Pith. sign in

REVIEW 4 major objections 6 minor 83 references

Is Your Model Fairly Certain? Uncertainty-Aware Fairness Evaluation for LLMs

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Accuracy-based fairness metrics miss how confident an LLM is in its biased answers; a new metric, UCerF, incorporates that confidence and changes which models are judged fair.

desk verdict Useful new metric and dataset for uncertainty-aware fairness auditing, but the central fairness claim rests on an unvalidated confidence scale and no comparison to the one prior method; still worth refereeing. read the letter →

arxiv 2505.23996 v1 pith:SA2M4VFB submitted 2025-05-29 cs.CL cs.AIcs.LG

Yinong Oliver Wang , Nivedha Sivakumar , Falaah Arif Khan , Rin Metcalf Susa , Adam Golinski , Natalie Mackraz , Barry-John Theobald , Luca Zappella
show 1 more author
Nicholas Apostoloff
This is my paper
classification cs.CLcs.AIcs.LG
keywords fairnessevaluationuncertaintyquantificationlargelanguagemodelsgenderbiascoreferenceresolutionbenchmarkdatasetequalizedodds
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

Standard accuracy-based fairness metrics for large language models treat every prediction as simply right or wrong, and this paper argues that the confidence behind each prediction is itself a fairness signal that those metrics throw away. It proposes UCerF, a metric that scores each minimal pair of samples by how close the model's 'desirability' is between groups — where a correct answer contributes its confidence and an incorrect answer contributes negative confidence — and averages this across the dataset, yielding 1 for perfectly fair and 0 for completely biased. To evaluate the metric on modern models, the paper introduces SynthBias, a 31,756-sample, human-validated pronoun–occupation coreference dataset that is roughly ten times larger and more diverse than WinoBias. On a ten-model benchmark, the headline observation is that Mistral-7B-Instruct looks acceptable under Equalized Odds yet is flagged by UCerF for high confidence in incorrect predictions, while the far less accurate but more cautious Pythia-1B ranks higher in fairness; if right, this means current fairness audits can miss a systematic harmful behavior.

What carries the argument

The load-bearing object is the linear scale of behavior preference (LSBP) with its desirability function $D(x_i)$: a rescaled perplexity gives a normalized certainty $c(x_i)=(k-\text{perplexity}(x_i))/(k-1)$, and predictions are placed on a continuum where correct answers sit at $+c$, incorrect answers at $-c$, and ambiguous asks without a correct answer at $+c$. UCerF is then the expected one-minus-half-distance between the desirability values of the two groups on this scale, which is what lets confidence asymmetries — a model being certain about one group and shaky about the other — show up as fairness deficits even when both predictions are right. The same scale also yields the group-wise variant $U_{\text{group}}=|\text{TPD}_{\text{pro}}-\text{TPD}_{\text{anti}}|+|\text{FPD}_{\text{pro}}-\text{FPD}_{\text{anti}}|$, an analog of Equalized Odds over desirabilities rather than error rates.

What would settle it

Run the ten models on the same SynthBias and WinoBias items, then replace perplexity-based certainty with a behavioral confidence proxy — for instance, the agreement rate across repeated temperature-1.0 samplings or the model's own self-reported probability — and recompute the UCerF rankings; if the ordering of Mistral-7B versus Pythia-1B flips, the headline result is an artifact of the certainty estimator rather than a property of the models' bias.

Watch

Extended reading notes

Core claim

UCerF is defined on a linear scale of behavior preference: each sample's certainty is computed from perplexity as $c(x_i)=(k-\text{perplexity}(x_i))/(k-1)\in[0,1]$, then mapped to a desirability $D(x_i)$ equal to $c(x_i)$ when the prediction is correct (or when no answer is objectively correct) and $-c(x_i)$ when it is incorrect. The metric is $U(X)=\mathbb{E}_{x_i}[1-\tfrac{1}{2}|D(x_i^A)-D(x_i^B)|]$, the expected closeness of desirabilities between pro- and anti-stereotypical versions of the same sentence, and the paper's central claim is that this quantity 'provides a better understanding of a model's fairness by accounting for uncertainty differences between groups.' The supporting evidence is a case study where Falcon-40B is correct for both groups yet nearly random in the anti-stereotypical case — a 0.552 UCerF despite a zero TPR gap — and a contrasting example where a confidently-wrong anti-stereotypical answer yields near-0.8 UCerF while Equalized Odds reports a 0.5 gap. Across the ten-model benchmark, UCerF penalizes confident-in-bias models such as Mistral-7B-Instruct and rewards cautious ones such as Pythia-1B, implying that accuracy-based fairness rankings can misorder models.

Load-bearing premise

The entire ranking depends on one measurement premise: that perplexity, rescaled to a 0-to-1 certainty score, faithfully reflects how confident the model really is about its answer, and that a confidently wrong answer is exactly as bad as the linear scale assumes.

Editorial extensions

If this is right

  • Fairness audits that adopt UCerF will rank models differently: confident-in-error models are penalized and cautious low-accuracy models are no longer automatically judged unfair, so the current best-fairness-model orderings change.
  • Equalized Odds can mislead in both directions on the same data — it reports a zero disparity where UCerF scores 0.55, and a 0.5 error-rate gap where UCerF scores 0.80 — so relying on right/wrong alone can both hide bias and exaggerate it.
  • The group-wise reformulation makes UCerF usable without minimal pairs, extending uncertainty-aware fairness evaluation to datasets such as BBQ Lite, where the authors measure fairness across race, religion, disability, age, and other attributes.
  • Dataset difficulty itself becomes a fairness-relevant design choice: SynthBias's harder, human-validated type-1 sentences expose hidden bias in models that look fair on WinoBias, such as Llama-3-70B-Instruct dropping from third to eighth in the type-1 ranking.

Reading between the lines

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

  • A natural stress test is to recompute UCerF with sampling-based or self-reported confidence estimators; if the Mistral-versus-Pythia ordering flips, the headline result should be attributed to the perplexity assumption rather than to model behavior.
  • The symmetric desirability scale bakes in a normative choice — that confident-wrong is the exact opposite of confident-right — and a deployment context that punishes confident errors more severely would reorder models, so the framework could be re-parameterized with an asymmetry weight.
  • Because the metric rewards cautious uncertainty, a maximally uninformative model can score as 'fair'; the paper's own joint fairness-performance product is therefore the practically relevant selection criterion in high-stakes settings, where fair-but-uninformative is not acceptable.
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 / 6 minor

Summary. The paper proposes UCerF, an uncertainty-aware fairness metric that combines prediction correctness with a normalized perplexity-based certainty score into a desirability value D(x) in [-1,1] (Eqs. 1-2), and defines fairness as the expected absolute desirability gap between pro- and anti-stereotypical minimal pairs (Eqs. 3-4), with a group-wise variant (Eq. 5). It introduces SynthBias, a 31,756-sample GPT-4o-generated and human-validated gender-occupation coreference dataset, and benchmarks ten open-weight LLMs on both WinoBias and SynthBias. The headline claim is that UCerF captures uncertainty-induced unfairness missed by Equalized Odds, with Mistral-7B-Instruct ranked substantially worse under UCerF than under EO because of high confidence in incorrect predictions, and that SynthBias is more diverse and challenging than WinoBias.

Significance. If the central claims hold, the paper makes a useful contribution: it introduces a public, large-scale fairness dataset with human validation, a modular metric that can be combined with different uncertainty estimators, a ten-model benchmark, and robustness probes using Rényi divergence and Fisher-Rao distance. The authors also provide a concrete dataset link and make falsifiable ranking claims. However, the central evaluative claim is not yet empirically established: the certainty scale in Eq. (1) is not calibrated against behavioral or sampling-based evidence, and the claimed superiority over EO rests on selected examples and on a definitional construction rather than on validation against an independent ground truth or an uncertainty-aware baseline. The significance is therefore conditional on the additional validation requested below.

major comments (4)
  1. [Section 3.3, Eq. (1)] The normalized certainty transform c(x) = (k - 2^H)/(k - 1) is strongly nonlinear: for a binary choice, p=0.75 maps to c≈0.245 whereas p=0.6 maps to c≈0.04, compressing moderate-confidence predictions near zero. All UCerF rankings, and the headline claim that Mistral-7B is unfair because of high confidence in incorrect predictions, depend on this transform being a faithful, approximately interval-scale measure of model confidence. The paper does not calibrate c against behavioral evidence, even though Sec. 3.1 motivates UCerF by arguing that repeated-sampling expected scores would reveal fairness differences; no comparison to repeated stochastic sampling, temperature-based consistency, or calibration curves is reported. Table 9 shows Mistral ranks last under Rényi divergence and Fisher-Rao distance, which supports some robustness across estimators, but it does not validate any of these scales as the quantity the fairness argument requires. Please add a calibration analysis and demonstrate that the qualitative rankings, especially the Mistral-versus-Pythia comparison, are stable under monotone transformations of c or under an alternative estimator with known calibration properties.
  2. [Section 5.2, Eq. (5)] The central claim that UCerF 'provides a better understanding of a model's fairness' is not empirically validated. The case study in Fig. 5 uses hand-picked examples, and the TPD/FPD quantities in Eq. (5) are averages of the same D used in UCerF, so observations such as TPR exceeding TPD in Fig. 5(a) follow by construction from the definition of D rather than from external evidence about fairness. No uncertainty-aware fairness baseline is compared (Kuzucu et al., 2023, is cited but not evaluated), no human fairness judgments or probe models with known confidence asymmetries are used as ground truth, and the paper's own repeated-sampling argument from Sec. 3.1 is never operationalized. I recommend adding a validation experiment, for example comparing UCerF and EO against human fairness rankings or against synthetic models with controlled confidence differences, before claiming that UCerF is more reflective of internal model bias.
  3. [Section 3.3 and Section 5.1 (type1 tasks)] For type1 samples, D(x)=c regardless of the model's prediction because there is no correct answer. This makes UCerF a pure certainty-parity measure and equates 'equally confident' with 'fair', even when a model is equally and highly confident in a stereotype-consistent resolution of an ambiguous pronoun. This normative choice drives the type1 rankings, for example Mistral-7B ranking first under UCerF in Table 4 type1 while also being the most certain model. The paper should either justify this normative claim explicitly or restrict the type1 conclusions to uncertainty parity rather than asserting that UCerF uncovers hidden bias in ambiguous settings.
  4. [Section 4.3 and Table 1] The SynthBias quality claim rests on crowd annotations with several threshold parameters: the annotator entrance-test cutoff of ≥80%, the coherence vote of >75%, and the 75% consensus over at least four annotators (or up to ten). The paper reports neither inter-annotator agreement nor sensitivity of the final dataset to these thresholds. Since the dataset is a main contribution and these thresholds are free parameters, please report agreement measures such as Fleiss' kappa and show that the filtering decisions are not knife-edge. This would also strengthen the comparison with WinoBias in Table 1, which currently relies primarily on embedding-derived diversity statistics.
minor comments (6)
  1. [Title and Abstract] The title 'Is Y our Model Fairly Certain?' appears to contain a typo; it should likely read 'Is Your Model Fairly Certain?'.
  2. [Section 1] The sentence 'In summary, we contribute the follows' needs correction, and the phrase 'AI systems are socially-beneficial' is missing a relative pronoun or should be rewritten.
  3. [Section 3.3, Eqs. (3)-(4)] The notation in Eq. (3) defines U(x_i) for a pair of samples but Eq. (4) averages over X; please clarify how the expectation is taken over minimal pairs and whether the group-wise formulation in Eq. (5) is intended for datasets without such pairs.
  4. [Figure 6] Figure 6 conveys rankings only through color intensity, which is difficult to read for individual models; please provide a numeric table alongside the figure or annotate the rank values.
  5. [Section J, Table 9] Table 9 reports UCerF scores for only eight of the ten benchmark models; please report all ten models or explain the omission, since the robustness claim would be stronger with complete coverage.
  6. [Tables 2, 4-7] The experiments are run on five random seeds, but no variance or confidence intervals are reported in the tables; given that the paper makes fine-grained ranking claims, standard deviations or CIs should be included.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: UCerF is a proposed definition and the empirical benchmark is self-contained.

full rationale

UCerF is constructed explicitly in Eqs. (1)-(4): c is normalized perplexity, D maps correctness and c to [-1,1], and U averages 1 - 0.5|D_pro - D_anti|. These are definitions, not fitted parameters, and the paper does not calibrate U to any ground-truth fairness label or to the models under test. The empirical evaluations (SynthBias construction and human annotation, ten-LLM benchmark, Table 9 alternatives) are self-contained measurements using the defined metric. The statement that confidence disparity should count toward fairness is a transparent normative choice introduced in Sec. 3.2 ('fairness can be intuitively defined as simply the distance between two groups on LSBP'), not a derived result that secretly presupposes its conclusion. Same-author citations (Santilli et al. in Sec. 2.2; Mackraz et al. and Patel et al. in Sec. 2.1) are related-work or plural supporting citations; the uncertainty-estimator choice is additionally supported by external Vashurin et al. and by direct re-evaluation with Rényi divergence and Fisher-Rao distance in Sec. J. The unvalidated perplexity-to-confidence mapping is a validation or calibration concern, not circularity. No fitted parameter is renamed as a prediction and no uniqueness claim is imported from same-author work. Score 0 reflects the absence of a circular step rather than endorsement of every normative or calibration choice in the paper.

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

The metric rests on a normative desirability scale and a perplexity-to-certainty mapping; the dataset rests on BLS stereotype labels, human annotation thresholds, and a minimal-pair assumption. The UCerF formulas themselves contain no fitted parameters.

free parameters (4)
  • annotator consensus threshold = 75%
    Set in Sec. 4.3; samples are kept only when at least 75% of annotators agree on coherence and ambiguity, shaping which sentences enter SynthBias.
  • annotator entrance-test cutoff = 80%
    Set in Sec. 4.3; only annotators scoring at least 80% on the 20-question test are admitted, influencing label quality and the final dataset.
  • occupation-pair stereotype difference cutoff = 10%
    Set in Sec. 4.2; occupation pairs whose BLS male/female distributions are within 10% are excluded, changing the pool of stereotyping probes.
  • dynamic annotation cap = 10 annotations per sample
    Set in Sec. 4.3; collection stops at 75% consensus among at least four raters or ten total responses, affecting label noise in SynthBias.
assumptions (6)
  • standard math Standard properties of expectation and absolute values used in Eq. (3)-(4) hold.
    The UCerF formulas are simple averages of 1 - 0.5*|D(pro)-D(anti)|; no nonstandard mathematics is needed.
  • domain assumption Perplexity over next-token occupation probabilities is a valid uncertainty estimator and maps linearly to certainty via c=(k-ppl)/(k-1).
    Invoked in Sec. 3.3 Eq. (1). The paper cites prior work that logit-based estimators are competitive but does not calibrate this mapping for fairness, and the Impact Statement acknowledges alternative estimators could change results.
  • domain assumption BLS workforce gender percentages are the correct ground truth for pro- and anti-stereotypical labels.
    Used in Sec. 4.2 and Table 3, inherited from WinoBias; this equates statistical workforce composition with stereotype and treats deviations from these statistics as model bias.
  • domain assumption Human annotators' majority judgments of coherence and ambiguity correctly classify type1/type2 and thus define the fairness task.
    Sec. 4.3; the dataset cleaning depends on 75% consensus thresholds, and no reliability analysis such as inter-annotator agreement is reported.
  • domain assumption Sentence pairs formed by swapping the pronoun are minimal pairs of equal difficulty.
    Sec. 3.3 and Sec. 4.2; the metric attributes D differences to group stereotypes rather than lexical or syntactic effects of the pronoun switch. Sec. G tests nearby sentences but does not establish minimal-pair equivalence.
  • ad hoc to paper The desirability mapping D(x)=c if correct, -c if incorrect, and c if no correct answer is the right normative scale for fairness.
    Sec. 3.2 Eq. (2); this is the value judgment that defines what UCerF counts as fair, and it is not derived from a fairness principle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Is Your Model Fairly Certain? Uncertainty-Aware Fairness Evaluation for LLMs." pith.science (2026). https://pith.science/paper/SA2M4VFB

@misc{pith2026250523996,
  author       = {Pith},
  title        = {Pith review of: Is Your Model Fairly Certain? Uncertainty-Aware Fairness Evaluation for LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SA2M4VFB}},
  note         = {Machine review of arXiv:2505.23996}
}
read the original abstract

The recent rapid adoption of large language models (LLMs) highlights the critical need for benchmarking their fairness. Conventional fairness metrics, which focus on discrete accuracy-based evaluations (i.e., prediction correctness), fail to capture the implicit impact of model uncertainty (e.g., higher model confidence about one group over another despite similar accuracy). To address this limitation, we propose an uncertainty-aware fairness metric, UCerF, to enable a fine-grained evaluation of model fairness that is more reflective of the internal bias in model decisions compared to conventional fairness measures. Furthermore, observing data size, diversity, and clarity issues in current datasets, we introduce a new gender-occupation fairness evaluation dataset with 31,756 samples for co-reference resolution, offering a more diverse and suitable dataset for evaluating modern LLMs. We establish a benchmark, using our metric and dataset, and apply it to evaluate the behavior of ten open-source LLMs. For example, Mistral-7B exhibits suboptimal fairness due to high confidence in incorrect predictions, a detail overlooked by Equalized Odds but captured by UCerF. Overall, our proposed LLM benchmark, which evaluates fairness with uncertainty awareness, paves the way for developing more transparent and accountable AI systems.

Figures

Figures reproduced from arXiv: 2505.23996 by the authors.

Figure 1
Figure 1. Overview of uncertainty impact in fairness evalua￾tion. Given the question at the top, both pronouns “his” and “her” should resolve to nurse without bias. Models A&B exhibit bias as the prediction is flipped by the pronoun. Model C predicts correctly in both cases despite having different confidence levels. Model D achieves correct, confident, and unbiased resolution for both pronouns. While correctness can reveal b… view at source ↗
Figure 2
Figure 2. Linear scale of behavior preference. From left to right, model behavior changes from confidently incorrect to confidently correct with unconfident behavior in between. and correctness, neglecting complex scenarios with varying correctness and uncertainty levels, which is essential for comprehensive fairness evaluation as shown in Sec. 3. 3 UCerF - Uncertainty-Aware Fairness 3.1 Overlooked Information - Uncertainty C… view at source ↗
Figure 3
Figure 3. Case study on LSBP. The four scenarios of the gender resolution task are shown. With the red pin and blue pin marking the model behaviors, the four cases illustrate the relation between model fairness and group distances on LSBP. referenced occupation aligns with the stereotypical pronoun and “anti-stereotypical” otherwise, we explore four sce￾narios in [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Visual comparison of text diversity. The t-SNE and PCA visualization of sentence embeddings in WinoBias and Syn￾thBias show better coverage in SynthBias on sentence context variations, indicating higher diversity. See Sec. E for details. for both Q2 and Q3, and (3) typ…
Figure 5
Figure 5. Figure 5: Case study of “nurse” and “physician”. To demonstrate the example from [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Model evaluation across metrics and datasets. We evaluate ten LLMs across four metrics on two task types in two datasets as configured in Sec. 5.1. We color code performance on WinoBias and SynthBias by their ranks in red and blue respectively. The lighter the color, t…
Figure 7
Figure 7. Figure 7: Per-occupation comparison between UCerF and EO. We break fairness down to examine per-occupation fairness of Falcon-40B on SynthBias. We corroborate observation in Sec. 5.2 that EO can exaggerate fairness scores compared to UCerF. (a) Pro-stereotypical D(x) (b) Anti-st…
Figure 8
Figure 8. Figure 8: Histograms of metrics between datasets. We compare the histograms of D(xi) and U(xi) between WinoBias and Synth￾Bias with Falcon-40B. The higher volume in undesirable cases in SynthBias show greater coverage of fairness scenarios. well under UCerF (second). Mistral-7B-…
Figure 9
Figure 9. Figure 9: 2D rankings of models on Type2 WinoBias [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]
Figure 11
Figure 11. Figure 11: Distribution of rater locales. We show the histogram of rater locales in each survey response. To further filter the samples by human rater consistency, we analyze the answers to each survey question. In the first question Q1 regarding sentence coherence, we find a to…
Figure 12
Figure 12. Figure 12: Distribution of answer consistency to Q2 and Q3 in the survey. We split the histogram by type1 and type2 tasks as well as “target” and “other” occupations corresponding to the answers to Q2 and Q3. The x-axis denotes the ratio of winning votes, i.e. the majority vote …
Figure 13
Figure 13. Figure 13: Case study of similar sentences. We select three pairs of samples where the sentence vocabularies differ by at most two words. The bold text highlights the sentence differences. The table shows that slight variations in the inputs do impact model prediction and the su…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

83 extracted references · 54 canonical work pages

  1. [1]

    The falcon series of open language models

    Almazrouei, E., Alobeidli, H., Alshamsi, A., Cappelli, A., Cojocaru, R., Debbah, M., Goffinet, \'E ., Hesslow, D., Launay, J., Malartic, Q., et al. The falcon series of open language models. arXiv preprint arXiv:2311.16867, 2023

    arXiv 2023

  2. [2]

    The silicon ceiling: Auditing gpt’s race and gender biases in hiring

    Armstrong, L., Liu, A., MacNeil, S., and Metaxa, D. The silicon ceiling: Auditing gpt’s race and gender biases in hiring. In Proceedings of the 4th ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization, 2024

    2024

  3. [3]

    V., and Pan, R

    Ba, Y., Mancenido, M. V., and Pan, R. Fill in the gaps: Model calibration and generalization with synthetic data. In EMNLP, 2024

    2024

  4. [4]

    G., Bradley, H., O'Brien, K., Hallahan, E., Khan, M

    Biderman, S., Schoelkopf, H., Anthony, Q. G., Bradley, H., O'Brien, K., Hallahan, E., Khan, M. A., Purohit, S., Prashanth, U. S., Raff, E., et al. Pythia: A suite for analyzing large language models across training and scaling. In ICML, 2023

    2023

  5. [5]

    Beyond the imitation game: Quantifying and extrapolating the capabilities of language models

    BIG-bench authors . Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. TMLR, 2023

    work page 2023

  6. [6]

    L., Barocas, S., Daum \'e III, H., and Wallach, H

    Blodgett, S. L., Barocas, S., Daum \'e III, H., and Wallach, H. Language (technology) is power: A critical survey of `` bias '' in NLP . In Jurafsky, D., Chai, J., Schluter, N., and Tetreault, J. (eds.), ACL, 2020

    work page 2020

  7. [7]

    A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M

    Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M. S., Bohg, J., Bosselut, A., Brunskill, E., et al. On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021

    arXiv 2021

  8. [8]

    Bridle, J. S. Probabilistic interpretation of feedforward classification network outputs, with relationships to statistical pattern recognition. In Neurocomputing, 1990

    work page 1990

Show all 83 references
  1. [9]

    Fairness in large language models: A taxonomic survey

    Chu, Z., Wang, Z., and Zhang, W. Fairness in large language models: A taxonomic survey. ACM SIGKDD explorations newsletter, 2024

    2024

  2. [10]

    Rainproof: An umbrella to shield text generators from out-of-distribution data

    Darrin, M., Piantanida, P., and Colombo, P. Rainproof: An umbrella to shield text generators from out-of-distribution data. In EMNLP, 2023

    2023

  3. [11]

    Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model, 2024

    DeepSeek-AI. Deepseek-v2: A strong, economical, and efficient mixture-of-experts language model, 2024

    2024

  4. [12]

    Bold: Dataset and metrics for measuring biases in open-ended language generation

    Dhamala, J., Sun, T., Kumar, V., Krishna, S., Pruksachatkun, Y., Chang, K.-W., and Gupta, R. Bold: Dataset and metrics for measuring biases in open-ended language generation. In ACM FAccT, 2021

    2021

  5. [13]

    Shifting attention to relevance: Towards the uncertainty estimation of large language models

    Duan, J., Cheng, H., Wang, S., Wang, C., Zavalny, A., Xu, R., Kailkhura, B., and Xu, K. Shifting attention to relevance: Towards the uncertainty estimation of large language models. In ACL, 2024

    2024

  6. [14]

    The llama 3 herd of models

    Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Yang, A., Fan, A., et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  7. [15]

    Is your classifier actually biased? measuring fairness under uncertainty with bernstein bounds

    Ethayarajh, K. Is your classifier actually biased? measuring fairness under uncertainty with bernstein bounds. In ACL, 2020

    2020

  8. [16]

    Fabris, A., Messina, S., Silvello, G., and Susto, G. A. Algorithmic fairness datasets: the story so far. In Data Mining and Knowledge Discovery, 2022

    2022

  9. [17]

    Lm-polygraph: Uncertainty estimation for language models

    Fadeeva, E., Vashurin, R., Tsvigun, A., Vazhentsev, A., Petrakov, S., Fedyanin, K., Vasilev, D., Goncharova, E., Panchenko, A., Panov, M., et al. Lm-polygraph: Uncertainty estimation for language models. In EMNLP, 2023

    2023

  10. [18]

    L., Waseem, Z., and Tsvetkov, Y

    Field, A., Blodgett, S. L., Waseem, Z., and Tsvetkov, Y. A survey of race, racism, and anti-racism in NLP . In ACL, 2021

    2021

  11. [19]

    Unsupervised quality estimation for neural machine translation

    Fomicheva, M., Sun, S., Yankovskaya, L., Blain, F., Guzm \'a n, F., Fishel, M., Aletras, N., Chaudhary, V., and Specia, L. Unsupervised quality estimation for neural machine translation. In ACL, 2020

    2020

  12. [20]

    Open llm leaderboard v2, 2024

    Fourrier, C., Habib, N., Lozovskaya, A., Szafer, K., and Wolf, T. Open llm leaderboard v2, 2024

    2024

  13. [21]

    and Ghahramani, Z

    Gal, Y. and Ghahramani, Z. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In ICML, 2016

    2016

  14. [22]

    O., Rossi, R

    Gallegos, I. O., Rossi, R. A., Barrow, J., Tanjim, M. M., Kim, S., Dernoncourt, F., Yu, T., Zhang, R., and Ahmed, N. K. Bias and fairness in large language models: A survey. Computational Linguistics, 2024

    2024

  15. [23]

    Gawlikowski, J., Tassi, C. R. N., Ali, M., Lee, J., Humt, M., Feng, J., Kruspe, A., Triebel, R., Jung, P., Roscher, R., et al. A survey of uncertainty in deep neural networks. Artificial Intelligence Review, 2023

    2023

  16. [24]

    Uncertainty-guided optimization on large language model search trees

    Grosse, J., Wu, R., Rashid, A., Hennig, P., Poupart, P., and Kristiadi, A. Uncertainty-guided optimization on large language model search trees. In AABI, 2024

    2024

  17. [25]

    and Chen, Y

    Guo, X. and Chen, Y. Generative ai for synthetic data generation: Methods, challenges and the future. arXiv preprint arXiv:2403.04190, 2024

    2024 arXiv

  18. [26]

    Guo, Y., Guo, M., Su, J., Yang, Z., Zhu, M., Li, H., Qiu, M., and Liu, S. S. Bias in large language models: Origin, evaluation, and mitigation. arXiv preprint arXiv:2411.10915, 2024

    2024 arXiv

  19. [27]

    Equality of opportunity in supervised learning

    Hardt, M., Price, E., and Srebro, N. Equality of opportunity in supervised learning. In NeurIPS, 2016

    2016

  20. [28]

    Toxigen: A large-scale machine-generated dataset for adversarial and implicit hate speech detection

    Hartvigsen, T., Gabriel, S., Palangi, H., Sap, M., Ray, D., and Kamar, E. Toxigen: A large-scale machine-generated dataset for adversarial and implicit hate speech detection. In ACL, 2022

    2022

  21. [29]

    and Gimpel, K

    Hendrycks, D. and Gimpel, K. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017

    2017

  22. [30]

    Measuring massive multitask language understanding

    Hendrycks, D., Burns, C., Basart, S., Zou, A., Mazeika, M., Song, D., and Steinhardt, J. Measuring massive multitask language understanding. In ICLR, 2021

    2021

  23. [31]

    Uncertainty in natural language processing: Sources, quantification, and applications

    Hu, M., Zhang, Z., Zhao, S., Huang, M., and Wu, B. Uncertainty in natural language processing: Sources, quantification, and applications. arXiv preprint arXiv:2306.04459, 2023

    2023 arXiv

  24. [32]

    Look before you leap: An exploratory study of uncertainty measurement for large language models

    Huang, Y., Song, J., Wang, Z., Zhao, S., Chen, H., Juefei-Xu, F., and Ma, L. Look before you leap: An exploratory study of uncertainty measurement for large language models. arXiv preprint arXiv:2307.10236, 2023

    2023 arXiv

  25. [33]

    L., Bahl, L

    Jelinek, F., Mercer, R. L., Bahl, L. R., and Baker, J. K. Perplexity—a measure of the difficulty of speech recognition tasks. The Journal of the Acoustical Society of America, 1977

    1977

  26. [34]

    Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D

    Jiang, A. Q., Sablayrolles, A., Mensch, A., Bamford, C., Chaplot, D. S., Casas, D. d. l., Bressand, F., Lengyel, G., Lample, G., Saulnier, L., et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023

    2023 arXiv

  27. [35]

    Q., Sablayrolles, A., Roux, A., Mensch, A., Savary, B., Bamford, C., Chaplot, D

    Jiang, A. Q., Sablayrolles, A., Roux, A., Mensch, A., Savary, B., Bamford, C., Chaplot, D. S., Casas, D. d. l., Hanna, E. B., Bressand, F., et al. Mixtral of experts. arXiv preprint arXiv:2401.04088, 2024

    2024 arXiv

  28. [36]

    and Martin, J

    Jurafsky, D. and Martin, J. H. Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition. Prentice Hall, 2000

    2000

  29. [37]

    Language models (mostly) know what they know

    Kadavath, S., Conerly, T., Askell, A., Henighan, T., Drain, D., Perez, E., Schiefer, N., Hatfield-Dodds, Z., DasSarma, N., Tran-Johnson, E., et al. Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221, 2022

    2022 arXiv

  30. [38]

    Uncertainty-aware predictive modeling for fair data-driven decisions

    Kaiser, P., Kern, C., and R \"u gamer, D. Uncertainty-aware predictive modeling for fair data-driven decisions. arXiv preprint arXiv:2211.02730, 2022

    2022 arXiv

  31. [39]

    and Gal, Y

    Kendall, A. and Gal, Y. What uncertainties do we need in bayesian deep learning for computer vision? NeurIPS, 2017

    2017

  32. [40]

    Gender bias and stereotypes in large language models

    Kotek, H., Dockum, R., and Sun, D. Gender bias and stereotypes in large language models. In Proceedings of the ACM collective intelligence conference, 2023

    2023

  33. [41]

    Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation

    Kuhn, L., Gal, Y., and Farquhar, S. Semantic uncertainty: Linguistic invariances for uncertainty estimation in natural language generation. In ICLR, 2023

    2023

  34. [42]

    Uncertainty estimation for debiased models: Does fairness hurt reliability? In Park, J

    Kuzmin, G., Vazhentsev, A., Shelmanov, A., Han, X., Suster, S., Panov, M., Panchenko, A., and Baldwin, T. Uncertainty estimation for debiased models: Does fairness hurt reliability? In Park, J. C., Arase, Y., Hu, B., Lu, W., Wijaya, D., Purwarianti, A., and Krisnadhi, A. A. (e...

    2023

  35. [43]

    Uncertainty as a fairness measure

    Kuzucu, S., Cheong, J., Gunes, H., and Kalkan, S. Uncertainty as a fairness measure. arXiv preprint arXiv:2312.11299, 2023

    2023 arXiv

  36. [44]

    Laskar, M. T. R., Bari, M. S., Rahman, M., Bhuiyan, M. A. H., Joty, S., and Huang, J. X. A systematic study and comprehensive evaluation of chatgpt on benchmark datasets. In ACL 2023, 2023

    2023

  37. [45]

    End-to-end neural coreference resolution

    Lee, K., He, L., Lewis, M., and Zettlemoyer, L. End-to-end neural coreference resolution. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, 2017

    2017

  38. [46]

    The winograd schema challenge

    Levesque, H., Davis, E., and Morgenstern, L. The winograd schema challenge. In KR, 2012

    2012

  39. [47]

    Collecting a large-scale gender bias dataset for coreference resolution and machine translation

    Levy, S., Lazar, K., and Stanovsky, G. Collecting a large-scale gender bias dataset for coreference resolution and machine translation. In Findings of the Association for Computational Linguistics: EMNLP 2021, 2021

    2021

  40. [48]

    A survey on fairness in large language models

    Li, Y., Du, M., Song, R., Wang, X., and Wang, Y. A survey on fairness in large language models. arXiv preprint arXiv:2308.10149, 2023

    2023 arXiv

  41. [49]

    Holistic evaluation of language models

    Liang, P., Bommasani, R., Lee, T., Tsipras, D., Soylu, D., Yasunaga, M., Zhang, Y., Narayanan, D., Wu, Y., Kumar, A., et al. Holistic evaluation of language models. arXiv preprint arXiv:2211.09110, 2022

    2022 arXiv

  42. [50]

    Liu, Z., Qiao, A., Neiswanger, W., Wang, H., Tan, B., Tao, T., Li, J., Wang, Y., Sun, S., Pangarkar, O., Fan, R., Gu, Y., Miller, V., Zhuang, Y., He, G., Li, H., Koto, F., Tang, L., Ranjan, N., Shen, Z., Ren, X., Iriondo, R., Mu, C., Hu, Z., Schulze, M., Nakov, P., Baldwin, T....

    2023 arXiv

  43. [51]

    On llms-driven synthetic data generation, curation, and evaluation: A survey

    Long, L., Wang, R., Xiao, R., Zhao, J., Ding, X., Chen, G., and Wang, H. On llms-driven synthetic data generation, curation, and evaluation: A survey. arXiv preprint arXiv:2406.15126, 2024

    2024 arXiv

  44. [52]

    Source2synth: Synthetic data generation and curation grounded in real data sources

    Lupidi, A., Gemmell, C., Cancedda, N., Dwivedi-Yu, J., Weston, J., Foerster, J., Raileanu, R., and Lomeli, M. Source2synth: Synthetic data generation and curation grounded in real data sources. arXiv preprint arXiv:2409.08239, 2024

    2024 arXiv

  45. [53]

    Evaluating gender bias transfer between pre-trained and prompt-adapted language models

    Mackraz, N., Sivakumar, N., Khorshidi, S., Patel, K., Theobald, B.-J., Zappella, L., and Apostoloff, N. Evaluating gender bias transfer between pre-trained and prompt-adapted language models. arXiv preprint arXiv:2412.03537, 2024

    2024 arXiv

  46. [54]

    A framework for automated measurement of responsible ai harms in generative ai applications

    Magooda, A., Helyar, A., Jackson, K., Sullivan, D., Atalla, C., Sheng, E., Vann, D., Edgar, R., Palangi, H., Lutz, R., et al. A framework for automated measurement of responsible ai harms in generative ai applications. arXiv preprint arXiv:2310.17750, 2023

    arXiv

  47. [55]

    and Gales, M

    Malinin, A. and Gales, M. Predictive uncertainty estimation via prior networks. In NeurIPS, 2018

    2018

  48. [56]

    and Gales, M

    Malinin, A. and Gales, M. Uncertainty estimation in autoregressive structured prediction. In ICLR, 2021

    2021

  49. [57]

    Evaluating the fairness of deep learning uncertainty estimates in medical image analysis

    Mehta, R., Shui, C., and Arbel, T. Evaluating the fairness of deep learning uncertainty estimates in medical image analysis. In Medical Imaging with Deep Learning, 2024

    2024

  50. [58]

    Generating bilingual example sentences with large language models as lexicography assistants

    Merx, R., Vylomova, E., and Kurniawan, K. Generating bilingual example sentences with large language models as lexicography assistants. arXiv preprint arXiv:2410.03182, 2024

    2024 arXiv

  51. [59]

    Hello gpt-4o

    OpenAI. Hello gpt-4o. https://openai.com, 2024

    2024

  52. [60]

    and Belinkov, Y

    Orgad, H. and Belinkov, Y. Choose your lenses: Flaws in gender bias evaluation. In Proceedings of the 4th Workshop on Gender Bias in Natural Language Processing (GeBNLP), 2022

    2022

  53. [61]

    and Dadu, T

    Pant, K. and Dadu, T. Incorporating subjectivity into gendered ambiguous pronoun ( GAP ) resolution using style transfer. In GeBNLP, 2022

    2022

  54. [62]

    M., and Bowman, S

    Parrish, A., Chen, A., Nangia, N., Padmakumar, V., Phang, J., Thompson, J., Htut, P. M., and Bowman, S. R. Bbq: A hand-built bias benchmark for question answering. In ACL, 2021

    2021

  55. [63]

    Fairness dynamics during training

    Patel, K., Sivakumar, N., Theobald, B.-J., Zappella, L., and Apostoloff, N. Fairness dynamics during training. Neurips Evaluating Evaluations: Examining Best Practices for Measuring Broader Impacts of Generative AI Workshop 2024, 2024

    2024

  56. [64]

    Rousseeuw, P. J. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. In Journal of computational and applied mathematics. Elsevier, 1987

    1987

  57. [65]

    Gender bias in coreference resolution

    Rudinger, R., Naradowsky, J., Leonard, B., and Van Durme, B. Gender bias in coreference resolution. In NAACL-HLT, 2018

    2018

  58. [66]

    On a spurious interaction between uncertainty scores and answer evaluation metrics in generative qa tasks

    Santilli, A., Xiong, M., Kirchhof, M., Rodriguez, P., Danieli, F., Suau, X., Zappella, L., Williamson, S., and Golinski, A. On a spurious interaction between uncertainty scores and answer evaluation metrics in generative qa tasks. In Neurips Safe Generative AI Workshop 2024, 2024

    2024

  59. [67]

    Revisiting uncertainty quantification evaluation in language models: Spurious interactions with response length bias results

    Santilli, A., Golinski, A., Kirchhof, M., Danieli, F., Blaas, A., Xiong, M., Zappella, L., and Williamson, S. Revisiting uncertainty quantification evaluation in language models: Spurious interactions with response length bias results. arXiv preprint arXiv:2504.13677, 2025

    2025 arXiv

  60. [68]

    Bridging the gulf of envisioning: Cognitive challenges in prompt based interactions with llms

    Subramonyam, H., Pea, R., Pondoc, C., Agrawala, M., and Seifert, C. Bridging the gulf of envisioning: Cognitive challenges in prompt based interactions with llms. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, 2024

    2024

  61. [69]

    Fairness through aleatoric uncertainty

    Tahir, A., Cheng, L., and Liu, H. Fairness through aleatoric uncertainty. In CIKM, 2023

    2023

  62. [70]

    Large language models are in-context semantic reasoners rather than symbolic reasoners

    Tang, X., Zheng, Z., Li, J., Meng, F., Zhu, S.-C., Liang, Y., and Zhang, M. Large language models are in-context semantic reasoners rather than symbolic reasoners. arXiv preprint arXiv:2305.14825, 2023

    2023 arXiv

  63. [71]

    Neutral rewriter: A rule-based and neural approach to automatic rewriting into gender-neutral alternatives

    Vanmassenhove, E., Emmery, C., and Shterionov, D. Neutral rewriter: A rule-based and neural approach to automatic rewriting into gender-neutral alternatives. In ACL, 2021

    2021

  64. [72]

    Benchmarking uncertainty quantification methods for large language models with lm-polygraph

    Vashurin, R., Fadeeva, E., Vazhentsev, A., Rvanova, L., Vasilev, D., Tsvigun, A., Petrakov, S., Xing, R., Sadallah, A., Grishchenkov, K., et al. Benchmarking uncertainty quantification methods for large language models with lm-polygraph. Transactions of the Association for Com...

    2025

  65. [73]

    Algorithmic learning in a random world

    Vovk, V., Gammerman, A., and Shafer, G. Algorithmic learning in a random world. Springer, 2005

    2005

  66. [74]

    Ceb: Compositional evaluation benchmark for fairness in large language models

    Wang, S., Wang, P., Zhou, T., Dong, Y., Tan, Z., and Li, J. Ceb: Compositional evaluation benchmark for fairness in large language models. arXiv preprint arXiv:2407.02408, 2024

    2024 arXiv

  67. [75]

    Mind the GAP : A balanced corpus of gendered ambiguous pronouns

    Webster, K., Recasens, M., Axelrod, V., and Baldridge, J. Mind the GAP : A balanced corpus of gendered ambiguous pronouns. Transactions of the Association for Computational Linguistics, 6, 2018

    2018

  68. [76]

    Qwen2 technical report, 2024 a

    Yang, A., Yang, B., Hui, B., Zheng, B., Yu, B., Zhou, C., Li, C., Li, C., Liu, D., Huang, F., Dong, G., Wei, H., Lin, H., Tang, J., Wang, J., Yang, J., Tu, J., Zhang, J., Ma, J., Yang, J., Xu, J., Zhou, J., Bai, J., He, J., Lin, J., Dang, K., Lu, K., Chen, K., Yang, K., Li, M....

    2024 arXiv

  69. [77]

    Assessing adversarial robustness of large language models: An empirical study

    Yang, Z., Meng, Z., Zheng, X., and Wattenhofer, R. Assessing adversarial robustness of large language models: An empirical study. In KDD, 2024 b

    2024

  70. [78]

    F., Yilmaz, E., Shi, S., and Tu, Z

    Ye, F., Yang, M., Pang, J., Wang, L., Wong, D. F., Yilmaz, E., Shi, S., and Tu, Z. Benchmarking llms via uncertainty quantification. arXiv preprint arXiv:2401.12794, 2024

    2024 arXiv

  71. [79]

    Learning uncertainty for unknown domains with zero-target-assumption

    Yu, Y., Sajjad, H., and Xu, J. Learning uncertainty for unknown domains with zero-target-assumption. In ICLR, 2022

    2022

  72. [80]

    Gender bias in coreference resolution: Evaluation and debiasing methods

    Zhao, J., Wang, T., Yatskar, M., Ordonez, V., and Chang, K.-W. Gender bias in coreference resolution: Evaluation and debiasing methods. In NAACL, 2018

    2018

  73. [81]

    D., Ren, X., and Sap, M

    Zhou, K., Hwang, J. D., Ren, X., and Sap, M. Relying on the unreliable: The impact of language models' reluctance to express uncertainty. In ACL, 2024

    2024

  74. [82]

    P ro SA : Assessing and understanding the prompt sensitivity of LLM s

    Zhuo, J., Zhang, S., Fang, X., Duan, H., Lin, D., and Chen, K. P ro SA : Assessing and understanding the prompt sensitivity of LLM s. In EMNLP, 2024

    2024

  75. [83]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

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