Pith. sign in

REVIEW 4 major objections 4 minor 52 references

LEX-EC: A Lexical Evidence-Channel Audit Framework for Zero-Shot LLM Personality Classification in Black-Box Settings

T0 review · 4 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read Zero-shot LLM personality prediction from a single text is not a unitary capability: it is configuration-specific, varying by genre, trait, and the lexical evidence that remains after masking.

desk verdict A careful, honest empirical audit whose main qualitative claim is believable, but the masking control gap and single-run statistics keep the quantitative conclusions brittle. read the letter →

arxiv 2607.24435 v2 pith:XSDI2VFQ submitted 2026-07-27 cs.CL cs.AI

Brittany Harbison , Ashok K. Goel This is my paper
classification cs.CLcs.AI
keywords zero-shotpersonalityclassificationBigFiveblack-boxinterpretabilitylexicalablationcontentmaskingLLMpsycholinguisticlexiconsagreementmetrics
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

Large language models readily assign Big Five labels to ordinary text, but the paper argues that a plausible label is not necessarily a valid measurement. It introduces an audit procedure that restricts inputs to a retained psycholinguistic channel—function words, affective terms, and cognitive-style vocabulary—and asks which trait associations survive. Across three genres, some associations (Extraversion in self-introductions, Conscientiousness in essays) weaken or vanish when topical and demographic content is masked, while others (Openness, Agreeableness in essays) remain detectable. Single-status microtexts yield little stable evidence even with balanced labels, suggesting a lower bound of content or length. The central conclusion is that zero-shot Big Five prediction is best characterized as a property of a model–genre–trait–lexicon configuration, not a general capability, so systems should be evaluated separately for each intended text condition.

What carries the argument

The load-bearing instrument is a content-masking ablation: every token outside a pre-specified retained channel—function-word POS classes, supplemental quantifiers and negations, and selected Empath and NRC affective and cognitive-style categories—is replaced with an underscore. This turns prediction into a retention test under restricted evidence. Association is measured by Spearman's rho (equal to the phi coefficient for binary labels) and Cohen's kappa for chance-corrected agreement, with Fisher or chi-square tests, compared before and after masking across models and prompts.

What would settle it

Mask the same texts with randomly selected tokens (or a reversed channel that keeps topical words and masks function words) and rerun the ablation; if trait associations persist equally under random masking, the reported persistence reflects the model's response to distortion, not to the retained lexical channel.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central result is not simply that performance was weak: associations differed in their sensitivity to lexical masking. Masking all tokens outside function-word, affective, and cognitive-style vocabulary attenuated some trait associations (Extraversion in student introductions, Conscientiousness in essays) while leaving others (Openness, Agreeableness in essays) comparatively intact. The pattern repeated across five closed models, suggesting that genre and lexical channel, rather than model identity, drove the variation. Prompts that asked for linguistic justifications shifted the distribution of self-explanations but did not eliminate topical reasoning, and expl

Load-bearing premise

The mapping that replaces every token outside the retained lexical channel with an underscore assumes the model reacts to the preserved words rather than to the systematic distortion itself; without a random-masking control, a surviving association could be an artifact of the transformation.

Editorial extensions

If this is right

  • If a trait association survives masking, it is recoverable from function words, affective terms, and cognitive-style vocabulary alone; if it attenuates, broader topical content contributed to the original judgment.
  • Single-status microtexts approach a lower boundary of usable evidence for zero-shot Big Five prediction, even when labels are balanced.
  • Prompt wording can change the distribution of model self-explanations without eliminating topical reasoning or consistently changing predictive accuracy.
  • No model showed a stable advantage across datasets and conditions, indicating that observed patterns may generalize to newer model families.
  • Statistically detectable associations did not translate into reliable individual-level classification, as chance-corrected agreement stayed low.

Reading between the lines

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

  • A direct control test would be random-token masking: if associations persist equally when non-psycholinguistic tokens are replaced at random, persistence reflects robustness to distortion rather than evidence carried by the retained lexical channel.
  • The framework could extend to other closed-model classification tasks where the question is whether a prediction rests on topical content or on recoverable style markers, treating persistence under restriction as a diagnostic of evidence channel.
  • Because the microtext condition confounds short length with genre, a length-matched control (truncated essays) could separate the lower-bound effect of content scarcity from genre-specific style.
  • The shift in self-explanations under the linguistic prompt is a behavioral change, not proof that the model's computations now use linguistic evidence; the paper's own framing supports this reserved reading.
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 / 4 minor

Summary. The paper introduces LEX-EC, a black-box audit framework for zero-shot LLM Big Five personality classification. The framework combines prevalence checks, item-level association (Spearman rho, Cohen's kappa), a lexical content-masking ablation that retains only function words and hand-selected Empath/NRC categories, and a prompt-sensitivity audit of model self-explanations. It is applied to three datasets (Pennebaker-King essays, graduate student introduction posts, myPersonality Facebook statuses), five closed models, and three prompt conditions. The paper reports heterogeneous results across genre, trait, model, and prompt: associations are generally weak; essay Conscientiousness attenuates after masking, post Extraversion attenuates, Facebook associations largely collapse, and the linguistic prompt shifts stated justifications toward linguistic content. The main conclusion is that zero-shot BFI prediction from a single text is not a unitary capability, and that different traits show different sensitivity to lexical restriction.

Significance. If the central claims hold, the paper makes a useful methodological contribution: it provides a reusable, code-shared black-box audit that treats LLM personality outputs as behaviors rather than measurements, and it offers a concrete caution against treating zero-shot personality prediction as a general model capability. The framework is transparent and carefully hedged, and the authors explicitly disavow any claim that persistence under masking demonstrates better prediction. The main value is in distinguishing marginal-distribution effects from item-level signal and in showing that ablation effects are trait- and genre-dependent. The strength of this contribution is currently limited by the absence of a control for the masking transformation itself; the central RQ2 inference is therefore not yet established beyond reasonable doubt.

major comments (4)
  1. [Content-Masking Ablation, Eq. (5); Discussion] The inference that a surviving association 'persists from the retained lexical channel' assumes the model is insensitive to the transformation itself. Because sigma collapses consecutive placeholders and all non-retained tokens become '_', ablated inputs are severely shortened and syntactically broken. The paper cites Feng et al. (2018) for exactly this failure mode, and the Limitations section concedes that removing material 'may disrupt correspondence with data consumed during training,' but no control condition is run. A random-mask condition (masking an equal fraction of tokens without regard to lexical category) or a reversed-channel condition (retaining the removed content and masking the psycholinguistic layer) is needed. Without it, the differential pattern in Tables 1-2 could be produced by trait-specific sensitivity to length reduction or underscore artifacts rather than by the
  2. [Methodology: Model Selection and Output Collection; Tables 1-2] The paper states that 'unless otherwise stated, all results were each gathered via a single run.' The main table deltas are small (e.g., full-corpus essay Openness rho .168 to .158; Facebook Neuroticism rho .184 to -.058), and some models are stochastic (reasoning models with no fixed seed). Without confidence intervals or repeated runs, the reported Delta-rho and Delta-kappa values cannot be distinguished from API noise. Additionally, the significance thresholds are unadjusted across roughly 150 trait/model/prompt cells, so the count of 'statistically detectable' pre-ablation associations is likely to include false positives. Reporting bootstrap intervals or at least a multiple-comparison sensitivity check would substantially strengthen the quantitative claims.
  3. [Tables 10-11; Experiments: Prediction and Ablation Results] Matched Facebook ablations were run only for GPT-4o (two prompts), although pre-ablation significant associations also appear for other models (e.g., Claude Haiku Conscientiousness under LINEX in Table 10). The Discussion's claim that 'the isolated Facebook associations disappeared after masking' is therefore based on a single model family. Either provide matched ablation analyses for all models with pre-ablation signal or explicitly restrict the claim to GPT-4o. This is important because the Facebook lower-boundary finding is one of the paper's headline results.
  4. [Content-Masking Ablation, Eq. (3)] The retained channel is constructed from a hand-selected set of Empath and NRC categories. The paper argues for construct validity and notes the high Empath-LIWC correlation, but there is no sensitivity analysis over the category set. Since the central claim is differential sensitivity to lexical masking, the persistence/attenuation pattern could be an artifact of the particular category choices (e.g., a too-narrow set for Conscientiousness, a too-broad set for Agreeableness). A robustness check varying the category set, or at least a per-category breakdown, would make the inference more secure.
minor comments (4)
  1. [Throughout] The trait name is spelled 'Extraversion' in the main text but 'Extroversion' in Tables 3-6, Listing 1, and Tables 10-11. Standardize the spelling.
  2. [Appendix Tables 10-11] The '???' entries indicate single-class prediction only, but this is not defined in the table notes. Specify how these cells were excluded from the median Delta calculations.
  3. [Abstract] The phrase 'some trait associations are recoverable from function words, affective terms, and cognitive-style vocabulary' overstates the finding, given the lack of a masking-transformation control. 'Remain detectable after masking' would be more accurate at this stage.
  4. [Discussion, Facebook lower-boundary claim] The 'possible lower bound of content or length' for Facebook statuses is stated informally. Because the matched matrix is GPT-4o-only and n=164, describe this as a hypothesis or condition-specific observation rather than a general boundary.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: empirical audit with a fixed a priori lexical channel; minor same-author citation is not load-bearing.

full rationale

LEX-EC is a behavioral audit, not a derivation. The retained lexical channel is fixed in advance from Empath and NRC categories (Eq. 3), no parameters are fitted to trait labels, and the central claims are descriptive before/after associations (Tables 1-2). The statement 'Persistence indicates that an association remains recoverable from the retained channel' is an operational definition of the ablation, not a hidden reduction of a prediction to its input. The paper explicitly acknowledges the main validity threat: 'by removing all but psycholinguistic material during ablation, this may disrupt correspondence with data consumed during training.' The absence of a random-mask control is a confounding/correctness concern, not circularity. The only same-author citation (Harbison et al. 2026) supplies the binarization midpoint and student corpus; it does not carry the central conclusion that trait associations vary by genre, lexical channel, and prompt. The empirical content is independent of its inputs, so the paper warrants a low non-circularity score.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The paper introduces no fitted coefficients and no new physical or mathematical entities. However, the central empirical findings depend on several hand-set analysis choices (midpoint tau, balancing delta, seed, and the retained lexicon set) and on domain assumptions about the validity of the masking channel. The weakest load-bearing premise is that masked placeholder text isolates lexical evidence rather than introducing distortion artifacts.

free parameters (4)
  • Trait binarization midpoint τ = 3 (fixed scale midpoint)
    Appendix Eq. (6); student-post BFI-44 scores dichotomized at >3. For O, C, A this yields 84-90% majority-class labels, making binary rho/kappa low-information.
  • Facebook trait-balancing tolerance δ = 0.05
    Eq. (1); MILP constraint keeps each trait's positive-label proportion in [0.45,0.55]. Chosen by hand; determines the n=164 sample.
  • Random seed 67 = 67
    Used for Facebook user sampling and the 226-item essay downsampling; no sensitivity analysis is shown.
  • Retained lexicon category set = Table 7: 17 Empath and 8 NRC subcategories
    Defines V_psych in Eq. (3). The central claim about which traits persist after masking is relative to this hand-selected set, guided by reported Empath-LIWC correlation r=.906 rather than an independent benchmark.
assumptions (6)
  • domain assumption Binary high/low BFI labels derived by midpoint threshold preserve meaningful trait classes.
    Appendix midpoint procedure; heavily skewed distributions for O/C/A in student posts make binary association measures weakly informative.
  • domain assumption The retained lexical channel V_psych plus function-word POS classes is a valid operationalization of psycholinguistic evidence.
    Eq. (3)-(5); relies on Empath-LIWC correlation r=.906 (Fast et al. 2016) and the selection of NRC categories.
  • domain assumption Masked placeholder '_' is lexically neutral and does not itself shift LLM trait judgments.
    Methodology, masking placeholder mu; the paper argues it avoids introducing associations but provides no control for distortion artifacts such as length reduction or broken syntax.
  • standard math Spearman's rho for binary variables equals phi and Cohen's kappa with observed marginals is an appropriate chance-corrected agreement measure.
    Eq. (2); standard definitions.
  • domain assumption LLM outputs parsed from required structured tool calls reflect the model's labels without parsing or decoding error.
    Algorithm 1; schema-constrained decoding was avoided because it degraded pilot association statistics, so reliance on tool-call parsing is load-bearing.
  • domain assumption Model snapshots are representative of their model families for this task.
    Model Selection; full-corpus essay analysis was run only with GPT-5.4-mini/LINEX, based on pilot similarity across models.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LEX-EC: A Lexical Evidence-Channel Audit Framework for Zero-Shot LLM Personality Classification in Black-Box Settings." pith.science (2026). https://pith.science/paper/XSDI2VFQ

@misc{pith2026260724435,
  author       = {Pith},
  title        = {Pith review of: LEX-EC: A Lexical Evidence-Channel Audit Framework for Zero-Shot LLM Personality Classification in Black-Box Settings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XSDI2VFQ}},
  note         = {Machine review of arXiv:2607.24435}
}
read the original abstract

Large language models may easily assign personality labels from text, but model interpretability remains an open problem. To address this gap, we introduce LEX-EC, a reusable black-box audit framework combining prevalence and agreement diagnostics with controlled lexical ablation to distinguish marginal-distribution effects from trait-associated signal recoverable under restricted evidence. Using this framework, we illustrate how various text genres may exhibit sharply different profiles: free-form essay text contains the broadest, but still weak, signal; in graduate student introductions, an observable Extraversion association weakened after masking; and single Facebook statuses yield little stable evidence even in a trait-balanced sample, indicating a possible lower bound of content or length. Masking topical and demographic content weakened some associations while leaving others detectable from function words, affective terms, and cognitive-style vocabulary. Linguistic prompting shifted model self-explanations but did not eliminate topical content. LEX-EC jointly evaluates classification prevalence, item-level association, chance-corrected agreement, persistence under lexical restriction, and prompt sensitivity in model-generated explanations. Across datasets, models, and prompts, LEX-EC characterizes how trait associations may vary with available lexical evidence, introducing a novel application of lexical methods to black-box interpretability in personality labeling.

Figures

Figures reproduced from arXiv: 2607.24435 by the authors.

Figure 1
Figure 1. Document prevalence and corpus-level density of retained Empath and NRC categories. Prevalence is the percentage [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. LEX-EC processing stages. Prevalence and item-level diagnostics address RQ1, full-versus-masked comparisons [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

52 extracted references · 2 linked inside Pith

  1. [1]

    , title =

    Goldberg, Lewis R. , title =. Journal of Personality and Social Psychology , year =

  2. [2]

    and John, Oliver P

    McCrae, Robert R. and John, Oliver P. , title =. Journal of Personality , year =

  3. [3]

    IEEE Transactions on Affective Computing , year =

    Vinciarelli, Alessandro and Mohammadi, Gelareh , title =. IEEE Transactions on Affective Computing , year =

  4. [4]

    and Kosinski, Michal and Nave, Gideon and Stillwell, David J

    Matz, Sandra C. and Kosinski, Michal and Nave, Gideon and Stillwell, David J. , title =. Proceedings of the National Academy of Sciences , year =

  5. [5]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , year =

    Rao, Haocong and Leung, Cyril and Miao, Chunyan , title =. Findings of the Association for Computational Linguistics: EMNLP 2023 , year =

    2023

  6. [6]

    , title =

    Peters, Heinrich and Matz, Sandra C. , title =. PNAS Nexus , year =

  7. [7]

    Derner, Erik and Ku. Can. Computers in Human Behavior: Artificial Humans , year =

  8. [8]

    Communications Psychology , year =

    Schoenegger, Philipp and Greenberg, Spencer and Grishin, Alexander and Lewis, Joshua and Caviola, Lucius , title =. Communications Psychology , year =

Show all 52 references
  1. [9]

    Computational Linguistics , year =

    Belinkov, Yonatan , title =. Computational Linguistics , year =

  2. [10]

    Distill , year =

    Olah, Chris and Cammarata, Nick and Schubert, Ludwig and Goh, Gabriel and Petrov, Michael and Carter, Shan , title =. Distill , year =

  3. [11]

    Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , year =

    Ribeiro, Marco Tulio and Singh, Sameer and Guestrin, Carlos , title =. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , year =

  4. [12]

    and Lee, Su-In , title =

    Lundberg, Scott M. and Lee, Su-In , title =. Advances in Neural Information Processing Systems 30 , year =

  5. [13]

    Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , year =

    Ribeiro, Marco Tulio and Wu, Tongshuang and Guestrin, Carlos and Singh, Sameer , title =. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , year =

  6. [14]

    and Costa, Jr., Paul T

    McCrae, Robert R. and Costa, Jr., Paul T. , title =. Journal of Personality and Social Psychology , year =

  7. [15]

    and Srivastava, Sanjay , title =

    John, Oliver P. and Srivastava, Sanjay , title =. Handbook of Personality: Theory and Research , editor =. 1999 , publisher =

    1999

  8. [16]

    and Costa, Jr., Paul T

    McCrae, Robert R. and Costa, Jr., Paul T. , title =. 1992 , publisher =

    1992

  9. [17]

    and King, Laura A

    Pennebaker, James W. and King, Laura A. , title =. Journal of Personality and Social Psychology , year =

  10. [18]

    Journal of Research in Personality , year =

    Yarkoni, Tal , title =. Journal of Research in Personality , year =

  11. [19]

    Andrew and Eichstaedt, Johannes C

    Schwartz, H. Andrew and Eichstaedt, Johannes C. and Kern, Margaret L. and Dziurzynski, Lukasz and Ramones, Stephanie M. and Agrawal, Megha and Shah, Achal and Kosinski, Michal and Stillwell, David J. and Seligman, Martin E. P. and Ungar, Lyle H. , title =. PLOS ONE , year =

  12. [20]

    , title =

    Fast, Ethan and Chen, Binbin and Bernstein, Michael S. , title =. Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems , year =

    2016

  13. [21]

    and Turney, Peter D

    Mohammad, Saif M. and Turney, Peter D. , title =. Computational Intelligence , year =

  14. [22]

    Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , year =

    Jacovi, Alon and Goldberg, Yoav , title =. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , year =

  15. [23]

    , title =

    Turpin, Miles and Michael, Julian and Perez, Ethan and Bowman, Samuel R. , title =. Advances in Neural Information Processing Systems 36 , year =

  16. [24]

    and Donahue, Eileen M

    John, Oliver P. and Donahue, Eileen M. and Kentle, Robert L. , title =

  17. [25]

    Frontiers in Artificial Intelligence , year =

    Piastra, Marco and Catellani, Patrizia , title =. Frontiers in Artificial Intelligence , year =

  18. [26]

    2020 IEEE International Conference on Data Mining (ICDM) , year =

    Mehta, Yash and Fatehi, Samin and Kazameini, Amirmohammad and Stachl, Clemens and Cambria, Erik and Eetemadi, Sauleh , title =. 2020 IEEE International Conference on Data Mining (ICDM) , year =

    2020

  19. [27]

    Neural Processing Letters , year =

    Mohades Deilami, Fatemeh and Sadr, Hossein and Tarkhan, Morteza , title =. Neural Processing Letters , year =

  20. [28]

    and Lin, Hongfei and Xu, Bo and Yang, Liang , title =

    Tadesse, Michael M. and Lin, Hongfei and Xu, Bo and Yang, Liang , title =. IEEE Access , year =

  21. [29]

    Proceedings of the International AAAI Conference on Web and Social Media , year =

    Celli, Fabio and Pianesi, Fabio and Stillwell, David and Kosinski, Michal , title =. Proceedings of the International AAAI Conference on Web and Social Media , year =

  22. [30]

    Goel , title =

    Brittany Harbison and Samuel Taubman and Travis Taylor and Ashok K. Goel , title =. INTED2026 Proceedings , year =

  23. [31]

    Details Omitted for Double-Blind Review , year =

  24. [32]

    and Pennebaker, James W

    Tausczik, Yla R. and Pennebaker, James W. , title =. Journal of Language and Social Psychology , volume =

  25. [33]

    First Monday , year =

    Hayes, Beatrice and Suleiman, Aiman and Watling, Dawn , title =. First Monday , year =

  26. [34]

    Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , year =

    Feng, Shi and Wallace, Eric and Grissom II, Alvin and Iyyer, Mohit and Rodriguez, Pedro and Boyd-Graber, Jordan , title =. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing , year =

    2018

  27. [35]

    arXiv preprint arXiv:1702.08608 , year =

    Doshi-Velez, Finale and Kim, Been , title =. arXiv preprint arXiv:1702.08608 , year =. 1702.08608 , archivePrefix =

    arXiv

  28. [36]

    and Barry, Jon and Cohn, Anthony G

    Blackwell, Robert E. and Barry, Jon and Cohn, Anthony G. , year =. Towards Reproducible. 2410.03492 , archivePrefix =

    arXiv

  29. [37]

    Register Studies , year =

    Liimatta, Aatu , title =. Register Studies , year =

  30. [38]

    International Journal of Corpus Linguistics , year =

    Liimatta, Aatu , title =. International Journal of Corpus Linguistics , year =

  31. [39]

    2024 , month = nov, pages =

    Text Length and the Function of Intentionality: A Case Study of Contrastive Subreddits , booktitle =. 2024 , month = nov, pages =. doi:10.18653/v1/2024.nlp4dh-1.1 , url =

    2024 doi

  32. [40]

    Findings of the Association for Computational Linguistics: ACL 2024 , month = aug, year =

    Are Self-Explanations from Large Language Models Faithful? , author =. Findings of the Association for Computational Linguistics: ACL 2024 , month = aug, year =

    2024

  33. [41]

    Journal of Medical Internet Research , year =

    Maharjan, Julina and Jin, Ruoming and Zhu, Jianfeng and Kenne, Deric , title =. Journal of Medical Internet Research , year =

  34. [42]

    Maurits and Bleich, Erik , title =

    van der Veen, A. Maurits and Bleich, Erik , title =. PLOS ONE , year =

  35. [43]

    2025 , month = aug, eprint =

    Compton, Thomas , title =. 2025 , month = aug, eprint =

    2025

  36. [44]

    When ``A Helpful Assistant'' Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models

    Zheng, Mingqian and Pei, Jiaxin and Logeswaran, Lajanugen and Lee, Moontae and Jurgens, David. When ``A Helpful Assistant'' Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models. Findings of the Association for Computational Lin...

    2024 doi

  37. [45]

    Better Zero-Shot Reasoning with Role-Play Prompting

    Kong, Aobo and Zhao, Shiwan and Chen, Hao and Li, Qicheng and Qin, Yong and Sun, Ruiqi and Zhou, Xin and Wang, Enzhi and Dong, Xiaohang. Better Zero-Shot Reasoning with Role-Play Prompting. Proceedings of the 2024 Conference of the North American Chapter of the Association for...

    2024 doi

  38. [46]

    Principled Personas: Defining and Measuring the Intended Effects of Persona Prompting on Task Performance

    Luz de Araujo, Pedro Henrique and R. Principled Personas: Defining and Measuring the Intended Effects of Persona Prompting on Task Performance. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10.18653/v1/2025.emnlp-main.1364

    2025 doi

  39. [47]

    2024 , howpublished =

    2024

  40. [48]

    Explainable and Interpretable Models in Computer Vision and Machine Learning , editor =

    Doshi-Velez, Finale and Kim, Been , title =. Explainable and Interpretable Models in Computer Vision and Machine Learning , editor =. 2018 , doi =

    2018

  41. [49]

    2023 , month = dec, address =

    Yang, Tao and Shi, Tianyuan and Wan, Fanqi and Quan, Xiaojun and Wang, Qifan and Wu, Bingzhe and Wu, Jiaxiang , booktitle =. 2023 , month = dec, address =. doi:10.18653/v1/2023.findings-emnlp.216 , url =

    2023 doi

  42. [50]

    Wright, Aidan G. C. and Ringwald, Whitney R. and Vize, Colin E. and Eichstaedt, Johannes C. and Angstadt, Mike and Taxali, Aman and Sripada, Chandra , journal =. Assessing Personality Using Zero-Shot Generative. 2026 , volume =. doi:10.1038/s41562-025-02389-x , url =

    2026 doi

  43. [51]

    On the Emergent Capabilities of

    Piastra, Marco and Catellani, Patrizia , journal =. On the Emergent Capabilities of. 2025 , volume =. doi:10.3389/frai.2025.1484260 , url =

    2025

  44. [52]

    and Woodard, Damon L

    Bhandarkar, Avanti and Wilson, Ronald and Swarup, Anushka and Webster, Gregory D. and Woodard, Damon L. , booktitle =. 2025 , month = may, address =. doi:10.18653/v1/2025.nlp4dh-1.14 , url =

    2025 doi

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.