(−α√|D|) derived from information-theoretic quantities; and Pareto opti

What carries the argument

VulnGuard Prompt is the load-bearing mechanism: a prompt template that prepends harmful/safe contrastive examples and an ethical reasoning checklist to the user query. It acts within a six-module architecture—input classification, VulnGuard prompting, ethical principles integration, external tool interaction, output validation, and user-system interaction—so that the guardrail is not a single instruction but an orchestrated pipeline. The mathematical engine is the composite loss L_total = αL_harm + βL_utility + γL_coherence, whose strong convexity (when the dataset is sufficiently covered) is said to give exponential convergence, with a divergence-based stability function and information-the

Load-bearing premise

The convergence and safety proofs assume that the harm-score function has one smooth, bowl-shaped minimum over all possible outputs and that every harm category has enough curated examples; the paper does not specify the output space or verify these conditions.

What would settle it

Run a fixed set of prompts targeting LGBTQ+ and single-parent scenarios through a baseline LLM with and without VulnGuard, using the paper's own loss definition and curated example set, and count harmful outputs with a pre-registered rubric. If the measured harm reduction over several random seeds does not approach 25–30%, or if empirical P(harm) exceeds exp(−α√|D|), the analytical claim is not supported.

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

If this is right

  • Harm probability under VulnGuard is bounded by exp(−α√|D|), where D is the size of the curated ethical example set; adding more curated examples tightens the bound.
  • VulnGuard converges to an ε-optimal harm-reduced state in O(log(1/ε)) iterations, making the guardrail feasible in interactive settings.
  • The framework can be layered onto existing LLMs without retraining or fine-tuning, since all protection is expressed through prompts and external checks.
  • The paper's expected analyses project PII leakage reductions of 70–85%, disparate-impact bias reductions of 40–60%, and jailbreak robustness improvements of 80–95% relative to unguarded baselines.
  • Pareto optimality extends the usual effectiveness-time-cost optimization to include harm prevention, giving a principled trade-off surface for ethical deployment.

Where Pith is reading between the lines

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

  • A natural next step is to turn the stated bounds into a benchmark: fix a harm taxonomy, instantiate L_harm, and measure whether observed harm rates obey exp(−α√|D|) as the curated example set grows; the paper does not report that experiment.
  • Because the convergence proof explicitly assumes at least k examples per harm category, a testable extension is to quantify how performance degrades when a category is underrepresented in the curated data.
  • If the prompt-only mechanism holds, the same orchestration could migrate to multimodal generation by replacing textual contrastive examples with multimodal safe/harmful pairs, a direction the authors list as future work.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes PromptGuard, a modular prompting framework whose core innovation is VulnGuard, a data-driven prompting technique intended to prevent harmful LLM outputs for vulnerable populations (LGBTQ+ individuals, single parents, etc.). The manuscript claims formal theoretical support: convergence proofs, information-theoretic safety bounds, Pareto optimality in a four-dimensional objective space, and an analytical 25-30% harm reduction relative to baseline prompting. It also describes a six-module system architecture, an LLM selection protocol, an evaluation methodology, and an extensive ethical-governance discussion. The reported quantitative results are framed as 'Expected analysis' rather than measured outcomes, and the theoretical section contains only proof sketches.

Significance. The problem addressed—proactive, population-specific harm prevention for LLM-generated text—is important, and the proposed modular architecture is a plausible organizing framework for future empirical work. The paper also includes thoughtful discussion of participatory design, dual-use risks, and community engagement. However, the central claimed contributions are not established. No machine-checked proofs, reproducible code, completed experiments, or derived quantitative bounds are provided. The 25-30% analytical harm-reduction figure, the exponential safety bound, and the Pareto optimality claim are asserted rather than proved. If the results were supported, the work would be significant; in its current form, the theoretical and empirical load-bearing elements are absent.

major comments (5)
  1. [Abstract, §I-B, §V-C (Theorem 2), §IX-A] The claimed 'formal proofs demonstrating 25-30% analytical harm reduction' are never derived. Theorem 2 states P(Harm) ≤ exp(-α sqrt(|D_GitHub|)) without proof, without defining α, and without specifying a baseline policy P_baseline. An absolute upper bound cannot yield a percentage reduction unless a baseline probability is defined and compared. The abstract, introduction, and conclusion repeat the 25-30% claim, but no equation in Section V connects the displayed inequalities to that number. This is a load-bearing gap in the paper's central contribution.
  2. [§V-B (Theorem 1), Eq. (2)-(3)] Theorem 1's proof sketch invokes 'strong convexity of the empirical risk minimization problem' and 'sufficient coverage' / 'at least k examples per harm category' without defining the output space Y, the loss L_harm, or the harm categories. The loss in Eq. (2)-(3) is not shown to be convex or strongly convex over Y, and no concrete convergence argument, Lyapunov function, or margin bound is supplied. The claimed O(log(1/ε)) convergence therefore does not follow from the text. The proof sketch restates the conclusion rather than providing a derivation.
  3. [§VI, §VII-A] Section VI describes an implemented framework, datasets (VulnPop-Synth, News-Derived Scenarios, Adversarial Test Suite), and a long list of metrics, but reports no measured outcomes. Section VII-A labels all quantitative results as 'Expected analysis' and gives ranges (70-85%, 40-60%, 50-70%, 80-95%, 5-15%) with no tables, confidence intervals, significance tests, or comparison to the baselines defined in Section VI-E. The conclusion's statement that PromptGuard 'demonstrates' these effects is unsupported by any empirical data in the manuscript.
  4. [§V-D (Theorem 3)] Pareto optimality is asserted without proof or a precise multi-objective formulation. The weighted loss in Eq. (2) uses constants α, β, γ, but no connection is made between scalarization and the Pareto frontier, and no trade-off conditions, existence argument, or characterization of the frontier is provided. Since Theorem 3 underpins the 'four-dimensional optimization' contribution, the absence of a derivation is a substantive gap.
  5. [§IX-D; §V-D closing sentence] The self-admitted limitations are themselves damaging to the claims. Section IX-D states the validation pipeline has 2.3x overhead, evaluation is Western-centric, and dynamic adaptation remains underdeveloped. Simultaneously, Section V-D claims the theoretical framework establishes 'rigorous mathematical foundations' and 'without requiring extensive empirical validation.' These statements are in tension: the paper presents only proof sketches and projected analyses, yet rejects the need for empirical confirmation. The manuscript's own text therefore flags that the central claims are not yet supported.
minor comments (6)
  1. [Eq. (3)] The notation P(safe | y, D_GitHub, p) is introduced but never defined concretely; it is unclear how this probability is computed from a prompt template or dataset. Also, the notation 'D GitHub' is rendered inconsistently as D_GitHub, DGitHub, and D GitHub.
  2. [§VII-A] Sections titled 'Quantitative Results' contain only expected/projected ranges, not results. If the authors intend to present a proposal, this should be relabeled and separated from any claim of validation.
  3. [§II-B] The claim of a 'systematic review of 247 papers (2020-2024)' is unsupported by a search strategy, inclusion criteria, or a reference list; it should be substantiated or removed.
  4. [§VI-A, Listing 1] Listing 1 is pseudocode with placeholders ('[Loaded from dataset]') and no details about actual GitHub-sourced examples, curation, or harm severity labeling. The reader cannot assess the data pipeline's validity from the presented material.
  5. [References] Several references lack complete bibliographic information (e.g., [22] and [29] missing venue or publisher details; some arXiv entries are not dated). This should be corrected before any resubmission.
  6. [Figures 1-2] The conceptual figures are high-level and do not include concrete prompt templates or module state transitions; they add little beyond the textual description. Including an actual worked example would improve clarity.

Circularity Check

0 steps flagged

No significant circularity: the 25–30% figure is asserted rather than derived, but no claim reduces to its inputs by construction or via self-citation.

full rationale

I examined the paper's derivation chain for the seven circularity patterns. VulnGuard is defined as the minimizer of a weighted loss (Eqs. 1-3); Theorem 1 asserts convergence of that minimization under strong convexity and sufficient coverage; Theorem 2 states an exponential bound P(Harm) ≤ exp(-α√|D_GitHub|) without proof or baseline; Theorem 3 asserts Pareto optimality without proof. None of these steps takes a fitted parameter and renames it a prediction, and the paper contains no self-citations or imported uniqueness theorems. The abstract's '25-30% analytical harm reduction' is never derived from the displayed inequalities; that is an unsupported quantitative claim, a correctness/rigor problem rather than a circular-reasoning problem, because the number does not equal any input by construction. Since the specific reduction required by the circularity definition cannot be exhibited, the circularity score is 0.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The framework rests on unproven convexity, unspecified coverage of external datasets, and an unknown constant alpha. No new physical or conceptual entities are introduced; the entire burden is on the missing derivations.

free parameters (3)
  • alpha (population-specific constant) = unspecified
    Theorem 2: P(Harm) <= exp(-alpha sqrt(|D|)), alpha is never estimated or bounded; the 25-30% reduction claim depends on alpha.
  • loss weights alpha, beta, gamma = not specified
    Equation (2): L_total = alpha*L_harm + beta*L_utility + gamma*L_coherence; these weights are free design choices that determine the optimum.
  • thresholds tau_utility, tau_coherence = not specified
    Equations (5)-(6) impose thresholds that shape the feasible set; no values are given.
axioms (4)
  • domain assumption L_harm is strongly convex
    Theorem 1's proof sketch relies on strong convexity of the empirical risk; not demonstrated for an LLM output space.
  • domain assumption GitHub dataset D has sufficient coverage and is balanced across harm categories
    Theorem 1 requires at least k examples per harm category; coverage and representativeness are asserted, not measured.
  • domain assumption LLMs reliably follow multi-instruction ethical prompts
    The entire framework assumes instruction-following accuracy high enough to enforce barriers; only future evaluation is proposed.
  • domain assumption Entropy and mutual information bounds translate to actual harm probabilities
    Equations (7)-(8) identify harm with information-theoretic quantities without calibration or empirical grounding.

pith-pipeline@v1.3.0-alltime-deepseek · 18726 in / 8163 out tokens · 78707 ms · 2026-08-04T19:59:00.794300+00:00 · methodology

0 comments
read the original abstract

The proliferation of Large Language Models (LLMs) in real-world applications poses unprecedented risks of generating harmful, biased, or misleading information to vulnerable populations including LGBTQ+ individuals, single parents, and marginalized communities. While existing safety approaches rely on post-hoc filtering or generic alignment techniques, they fail to proactively prevent harmful outputs at the generation source. This paper introduces PromptGuard, a novel modular prompting framework with our breakthrough contribution: VulnGuard Prompt, a hybrid technique that prevents harmful information generation using real-world data-driven contrastive learning. VulnGuard integrates few-shot examples from curated GitHub repositories, ethical chain-of-thought reasoning, and adaptive role-prompting to create population-specific protective barriers. Our framework employs theoretical multi-objective optimization with formal proofs demonstrating 25-30% analytical harm reduction through entropy bounds and Pareto optimality. PromptGuard orchestrates six core modules: Input Classification, VulnGuard Prompting, Ethical Principles Integration, External Tool Interaction, Output Validation, and User-System Interaction, creating an intelligent expert system for real-time harm prevention. We provide comprehensive mathematical formalization including convergence proofs, vulnerability analysis using information theory, and theoretical validation framework using GitHub-sourced datasets, establishing mathematical foundations for systematic empirical research.

Figures

Figures reproduced from arXiv: 2509.08910 by Lam Nguyen, Quynh Dao, Tung Vu.

Figure 1
Figure 1. Figure 1: Conceptual Architecture of the PromptGuard Framework, illustrating the distinct modules, their primary interactions, and the controlled flow of data [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of the PromptGuard execution plan for generating a [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

61 extracted references · 21 linked inside Pith

  1. [1]

    Language models are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askellet al., “Language models are few-shot learners,”Advances in Neural Information Processing Systems, vol. 33, pp. 1877–1901, 2020

    1901

  2. [2]

    Privacy- preserving synthetic data generation for healthcare applications,

    M. Hernandez, G. Epelde, A. Alberdi, R. Cilla, and D. Rankin, “Privacy- preserving synthetic data generation for healthcare applications,” pp. 532–555, 2022

    2022

  3. [3]

    Synthea: An approach, method, and software mechanism for generating synthetic patients and the synthetic electronic health care record,

    J. Walonoski, M. Kramer, J. Nichols, A. Quina, C. Moesel, D. Hall, C. Duffett, K. Dube, T. Gallagher, and S. McLachlan, “Synthea: An approach, method, and software mechanism for generating synthetic patients and the synthetic electronic health care record,” pp. 230–238, 2018

    2018

  4. [4]

    On the dangers of stochastic parrots: Can language models be too big?

    E. M. Bender, T. Gebru, A. McMillan-Major, and S. Shmitchell, “On the dangers of stochastic parrots: Can language models be too big?” in Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 2021, pp. 610–623

    2021

  5. [5]

    Algorithmic injustice: A relational ethics approach,

    A. Birhane, “Algorithmic injustice: A relational ethics approach,” p. 100205, 2021

    2021

  6. [6]

    Ai4peoplean ethical framework for a good ai society: Opportunities, risks, principles, and recommendations,

    L. Floridi, J. Cowls, M. Beltrametti, R. Chatila, P. Chazerand, V . Dignum, C. Luetge, R. Madelin, U. Pagallo, F. Rossiet al., “Ai4peoplean ethical framework for a good ai society: Opportunities, risks, principles, and recommendations,”Minds and Machines, vol. 28, no. 4, pp. 689–707, 2018

    2018

  7. [7]

    Livingstone, G

    S. Livingstone, G. Mascheroni, and E. Staksrud,Children’s Online Risks and Safety: A Review of the Available Evidence. LSE Research Online, 2019

    2019

  8. [8]

    Eubanks,Automating Inequality: How High-Tech Tools Profile, Po- lice, and Punish the Poor

    V . Eubanks,Automating Inequality: How High-Tech Tools Profile, Po- lice, and Punish the Poor. St. Martin’s Press, 2018

    2018

  9. [9]

    The misgendering machines: Trans/hci implications of automatic gender recognition,

    O. Keyes, “The misgendering machines: Trans/hci implications of automatic gender recognition,”Proceedings of the ACM on Human- Computer Interaction, vol. 2, no. CSCW, pp. 1–22, 2018

    2018

  10. [10]

    Technocolonialism: Digital innovation and data practices in the humanitarian response to refugee crises,

    M. Madianou, “Technocolonialism: Digital innovation and data practices in the humanitarian response to refugee crises,”Social Media + Society, vol. 5, no. 3, 2019

    2019

  11. [11]

    Under- standing social media disclosures of sexual abuse through the lenses of support seeking and anonymity,

    N. Andalibi, O. L. Haimson, M. De Choudhury, and A. Forte, “Under- standing social media disclosures of sexual abuse through the lenses of support seeking and anonymity,” pp. 3906–3918, 2016

    2016

  12. [12]

    S. K. Kapp, Ed.,Autistic Community and the Neurodiversity Movement: Stories from the Frontline. Singapore: Springer Nature, 2020

    2020

  13. [13]

    Kukutai and J

    T. Kukutai and J. Taylor, Eds.,Indigenous Data Sovereignty: Toward an Agenda. Canberra, Australia: ANU Press, 2016

    2016

  14. [14]

    Language (technology) is power: A critical survey of

    S. L. Blodgett, S. Barocas, H. Daum III, and H. Wallach, “Language (technology) is power: A critical survey of ”bias” in nlp,” pp. 5454– 5476, 2020

    2020

  15. [15]

    Realtoxicityprompts: Evaluating neural toxic degeneration in language models,

    S. Gehman, S. Gururangan, M. Sap, Y . Choi, and N. A. Smith, “Realtoxicityprompts: Evaluating neural toxic degeneration in language models,” pp. 3356–3369, 2020

    2020

  16. [16]

    Ethical and social risks of harm from language models,

    L. Weidinger, J. Mellor, M. Rauh, C. Griffin, J. Uesato, P.-S. Huang, M. Cheng, A. Glaese, B. Balle, A. Kasirzadehet al., “Ethical and social risks of harm from language models,” 2021. [Online]. Available: https://arxiv.org/abs/2112.04359

    Pith/arXiv arXiv 2021

  17. [17]

    Automated hate speech detection and the problem of offensive language,

    T. Davidson, D. Warmsley, M. Macy, and I. Weber, “Automated hate speech detection and the problem of offensive language,” inProceedings of the Eleventh International AAAI Conference on Web and Social Media, 2017, pp. 512–515

    2017

  18. [18]

    Constitutional ai: Harmlessness from ai feedback,

    Y . Bai, S. Kadavath, S. Kundu, A. Askell, J. Kernion, A. Jones, A. Chen, A. Goldie, A. Mirhoseini, C. McKinnonet al., “Constitutional ai: Harmlessness from ai feedback,” 2022. [Online]. Available: https://arxiv.org/abs/2212.08073

    Pith/arXiv arXiv 2022

  19. [19]

    Ethically aligned design: A vision for prioritizing human well-being with autonomous and intelligent systems,

    IEEE Standards Association, “Ethically aligned design: A vision for prioritizing human well-being with autonomous and intelligent systems,” 2021. [Online]. Available: https: //standards.ieee.org/industry-connections/ec/autonomous-systems/

    2021

  20. [20]

    Russell and P

    S. Russell and P. Norvig,Artificial Intelligence: A Modern Approach, 4th ed. Pearson, 2020

    2020

  21. [21]

    Training language models to follow instructions with human feedback,

    L. Ouyang, J. Wu, X. Jiang, D. Almeida, C. Wainwright, P. Mishkin, C. Zhang, S. Agarwal, K. Slama, A. Rayet al., “Training language models to follow instructions with human feedback,” pp. 27 730–27 744, 2022

    2022

  22. [22]

    Dspy: Compiling declarative language model calls into self-improving pipelines,

    O. Khattab, C. Potts, and M. Zaharia, “Dspy: Compiling declarative language model calls into self-improving pipelines,” 2023. [Online]. Available: https://arxiv.org/abs/2310.03714

    Pith/arXiv arXiv 2023

  23. [23]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. Chi, Q. V . Le, and D. Zhou, “Chain-of-thought prompting elicits reasoning in large language models,”Advances in Neural Information Processing Systems, vol. 35, pp. 24 824–24 837, 2022

    2022

  24. [24]

    Tree of thoughts: Deliberate problem solving with large language models,

    S. Yao, D. Yu, J. Zhao, I. Shafran, T. L. Griffiths, Y . Cao, and K. Narasimhan, “Tree of thoughts: Deliberate problem solving with large language models,”Advances in Neural Information Processing Systems, vol. 36, 2023

    2023

  25. [25]

    Take a step back: Evoking reasoning via abstraction in large language models,

    H. S. Zheng, S. Mishra, X. Chen, H.-T. Cheng, E. H. Chi, Q. V . Le, and D. Zhou, “Take a step back: Evoking reasoning via abstraction in large language models,” 2023. [Online]. Available: https://arxiv.org/abs/2310.06117

    Pith/arXiv arXiv 2023

  26. [26]

    React: Synergizing reasoning and acting in language models,

    S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y . Cao, “React: Synergizing reasoning and acting in language models,” in International Conference on Learning Representations, 2023. [Online]. Available: https://arxiv.org/abs/2210.03629

    Pith/arXiv arXiv 2023

  27. [27]

    Self-critiquing models for assisting human evaluators,

    W. Saunders, C. Yeh, J. Wu, S. Bills, L. Ouyang, J. Ward, J. Gray, G. Krueger, R. Lowe, J. Leikeet al., “Self-critiquing models for assisting human evaluators,” 2022. [Online]. Available: https://arxiv.org/abs/2206.05802

    Pith/arXiv arXiv 2022

  28. [28]

    Self-refine: Iterative refinement with self-feedback,

    A. Madaan, N. Tandon, P. Gupta, S. Hallinan, L. Gao, S. Wiegreffe, U. Alon, N. Dziri, S. Prabhumoye, Y . Yanget al., “Self-refine: Iterative refinement with self-feedback,” 2023. [Online]. Available: https://arxiv.org/abs/2303.17651

    Pith/arXiv arXiv 2023

  29. [29]

    A prompt pattern catalog to enhance prompt engineering with chatgpt,

    J. White, Q. Fu, S. Hays, M. Sandborn, C. Olea, H. Gilbert, A. Elnashar, J. Spencer-Smith, and D. C. Schmidt, “A prompt pattern catalog to enhance prompt engineering with chatgpt,” 2023. [Online]. Available: https://arxiv.org/abs/2302.11382

    Pith/arXiv arXiv 2023

  30. [30]

    Prompt programming for large language models: Beyond the few-shot paradigm,

    L. Reynolds and K. McDonell, “Prompt programming for large language models: Beyond the few-shot paradigm,” pp. 1–7, 2021

    2021

  31. [31]

    Calibrating noise to sensitivity in private data analysis,

    C. Dwork, F. McSherry, K. Nissim, and A. Smith, “Calibrating noise to sensitivity in private data analysis,” inTheory of Cryptography Conference. Springer, 2006, pp. 265–284

    2006

  32. [32]

    Extracting training data from large language models,

    N. Carlini, F. Tramr, E. Wallace, M. Jagielski, A. Herbert-V oss, K. Lee, A. Roberts, T. Brown, D. Song, . Erlingssonet al., “Extracting training data from large language models,” pp. 2633–2650, 2021

    2021

  33. [33]

    An empirical survey of the effectiveness of debiasing techniques for pre-trained language models,

    N. Meade, E. Poole-Dayan, and S. Reddy, “An empirical survey of the effectiveness of debiasing techniques for pre-trained language models,” pp. 1878–1898, 2021

    2021

  34. [34]

    Racial microaggressions in everyday life: Implications for clinical practice,

    D. W. Sue, C. M. Capodilupo, G. C. Torino, J. M. Bucceri, A. M. Holder, K. L. Nadal, and M. Esquilin, “Racial microaggressions in everyday life: Implications for clinical practice,”American Psychologist, vol. 62, no. 4, pp. 271–286, 2007

    2007

  35. [35]

    Machine ethics: Creating an ethical intelligent agent,

    M. Anderson and S. L. Anderson, “Machine ethics: Creating an ethical intelligent agent,”AI Magazine, vol. 28, no. 4, pp. 15–26, 2007

    2007

  36. [36]

    Wallach and C

    W. Wallach and C. Allen,Moral Machines: Teaching Robots Right from Wrong. Oxford University Press, 2008

    2008

  37. [37]

    Microsoft presidio,

    Microsoft, “Microsoft presidio,” https://microsoft.github.io/presidio/, 2024, accessed: 2024-05-13

    2024

  38. [38]

    Lost in PII: Towards fair and private text data with large language models,

    Y . Li, T. Baldwin, and T. Cohn, “Lost in PII: Towards fair and private text data with large language models,” 2023. [Online]. Available: https://arxiv.org/abs/2310.10716

    Pith/arXiv arXiv 2023

  39. [39]

    A survey on content moderation and its challenges in social media,

    A. R. Fabbri, W. Kryscinski, B. McCann, C. Xiong, R. Socher, and D. Radev, “A survey on content moderation and its challenges in social media,”ACM Computing Surveys, vol. 54, no. 3, pp. 1–36, 2021

    2021

  40. [40]

    k-anonymity: A model for protecting privacy,

    L. Sweeney, “k-anonymity: A model for protecting privacy,”Interna- tional Journal of Uncertainty, Fuzziness and Knowledge-Based Systems, vol. 10, no. 05, pp. 557–570, 2002

    2002

  41. [41]

    Perspective API,

    Jigsaw, “Perspective API,” 2025, accessed: 2025-05-13. [Online]. Available: https://www.perspectiveapi.com/

    2025

  42. [42]

    Rephrase and respond: Let large language models ask better questions for themselves,

    Y . Deng, W. Zhang, Z. Chen, and Q. Gu, “Rephrase and respond: Let large language models ask better questions for themselves,” 2023. [Online]. Available: https://arxiv.org/abs/2311.04205

    Pith/arXiv arXiv 2023

  43. [43]

    Large language models understand and can be enhanced by emotional stimuli,

    C. Li, J. Wang, Y . Zhang, K. Zhu, W. Wang, P. Cao, and X. Xie, “Large language models understand and can be enhanced by emotional stimuli,” 2023. [Online]. Available: https://arxiv.org/abs/2307.11760

    Pith/arXiv arXiv 2023

  44. [44]

    Schuler and A

    D. Schuler and A. Namioka, Eds.,Participatory Design: Principles and Practices. Boca Raton, FL: CRC Press, 1993

    1993

  45. [45]

    Self-instruct: Aligning language models with self-generated instructions,

    Y . Wang, Y . Kordi, S. Mishra, A. Liu, N. A. Smith, D. Khashabi, and H. Hajishirzi, “Self-instruct: Aligning language models with self-generated instructions,” 2022. [Online]. Available: https://arxiv.org/ abs/2212.10560

    Pith/arXiv arXiv 2022

  46. [46]

    [Online]

    National Library of Medicine, “PubMed,” 2025, accessed: 2025-05-13. [Online]. Available: https://pubmed.ncbi.nlm.nih.gov/

    2025

  47. [47]

    UNHCR refugee data finder,

    UNHCR, “UNHCR refugee data finder,” 2025, accessed: 2025-05-13. [Online]. Available: https://www.unhcr.org/refugee-statistics/

    2025

  48. [48]

    Aequitas: A bias and fairness audit toolkit,

    P. Saleiro, B. Kuester, L. Hinkson, J. London, A. Stevens, A. Anisfeld, K. T. Rodolfa, and R. Ghani, “Aequitas: A bias and fairness audit toolkit,” 2018. [Online]. Available: https://arxiv.org/abs/1811.05577

    Pith/arXiv arXiv 2018

  49. [49]

    Active retrieval augmented generation,

    Z. Jiang, F. F. Xu, J. Araki, and G. Neubig, “Active retrieval augmented generation,” 2023. [Online]. Available: https://arxiv.org/abs/2305.06983

    Pith/arXiv arXiv 2023

  50. [50]

    Judging LLM-as-a- judge with MT-bench and chatbot arena,

    L. Zheng, W.-L. Chiang, Y . Sheng, S. Zhuang, Z. Wu, Y . Zhuang, Z. Lin, Z. Li, D. Li, E. P. Xinget al., “Judging LLM-as-a- judge with MT-bench and chatbot arena,” 2023. [Online]. Available: https://arxiv.org/abs/2306.05685

    Pith/arXiv arXiv 2023

  51. [51]

    Larger language models do in-context learning differently,

    J. Wei, J. Wei, Y . Tay, D. Tran, A. Webson, Y . Lu, X. Chen, H. Liu, D. Huang, D. Zhouet al., “Larger language models do in-context learning differently,” 2023. [Online]. Available: https: //arxiv.org/abs/2303.03846

    Pith/arXiv arXiv 2023

  52. [52]

    What makes good in-context examples for gpt-3?

    J. Liu, D. Shen, Y . Zhang, B. Dolan, L. Carin, and W. Chen, “What makes good in-context examples for gpt-3?” 2022. [Online]. Available: https://arxiv.org/abs/2101.06804

    Pith/arXiv arXiv 2022

  53. [53]

    Seer: A knapsack approach to exemplar selection for in-context learning,

    J. Su, T. Y . Zhuo, D. Wang, and P. Nakov, “Seer: A knapsack approach to exemplar selection for in-context learning,” inProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 2023, pp. 13 520–13 534

    2023

  54. [54]

    Measuring and narrowing the compositionality gap in language models,

    O. Press, M. Zhang, S. Min, L. Schmidt, N. A. Smith, and M. Lewis, “Measuring and narrowing the compositionality gap in language models,” 2022. [Online]. Available: https://arxiv.org/abs/2210.03350

    Pith/arXiv arXiv 2022

  55. [55]

    Self-consistency improves chain of thought reasoning in language models,

    X. Wang, J. Wei, D. Schuurmans, Q. Le, E. Chi, S. Narang, A. Chowdhery, and D. Zhou, “Self-consistency improves chain of thought reasoning in language models,” 2022. [Online]. Available: https://arxiv.org/abs/2203.11171

    Pith/arXiv arXiv 2022

  56. [56]

    Large language models as analogical reasoners,

    M. Yasunaga and J. Leskovec, “Large language models as analogical reasoners,” 2023. [Online]. Available: https://arxiv.org/abs/2310.01714

    Pith/arXiv arXiv 2023

  57. [57]

    A survey of large language models,

    W. X. Zhao, K. Zhou, J. Li, T. Tang, X. Wang, Y . Hou, Y . Min, B. Zhang, J. Zhang, Z. Donget al., “A survey of large language models,” 2023. [Online]. Available: https://arxiv.org/abs/2303.18223

    Pith/arXiv arXiv 2023

  58. [58]

    On measuring social biases in sentence encoders,

    C. May, A. Wang, S. Bordia, S. R. Bowman, and R. Rudinger, “On measuring social biases in sentence encoders,” inProceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2019, pp. 622–628

    2019

  59. [59]

    Semantics derived au- tomatically from language corpora contain human-like biases,

    A. Caliskan, J. J. Bryson, and A. Narayanan, “Semantics derived au- tomatically from language corpora contain human-like biases,”Science, vol. 356, no. 6334, pp. 183–186, 2017

    2017

  60. [60]

    Bertscore: Evaluating text generation with bert,

    T. Zhang, V . Kishore, F. Wu, K. Q. Weinberger, and Y . Artzi, “Bertscore: Evaluating text generation with bert,” inInternational Conference on Learning Representations, 2020

    2020

  61. [61]

    Co-creation and the new landscapes of design,

    E. B.-N. Sanders and P. J. Stappers, “Co-creation and the new landscapes of design,”CoDesign, vol. 4, no. 1, pp. 5–18, 2008

    2008

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.