Pith. sign in

REVIEW 2 major objections 4 minor 142 references

RPAM: A Principled Metric for Evaluating Associations in Language Models with High Predictive Validity in Downstream Outputs

T0 review · 2 major / 4 minor · reviewed 2026-07-11 · grok-4.5

Pith's one-line read Relative continuation probabilities in language models track human associations and the models' own generated outputs.

desk verdict RPAM is a clean, usable relative-probability metric that finally shows strong upstream-to-downstream association links on open generative models; the main soft spot is template tuning on the same human signals used for validation. read the letter →

arxiv 2607.05679 v1 pith:ODZVNJM6 submitted 2026-07-06 cs.CL cs.AI

classification cs.CLcs.AI
keywords languagemodelsassociationmeasurementbiasevaluationcontinuationprobabilityupstreammetricsWEATvalenceRPAM
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

Language models carry statistical associations that can surface as stereotypes or biased decisions, yet measuring them cleanly has been hard. Downstream checks on generated text need model-specific prompts and datasets that do not transfer. Prior upstream metrics that look at embeddings or raw next-token probabilities have shown only weak links to real-world associations or to what the same models later produce. This paper introduces RPAM, which measures the relative probability that a model continues a target with one attribute word versus a whole set of attributes, then normalizes those probabilities. Across three models of different sizes and purposes and four standard human datasets, the upstream scores line up with implicit human associations, explicit human ratings of similarity and valence, and the models' own downstream ratings and classifications. The result is a single upstream procedure that works across model types and that predicts the associations people actually see in generated text.

What carries the argument

RPAM: the softmax-normalized continuation probability of an attribute word given a target, taken relative to a full set of attributes rather than as an absolute score; the same relative scores are then aggregated into WEAT-style effect sizes or valence scores.

What would settle it

Re-run the same WEAT-WS, Bellezza, WS-353 and SST2 comparisons after replacing the two templates with a fresh set of semantically neutral templates that were never tuned on the validation models; if the correlations collapse, the central claim does not hold.

Watch

Extended reading notes

Core claim

RPAM is the first upstream association metric for generative language models whose measurements show a strong relationship with both human associations (implicit and explicit) and associations measured in the same models' generated text, outperforming prior record values on the same validation tasks where comparisons are possible.

Load-bearing premise

The two hand-tuned templates that insert the target word were chosen by matching human scores on GPT-2; if those templates overfit the validation signals, the reported correlations partly reflect the template choice rather than a general property of relative probabilities.

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

2 major / 4 minor

Summary. The paper introduces RPAM, an upstream association metric for generative LMs that computes softmax-normalized continuation probabilities of attribute words given a target inserted into one of two fixed templates (TP1 for unigrams, TP2 for N-grams). RPAM is extended to RPAM Test (WEAT-style effect sizes) and RPAM Valence (SC-WEAT-style valence). On Mistral-7B-Instruct, Mistral-7B and GPT-2, and on WEAT-WS, WS-353, Bellezza and SST2, the authors report strong congruence with implicit and explicit human associations (e.g., WEAT-WS F1 = 1.0, Bellezza Pearson ρ up to 0.85) and with associations measured in the same models’ generated text (Spearman ρ = 0.73 on Bellezza; F1 ≥ 0.74 on SST2). An ablation without normalization shows a large drop, and results are claimed to outperform prior upstream metrics where comparable.

Significance. If the predictive-validity claims hold after addressing template selection, RPAM would be a practically useful, model-agnostic upstream probe that links continuation probabilities to both human association norms and downstream generation behavior. The formal definitions (Eqs. 1–2), the controlled same-model upstream/downstream design, the multi-model evaluation, and the explicit normalization ablation are concrete strengths that advance the literature beyond absolute-probability or embedding-based WEAT variants. Public code would further increase impact for bias auditing and mitigation studies.

major comments (2)
  1. Appendix D states that TP1 and TP2 were iteratively optimized by comparing RPAM outputs to the same human signals (WS-353, Bellezza, WEAT-WS) later used for validation, and only on GPT-2. Because every reported correlation (Tables 2–4, Fig. 2) is obtained inside these fixed, human-tuned templates, the strongest claim—“first upstream metric with strong real-world predictive validity”—is vulnerable to selection effects. An ablation that freezes templates before any human-signal comparison, or that re-optimizes them on a disjoint stimulus set and re-evaluates transfer to Mistral and Experiment 3, is required to substantiate that the gains arise from relative probabilities rather than template–signal alignment.
  2. Experiment 3 (Table 4) reports high congruence with LM-generated text, yet the downstream tasks themselves are LM-specific (zero-shot rating for Mistral-Instruct, zero-shot classification for Mistral, fine-tuned GPT-2-Sentiment). While the paper correctly notes that specialized prompts are a limitation of downstream metrics, the claim that RPAM “predicts downstream behavior” therefore rests on three different operationalizations. A single, model-agnostic generation protocol (or at least a sensitivity analysis across prompt variants) would strengthen the load-bearing claim that upstream RPAM generalizes to real-world outputs.
minor comments (4)
  1. Table 3 mixes Spearman and Pearson ρ without consistent labeling; clarify which coefficient is used for each row and why the switch occurs relative to Wolfe & Caliskan (2022).
  2. Figure 4 caption and surrounding text discuss prompt refusal and ambiguous generation; the figure itself is hard to parse and could be simplified or moved to the appendix.
  3. Multi-token attribute aggregation (Appendix E) follows Kurita et al. (2019) but is not ablated; a short note on sensitivity would help.
  4. Typographical inconsistencies appear (e.g., “sof tmax”, “Sch ütze”, “WS-WEAT” vs. “WEAT-WS”); a careful pass would improve readability.

Circularity Check

1 steps flagged · score 3.0 of 10

Mild circularity from templates iteratively selected on the same human association datasets later used for validation claims.

  1. fitted input called prediction [Section 4 (Prompting Templates) and Appendix D]
    "we optimized the templates based on preliminary comparative measurements with human associations on GPT-2. Details are in Appendix D. ... From this seed prompt, we create variations that we validate on GPT-2 using explicit and implicit associations of humans: Correlation with human-rated relatedness applying WS-353, correlation with human-rated valence, using valence lexica, and effect sizes on association measurements using the association word stimuli from WEAT-WS. With this iterative approach, we create our final prompts."

    TP1 and TP2 were chosen by maximizing agreement with the exact human association quantities (WS-353 correlations, Bellezza valence, WEAT-WS effect sizes) that Experiments 1-2 later treat as external validation of RPAM's predictive validity and outperformance. The high reported figures (e.g., WEAT-WS F1=1.0, Bellezza Pearson ρ up to 0.85 on GPT-2) therefore partly reflect this selection rather than an independent test of relative probabilities; no pre-selection hold-out or frozen template is reported.

full rationale

The core of RPAM (softmax-normalized continuation probabilities of attribute words given a target in a fixed template) is not definitionally equal to the external human ratings or downstream labels; those remain independent measurements, so there is no self-definitional loop or forced identity of prediction with input. The only circularity is methodological: Appendix D and Section 4 state that the two templates were created by iterative variation and selection against precisely the WS-353, Bellezza valence, and WEAT-WS signals later reported as strong predictive validity (Tables 2-3, Fig. 2). This is a form of fitting the measurement procedure to the evaluation targets on GPT-2, which inflates the reported outperformance relative to prior metrics on the same model. Transfer results on Mistral models and the controlled upstream-downstream comparisons of Experiment 3 still supply independent content once the templates are frozen, so the central claim is not wholly reduced to the fit. No uniqueness theorem, self-citation chain, or renaming of a known result is load-bearing. Score 3 reflects this limited selection leakage without full constructional circularity.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

RPAM rests on standard next-token probability semantics, the WEAT/IAT differential-association construction, and a small set of design choices (two templates, softmax over the attribute set, multi-token product heuristic). No new physical or mathematical entities are postulated. The free parameters are the hand-chosen templates and the conventional WEAT stimulus sets; the main domain assumptions are that relative continuation probabilities capture the same associations that appear in free generation and that the chosen stimuli adequately represent the intended concepts.

free parameters (3)
  • TP1 / TP2 prompting templates
    Two templates were iteratively optimized on GPT-2 by maximizing correlation with human association scores (Appendix D); they are free design choices that affect every reported number.
  • WEAT-WS / Bellezza / WS-353 / SST-2 stimulus sets
    Standard but still hand-curated word and sentence lists; measurement outcomes depend on which stimuli represent each concept (acknowledged via Antoniak & Mimno 2021).
  • Multi-token attribute probability aggregation
    Product of first-token normalized probability and average of remaining subword probabilities (Appendix E); a heuristic choice not derived from first principles.
assumptions (4)
  • ad hoc to paper Softmax-normalized continuation probabilities over a fixed attribute set are a valid measure of relative association strength in generative LMs.
    Core modeling choice of RPAM (Section 4); motivated by psychology relative-comparison findings and by Schick et al. (2021) but not independently proven.
  • domain assumption WEAT/IAT differential association (effect size d) and SC-WEAT valence correctly operationalize the associations of interest.
    Inherited from Caliskan et al. (2017) and Toney-Wails & Caliskan (2021); used throughout Experiments 1–3.
  • domain assumption Semantically bleached templates of the form 'These words are associated: [TARGET] and' isolate association without introducing confounding semantics.
    Assumed after Gonen et al. (2022) and empirical tuning (Appendix D).
  • domain assumption Human-rated association scores (WS-353, Bellezza, SST-2, WEAT-WS) are valid external ground truth for validating LM associations.
    Standard validation practice in the bias-measurement literature; stated in Section 4 Validation framework.
invented entities (1)
  • RPAM (Relative Probability Association Metric)
    purpose: Provide a single upstream scalar (normalized continuation probability or WEAT-style d) that predicts both human associations and same-model downstream behavior.
    The metric is a new named procedure assembled from known components; independent evidence is the reported correlations themselves, not an external physical handle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RPAM: A Principled Metric for Evaluating Associations in Language Models with High Predictive Validity in Downstream Outputs." pith.science (2026). https://pith.science/paper/ODZVNJM6

@misc{pith2026260705679,
  author       = {Pith},
  title        = {Pith review of: RPAM: A Principled Metric for Evaluating Associations in Language Models with High Predictive Validity in Downstream Outputs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ODZVNJM6}},
  note         = {Machine review of arXiv:2607.05679}
}
read the original abstract

Language models (LMs) exhibit problematic biases, such as stereotypes. Effectively analyzing and mitigating such biases requires accurate and generalizable evaluation methods of the underlying associations. Some existing approaches focus on downstream metrics that analyze associations in generated text. Since generated text content can vary drastically across LMs, such metrics often require specialized evaluation datasets, which limits the generalization of such downstream metrics. In contrast, upstream metrics examine LMs at the fundamental level of embeddings or continuation probabilities, enabling principled association analyses across LMs. Yet, to date, no upstream metric for generative LMs has uncovered a strong relationship with real-world associations, including those measured in generated text. To address this gap, we introduce the Relative Probability Association Metric (RPAM), an association evaluation metric for generative LMs. For three LMs of different quality of language generation and purpose (Mistral-7B-Instruct, Mistral-7B, and GPT-2) and well-studied evaluation datasets (WEAT-WS, Bellezza, WS-353, and SST2), we find a strong relationship between upstream RPAM measurements and corresponding implicit and explicit associations observed in humans, as well as biases measured downstream with LM-specific tasks, outperforming prior record values where applicable.

Figures

Figures reproduced from arXiv: 2607.05679 by the authors.

Figure 1
Figure 1. (a) RPAM [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. RPAM Test replicates implicit human associations [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Prompts used for downstream tasks in Experiment [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Challenges in measuring associations downstream: [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Our metric RPAM in comparison to prior work [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: RPAM replicates human-like associations in LMs [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

142 extracted references · 142 canonical work pages

  1. [1]

    , title =

    Jurafsky, Daniel and Martin, James H. , title =. 2000 , isbn =

  2. [2]

    The Pile: An 800GB Dataset of Diverse Text for Language Modeling

    The pile: An 800gb dataset of diverse text for language modeling , author=. arXiv preprint arXiv:2101.00027 , year=

  3. [3]

    Demarginalizing the intersection of race and sex: A black feminist critique of antidiscrimination doctrine, feminist theory and antiracist politics , author=. u. Chi. Legal f. , pages=. 1989 , publisher=

  4. [4]

    OpenAI blog , volume=

    Language models are unsupervised multitask learners , author=. OpenAI blog , volume=

  5. [5]

    Brown, Tom B. and Mann, Benjamin and Ryder, Nick and Subbiah, Melanie and Kaplan, Jared and Dhariwal, Prafulla and Neelakantan, Arvind and Shyam, Pranav and Sastry, Girish and Askell, Amanda and Agarwal, Sandhini and Herbert-Voss, Ariel and Krueger, Gretchen and Henighan, Tom and Child, Rewon and Ramesh, Aditya and Ziegler, Daniel M. and Wu, Jeffrey and W...

  6. [6]

    Advances in neural information processing systems , volume=

    Man is to computer programmer as woman is to homemaker? debiasing word embeddings , author=. Advances in neural information processing systems , volume=

  7. [7]

    , author=

    Intersectional implicit bias: Evidence for asymmetrically compounding bias and the predominance of target gender. , author=. Journal of Personality and Social Psychology , year=

  8. [8]

    , author=

    Measuring individual differences in implicit cognition: the implicit association test. , author=. Journal of personality and social psychology , volume=. 1998 , publisher=

Show all 142 references
  1. [9]

    Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP) , pages=

    Glove: Global vectors for word representation , author=. Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP) , pages=

  2. [10]

    1998 , publisher=

    WordNet: An electronic lexical database , author=. 1998 , publisher=

  3. [11]

    The Journal of Machine Learning Research , volume=

    Exploring the limits of transfer learning with a unified text-to-text transformer , author=. The Journal of Machine Learning Research , volume=. 2020 , publisher=

  4. [12]

    arXiv preprint arXiv:1910.03771 , year=

    Huggingface's transformers: State-of-the-art natural language processing , author=. arXiv preprint arXiv:1910.03771 , year=

  5. [13]

    huggingface.co. Models. 2023

  6. [14]

    Proceedings of the 10th international conference on World Wide Web , pages=

    Placing search in context: The concept revisited , author=. Proceedings of the 10th international conference on World Wide Web , pages=

  7. [15]

    arXiv preprint arXiv:1605.02276 , year=

    Problems with evaluation of word embeddings using word similarity tasks , author=. arXiv preprint arXiv:1605.02276 , year=

  8. [16]

    Bureau of Labor Statistics

    U.S. Bureau of Labor Statistics. Employed persons by detailed occupation, sex, race, and Hispanic or Latino ethnicity. 2023

  9. [17]

    1999 , institution=

    Affective norms for English words (ANEW): Instruction manual and affective ratings , author=. 1999 , institution=

  10. [18]

    arXiv preprint arXiv:1810.04805 , year=

    Bert: Pre-training of deep bidirectional transformers for language understanding , author=. arXiv preprint arXiv:1810.04805 , year=

  11. [19]

    arXiv preprint arXiv:2205.05092 , year=

    Problems with Cosine as a Measure of Embedding Similarity for High Frequency Words , author=. arXiv preprint arXiv:2205.05092 , year=

  12. [20]

    2013 , publisher=

    Statistical power analysis for the behavioral sciences , author=. 2013 , publisher=

  13. [21]

    Huang, Jiaxin and Gu, Shixiang Shane and Hou, Le and Wu, Yuexin and Wang, Xuezhi and Yu, Hongkun and Han, Jiawei , keywords =

  14. [22]

    , month = sep, year =

    Gehman, Samuel and Gururangan, Suchin and Sap, Maarten and Choi, Yejin and Smith, Noah A. , month = sep, year =

  15. [23]

    Science , author =

    Semantics derived automatically from language corpora contain human-like biases , volume =. Science , author =. 2017 , pages =. doi:10.1126/science.aal4230 , abstract =

  16. [24]

    Documenting

    Dodge, Jesse and Sap, Maarten and Marasović, Ana and Agnew, William and Ilharco, Gabriel and Groeneveld, Dirk and Mitchell, Margaret and Gardner, Matt , month = sep, year =. Documenting

  17. [25]

    Wolfe, Robert and Caliskan, Aylin , month = mar, year =

  18. [26]

    Leveraging

    Mekala, Dheeraj and Vu, Tu and Schick, Timo and Shang, Jingbo , pages =. Leveraging

  19. [27]

    Proceedings of the 59th

    Nadeem, Moin and Bethke, Anna and Reddy, Siva , year =. Proceedings of the 59th. doi:10.18653/v1/2021.acl-long.416 , abstract =

  20. [28]

    Detecting

    Guo, Wei and Caliskan, Aylin , month = jul, year =. Detecting. Proceedings of the 2021. doi:10.1145/3461702.3462536 , abstract =

  21. [29]

    Generating

    Schick, Timo and Schütze, Hinrich , month = oct, year =. Generating

  22. [30]

    Schick, Timo and Udupa, Sahana and Schütze, Hinrich , month = sep, year =. Self-

  23. [31]

    Brown, Tom B. and Mann, Benjamin and Ryder, Nick and Subbiah, Melanie and Kaplan, Jared and Dhariwal, Prafulla and Neelakantan, Arvind and Shyam, Pranav and Sastry, Girish and Askell, Amanda and Agarwal, Sandhini and Herbert-Voss, Ariel and Krueger, Gretchen and Henighan, Tom ...

  24. [32]

    All-but-the-

    Mu, Jiaqi and Bhat, Suma and Viswanath, Pramod , month = mar, year =. All-but-the-

  25. [33]

    and Shtedritski, Aleksandar and Asano, Yuki M

    Kirk, Hannah and Jun, Yennie and Iqbal, Haider and Benussi, Elias and Volpin, Filippo and Dreyer, Frederic A. and Shtedritski, Aleksandar and Asano, Yuki M. , month = oct, year =. Bias

  26. [34]

    and Rudinger, Rachel , year =

    May, Chandler and Wang, Alex and Bordia, Shikha and Bowman, Samuel R. and Rudinger, Rachel , year =. On. Proceedings of the 2019. doi:10.18653/v1/N19-1063 , abstract =

  27. [35]

    Measuring

    Kurita, Keita and Vyas, Nidhi and Pareek, Ayush and Black, Alan W and Tsvetkov, Yulia , year =. Measuring. Proceedings of the. doi:10.18653/v1/W19-3823 , abstract =

  28. [36]

    Contemporary Political Theory , author =

    Intersectionality as. Contemporary Political Theory , author =. 2021 , pages =. doi:10.1057/s41296-021-00490-0 , language =

  29. [37]

    and Zettlemoyer, Luke , month = dec, year =

    Gonen, Hila and Iyer, Srini and Blevins, Terra and Smith, Noah A. and Zettlemoyer, Luke , month = dec, year =. Demystifying

  30. [38]

    Language

    Radford, Alec and Wu, Jeffrey and Child, Rewon and Luan, David and Amodei, Dario and Sutskever, Ilya , year =. Language

  31. [39]

    Psychology of Women Quarterly , author =

    An. Psychology of Women Quarterly , author =. 2013 , pages =. doi:10.1177/0361684312464203 , abstract =

  32. [40]

    and Sanh, Victor and Yong, Zheng-Xin and Webson, Albert and Raffel, Colin and Nayak, Nihal V

    Bach, Stephen H. and Sanh, Victor and Yong, Zheng-Xin and Webson, Albert and Raffel, Colin and Nayak, Nihal V. and Sharma, Abheesht and Kim, Taewoon and Bari, M. Saiful and Fevry, Thibault and Alyafeai, Zaid and Dey, Manan and Santilli, Andrea and Sun, Zhiqing and Ben-David, S...

  33. [41]

    , year =

    Nangia, Nikita and Vania, Clara and Bhalerao, Rasika and Bowman, Samuel R. , year =. Proceedings of the 2020. doi:10.18653/v1/2020.emnlp-main.154 , abstract =

  34. [42]

    Exploring

    Zhuo, Terry Yue and Huang, Yujin and Chen, Chunyang and Xing, Zhenchang , month = feb, year =. Exploring

  35. [43]

    Journal of Machine Learning Research , author =

    Exploring the. Journal of Machine Learning Research , author =. 2020 , pages =

  36. [44]

    Intersectional

    Connor, Paul , year =. Intersectional

  37. [45]

    Toney-Wails, Autumn and Caliskan, Aylin , month = nov, year =

  38. [46]

    ACM Transactions on Information Systems , author =

    A similarity measure for indefinite rankings , volume =. ACM Transactions on Information Systems , author =. 2010 , pages =. doi:10.1145/1852102.1852106 , abstract =

  39. [47]

    Husse, Silke and Spitz, Andreas , month = nov, year =. Mind

  40. [48]

    Cao, Yang Trista and Pruksachatkun, Yada and Chang, Kai-Wei and Gupta, Rahul and Kumar, Varun and Dhamala, Jwala and Galstyan, Aram , month = mar, year =. On the

  41. [49]

    Viswanath, Hrishikesh and Zhang, Tianyi , month = feb, year =

  42. [50]

    Problems

    Faruqui, Manaal and Tsvetkov, Yulia and Rastogi, Pushpendre and Dyer, Chris , month = jun, year =. Problems

  43. [51]

    and Gebru, Timnit and McMillan-Major, Angelina and Shmitchell, Shmargaret , month = mar, year =

    Bender, Emily M. and Gebru, Timnit and McMillan-Major, Angelina and Shmitchell, Shmargaret , month = mar, year =. On the. Proceedings of the 2021. doi:10.1145/3442188.3445922 , abstract =

  44. [52]

    Assessing

    Tan, Yi Chern and Celis, L Elisa , year =. Assessing

  45. [53]

    Intersectional

    Magee, Liam and Ghahremanlou, Lida and Soldatic, Karen and Robertson, Shanthi , month = jul, year =. Intersectional

  46. [54]

    Social and Personality Psychology Compass , author =

    An intersectional approach to understanding how race and social class affect intergroup processes , volume =. Social and Personality Psychology Compass , author =. 2019 , keywords =. doi:10.1111/spc3.12426 , abstract =

  47. [55]

    De-Arteaga, Maria and Romanov, Alexey and Wallach, Hanna and Chayes, Jennifer and Borgs, Christian and Chouldechova, Alexandra and Geyik, Sahin and Kenthapadi, Krishnaram and Kalai, Adam Tauman , month = jan, year =. Bias in. Proceedings of the. doi:10.1145/3287560.3287572 , a...

  48. [56]

    Measuring

    Greenwald, Anthony G and McGhee, Debbie E and Schwartz, Jordan L K , year =. Measuring

  49. [57]

    Antoniak, Maria and Mimno, David , year =. Bad. Proceedings of the 59th. doi:10.18653/v1/2021.acl-long.148 , abstract =

  50. [58]

    Gohar, Usman and Cheng, Lu , keywords =. A

  51. [59]

    Debiasing

    Cheng, Lu and Kim, Nayoung and Liu, Huan , month = aug, year =. Debiasing

  52. [60]

    Intrinsic

    Goldfarb-Tarrant, Seraphina and Marchant, Rebecca and Sanchez, Ricardo Muñoz and Pandya, Mugdha and Lopez, Adam , month = jun, year =. Intrinsic

  53. [61]

    Measuring

    Delobelle, Pieter and Tokpo, Ewoenam Kwaku and Calders, Toon and Berendt, Bettina , month = dec, year =. Measuring

  54. [62]

    Zhao, Jieyu and Wang, Tianlu and Yatskar, Mark and Cotterell, Ryan and Ordonez, Vicente and Chang, Kai-Wei , month = apr, year =. Gender

  55. [63]

    Ghosh, Sourojit and Caliskan, Aylin , year =

  56. [64]

    Problems with

    Zhou, Kaitlyn and Ethayarajh, Kawin and Card, Dallas and Jurafsky, Dan , year =. Problems with. Proceedings of the 60th. doi:10.18653/v1/2022.acl-short.45 , abstract =

  57. [65]

    An, Haozhe and Li, Zongxia and Zhao, Jieyu and Rudinger, Rachel , month = feb, year =

  58. [66]

    , month = aug, year =

    Gupta, Vipul and Venkit, Pranav Narayanan and Laurençon, Hugo and Wilson, Shomir and Passonneau, Rebecca J. , month = aug, year =

  59. [67]

    Goldfarb-Tarrant, Seraphina and Ungless, Eddie and Balkir, Esma and Blodgett, Su Lin , month = may, year =. This

  60. [68]

    Mitigating

    Ahn, Jaimeen and Oh, Alice , file =. Mitigating

  61. [69]

    AI & SOCIETY , author =

    Aligning artificial intelligence with human values: reflections from a phenomenological perspective , volume =. AI & SOCIETY , author =. 2022 , pages =. doi:10.1007/s00146-021-01247-4 , abstract =

  62. [70]

    Language (

    Blodgett, Su Lin and Barocas, Solon and Daumé Iii, Hal and Wallach, Hanna , year =. Language (. Proceedings of the 58th. doi:10.18653/v1/2020.acl-main.485 , abstract =

  63. [71]

    OpenAI , month = mar, year =

  64. [72]

    Touvron, Hugo and Lavril, Thibaut and Izacard, Gautier and Martinet, Xavier and Lachaux, Marie-Anne and Lacroix, Timothée and Rozière, Baptiste and Goyal, Naman and Hambro, Eric and Azhar, Faisal and Rodriguez, Aurelien and Joulin, Armand and Grave, Edouard and Lample, Guillau...

  65. [73]

    Evaluating and

    Tamkin, Alex and Askell, Amanda and Lovitt, Liane and Durmus, Esin and Joseph, Nicholas and Kravec, Shauna and Nguyen, Karina and Kaplan, Jared and Ganguli, Deep , month = dec, year =. Evaluating and

  66. [74]

    Training language models to follow instructions with human feedback , abstract =

    Ouyang, Long and Wu, Jeff and Jiang, Xu and Almeida, Diogo and Wainwright, Carroll L and Mishkin, Pamela and Zhang, Chong and Agarwal, Sandhini and Slama, Katarina and Ray, Alex and Schulman, John and Hilton, Jacob and Kelton, Fraser and Miller, Luke and Simens, Maddie and Ask...

  67. [75]

    Gender bias and stereotypes in

    Kotek, Hadas and Dockum, Rikker and Sun, David , month = nov, year =. Gender bias and stereotypes in. Proceedings of. doi:10.1145/3582269.3615599 , abstract =

  68. [76]

    Article in press,

    Axt, Jordan R , file =. Article in press,

  69. [77]

    Prompting is not a substitute for probability measurements in large language models , url =

    Hu, Jennifer and Levy, Roger , month = oct, year =. Prompting is not a substitute for probability measurements in large language models , url =

  70. [78]

    Behavior Research Methods , author =

    The project implicit international dataset:. Behavior Research Methods , author =. 2022 , pages =. doi:10.3758/s13428-022-01851-2 , abstract =

  71. [79]

    and Ritter, Alan and Xu, Wei , month = mar, year =

    Naous, Tarek and Ryan, Michael J. and Ritter, Alan and Xu, Wei , month = mar, year =. Having

  72. [80]

    Journal of Data and Information Quality , author =

    Biases in. Journal of Data and Information Quality , author =. 2023 , pages =. doi:10.1145/3597307 , abstract =

  73. [81]

    Salinas, Abel and Penafiel, Louis and McCormack, Robert and Morstatter, Fred , month = oct, year =. "

  74. [82]

    Evaluating

    Omrani Sabbaghi, Shiva and Wolfe, Robert and Caliskan, Aylin , month = aug, year =. Evaluating. Proceedings of the 2023. doi:10.1145/3600211.3604666 , abstract =

  75. [83]

    , month = nov, year =

    Wan, Yuxuan and Wang, Wenxuan and He, Pinjia and Gu, Jiazhen and Bai, Haonan and Lyu, Michael R. , month = nov, year =. Proceedings of the 31st. doi:10.1145/3611643.3616310 , abstract =

  76. [84]

    PNAS Nexus , author =

    Extracting intersectional stereotypes from embeddings:. PNAS Nexus , author =. 2024 , pages =. doi:10.1093/pnasnexus/pgae089 , abstract =

  77. [85]

    , volume =

    Valence biases and emergence in the stereotype content of intersecting social categories. , volume =. Journal of Experimental Psychology: General , author =. 2023 , pages =. doi:10.1037/xge0001416 , abstract =

  78. [86]

    Gupta, Shashank and Shrivastava, Vaishnavi and Deshpande, Ameet and Kalyan, Ashwin and Clark, Peter and Sabharwal, Ashish and Khot, Tushar , file =

  79. [87]

    Lucy, Li and Blodgett, Su Lin and Shokouhi, Milad and Wallach, Hanna and Olteanu, Alexandra , month = oct, year =. "

  80. [88]

    Current Issues in Criminal Justice , author =

    Predicting risk in criminal procedure: actuarial tools, algorithms,. Current Issues in Criminal Justice , author =. 2020 , pages =. doi:10.1080/10345329.2019.1658694 , abstract =

  81. [89]

    AI and Ethics , author =

    Is. AI and Ethics , author =. 2023 , keywords =. doi:10.1007/s43681-022-00166-4 , abstract =

  82. [90]

    Technology Analysis & Strategic Management , author =

    Artificial intelligence (. Technology Analysis & Strategic Management , author =. 2023 , pages =. doi:10.1080/09537325.2021.1971188 , abstract =

  83. [91]

    Tech Policy Press , author =

    Using. Tech Policy Press , author =. 2024 , file =

  84. [92]

    Northern European Journal of Language Technology , author =

    Bias. Northern European Journal of Language Technology , author =

  85. [93]

    Behavior Research Methods, Instruments, & Computers , author =

    Words high and low in pleasantness as rated by male and female college students , volume =. Behavior Research Methods, Instruments, & Computers , author =. 1986 , pages =. doi:10.3758/BF03204403 , language =

  86. [94]

    Jiang, Albert Q. and Sablayrolles, Alexandre and Mensch, Arthur and Bamford, Chris and Chaplot, Devendra Singh and Casas, Diego de las and Bressand, Florian and Lengyel, Gianna and Lample, Guillaume and Saulnier, Lucile and Lavaud, Lélio Renard and Lachaux, Marie-Anne and Stoc...

  87. [95]

    Jiang, Albert Q. and Sablayrolles, Alexandre and Roux, Antoine and Mensch, Arthur and Savary, Blanche and Bamford, Chris and Chaplot, Devendra Singh and Casas, Diego de las and Hanna, Emma Bou and Bressand, Florian and Lengyel, Gianna and Bour, Guillaume and Lample, Guillaume ...

  88. [96]

    Bianchi, Federico and Kalluri, Pratyusha and Durmus, Esin and Ladhak, Faisal and Cheng, Myra and Nozza, Debora and Hashimoto, Tatsunori and Jurafsky, Dan and Zou, James and Caliskan, Aylin , month = jun, year =. Easily. 2023. doi:10.1145/3593013.3594095 , language =

  89. [97]

    Recursive

    Socher, Richard and Perelygin, Alex and Wu, Jean and Chuang, Jason and Manning, Christopher D and Ng, Andrew and Potts, Christopher , year =. Recursive

  90. [98]

    Rudinger, Rachel and Naradowsky, Jason and Leonard, Brian and Van Durme, Benjamin , month = apr, year =. Gender

  91. [99]

    2024 , note =

    ". 2024 , note =. doi:10.13140/RG.2.2.26397.12006 , language =

  92. [100]

    , month = feb, year =

    Bai, Xuechunzi and Wang, Angelina and Sucholutsky, Ilia and Griffiths, Thomas L. , month = feb, year =. Measuring

  93. [101]

    The Guardian , author =

    New. The Guardian , author =. 2019 , keywords =

  94. [102]

    and Arora, Simran and Mazeika, Mantas and Hendrycks, Dan and Lin, Zinan and Cheng, Yu and Koyejo, Sanmi and Song, Dawn and Li, Bo , month = feb, year =

    Wang, Boxin and Chen, Weixin and Pei, Hengzhi and Xie, Chulin and Kang, Mintong and Zhang, Chenhui and Xu, Chejian and Xiong, Zidi and Dutta, Ritik and Schaeffer, Rylan and Truong, Sang T. and Arora, Simran and Mazeika, Mantas and Hendrycks, Dan and Lin, Zinan and Cheng, Yu an...

  95. [103]

    Automaticity of

    Bargh, John A and Chen, Mark and Burrows, Lara , year =. Automaticity of

  96. [104]

    Psychological Review , author =

    Implicit social cognition:. Psychological Review , author =. 1995 , pages =. doi:10.1037/0033-295X.102.1.4 , language =

  97. [105]

    Bias\_Prop\_\_\_AIES\_2024\_Submission.pdf , file =

  98. [106]

    Proceedings of the 2021

    Dhamala, Jwala and Sun, Tony and Kumar, Varun and Krishna, Satyapriya and Pruksachatkun, Yada and Chang, Kai-Wei and Gupta, Rahul , month = mar, year =. Proceedings of the 2021. doi:10.1145/3442188.3445924 , abstract =

  99. [107]

    Weak-to-

    Burns, Collin and Izmailov, Pavel and Kirchner, Jan Hendrik and Baker, Bowen and Gao, Leo and Aschenbrenner, Leopold and Chen, Yining and Ecoffet, Adrien and Joglekar, Manas and Leike, Jan and Sutskever, Ilya and Wu, Jeff , month = dec, year =. Weak-to-

  100. [108]

    Journal of Personality and Social Psychology , author =

    A spontaneous stereotype content model:. Journal of Personality and Social Psychology , author =. 2022 , pages =. doi:10.1037/pspa0000312 , abstract =

  101. [109]

    Journal of Social Issues , author =

    Warmth‐. Journal of Social Issues , author =. 2017 , pages =. doi:10.1111/josi.12210 , abstract =

  102. [110]

    Journal of Personality and Social Psychology , author =

    A model of (often mixed) stereotype content:. Journal of Personality and Social Psychology , author =. 2002 , pages =. doi:10.1037/0022-3514.82.6.878 , language =

  103. [111]

    Jeoung, Sullam and Ge, Yubin and Diesner, Jana , month = oct, year =

  104. [112]

    Charlesworth, Tessa E. S. and Banaji, Mahzarin R. , editor =. Evidence of. Handbook on. 2022 , doi =

  105. [113]

    Psychological Inquiry , author =

    Beyond. Psychological Inquiry , author =. 2017 , pages =. doi:10.1080/1047840X.2017.1373546 , language =

  106. [114]

    Context and the

    Brownstein, Michael , editor =. Context and the. Implicit. 2016 , doi =. doi:10.1093/acprof:oso/9780198766179.003.0010 , pages =

  107. [115]

    CNS Spectrums , author =

    Risk factors for recidivism in individuals receiving community sentences: a systematic review and meta-analysis , volume =. CNS Spectrums , author =. 2020 , pages =. doi:10.1017/S1092852919001056 , abstract =

  108. [116]

    Unpacking

    Bravo-Marquez, Felipe and Zambrano, Maria Jose , file =. Unpacking

  109. [117]

    Fort, Karën and Alemany, Laura Alonso and Benotti, Luciana and Bezançon, Julien and Borg, Claudia and Borg, Marthese and Chen, Yongjian and Ducel, Fanny and Dupont, Yoann and Ivetta, Guido and Li, Zhijian and Mieskes, Margot and Naguib, Marco and Qian, Yuyan and Radaelli, Matt...

  110. [118]

    Zhao, Yachao and Wang, Bo and Wang, Yan and Zhao, Dongming and Jin, Xiaojia and Zhang, Jijun and He, Ruifang and Hou, Yuexian , file =. A

  111. [119]

    Scarone, Bruno and Viola, Alfredo and Baeza-Yates, Ricardo , month = may, year =. A

  112. [120]

    Computational Linguistics , author =

    Bias and. Computational Linguistics , author =. 2024 , pages =. doi:10.1162/coli_a_00524 , abstract =

  113. [121]

    Cabello, Laura and Jørgensen, Anna Katrine and Søgaard, Anders , month = jun, year =. On the. 2023. doi:10.1145/3593013.3594004 , abstract =

  114. [122]

    and Kirchhoff, Katrin , year =

    Salazar, Julian and Liang, Davis and Nguyen, Toan Q. and Kirchhoff, Katrin , year =. Masked. Proceedings of the 58th. doi:10.18653/v1/2020.acl-main.240 , abstract =

  115. [123]

    American Anthropologist , author =

    Semantic. American Anthropologist , author =. 1964 , pages =. doi:10.1525/aa.1964.66.3.02a00880 , language =

  116. [124]

    Feng, Shangbin and Park, Chan Young and Liu, Yuhan and Tsvetkov, Yulia , file =. From

  117. [125]

    Grounding and

    Kenthapadi, Krishnaram and Sameki, Mehrnoosh and Taly, Ankur , month = aug, year =. Grounding and. Proceedings of the 30th. doi:10.1145/3637528.3671467 , abstract =

  118. [126]

    Explicit and

    Giorgi, Salvatore and Liu, Tingting and Aich, Ankit and Isman, Kelsey and Sherman, Garrick and Fried, Zachary and Sedoc, João and Ungar, Lyle H and Curtis, Brenda , file =. Explicit and

  119. [127]

    Salewski, Leonard and Alaniz, Stephan and Rio-Torto, Isabel and Schulz, Eric and Akata, Zeynep , year =. In-

  120. [128]

    Cao, Yang Trista and Sotnikova, Anna and Daumé III, Hal and Rudinger, Rachel and Zou, Linda , month = jun, year =. Theory-

  121. [129]

    Upstream

    Steed, Ryan and Panda, Swetasudha and Kobren, Ari and Wick, Michael , year =. Upstream. Proceedings of the 60th. doi:10.18653/v1/2022.acl-long.247 , language =

  122. [130]

    Crosby, Faye and Bromley, Stephanie and Saxe, Leonard , year =. Recent

  123. [131]

    Language

    Kumar, Sachin and Balachandran, Vidhisha and Njoo, Lucille and Anastasopoulos, Antonios and Tsvetkov, Yulia , year =. Language. Proceedings of the 17th. doi:10.18653/v1/2023.eacl-main.241 , abstract =

  124. [132]

    Measuring

    Delobelle, Pieter and Tokpo, Ewoenam and Calders, Toon and Berendt, Bettina , year =. Measuring. Proceedings of the 2022. doi:10.18653/v1/2022.naacl-main.122 , abstract =

  125. [133]

    Measuring bias in

    Onorati, Dario and Ruzzetti, Elena and Venditti, Davide and Ranaldi, Leonardo and Zanzotto, Fabio , year =. Measuring bias in. Findings of the. doi:10.18653/v1/2023.findings-emnlp.539 , abstract =

  126. [134]

    Collapse of

    Herel, David and Mikolov, Tomas , month = apr, year =. Collapse of

  127. [135]

    Mei, Katelyn and Fereidooni, Sonia and Caliskan, Aylin , month = jun, year =. Bias. Proceedings of the 2023. doi:10.1145/3593013.3594109 , abstract =

  128. [136]

    and Nejadgholi, Isar and Kiritchenko, Svetlana , month = jun, year =

    Fraser, Kathleen C. and Nejadgholi, Isar and Kiritchenko, Svetlana , month = jun, year =. Understanding and

  129. [137]

    2024 , note =

    Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society , author =. 2024 , note =. doi:10.1609/aies.v7i1.31751 , abstract =

  130. [138]

    2024 , note =

    Nature , author =. 2024 , note =. doi:10.1038/s41586-024-07856-5 , abstract =

  131. [139]

    Nature Computational Science , author =

    Generative language models exhibit social identity biases , volume =. Nature Computational Science , author =. 2024 , pages =. doi:10.1038/s43588-024-00741-1 , abstract =

  132. [140]

    , volume =

    Racial stereotypes of one hundred college students. , volume =. The Journal of Abnormal and Social Psychology , author =. 1933 , note =. doi:10.1037/h0074049 , abstract =

  133. [141]

    Mendelsohn, Julia and Budak, Ceren , month = feb, year =. When. doi:10.48550/arXiv.2502.13246 , abstract =

  134. [142]

    Xu, Xiaohan and Li, Ming and Tao, Chongyang and Shen, Tao and Cheng, Reynold and Li, Jinyang and Xu, Can and Tao, Dacheng and Zhou, Tianyi , month = oct, year =. A. doi:10.48550/arXiv.2402.13116 , abstract =

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.