Pith. sign in

REVIEW 4 major objections 6 minor 23 references

Sampling Preferences Yields Simple Trustworthiness Scores

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 Sampling preference weights yields a single trustworthiness score and a winner every time.

desk verdict Simple aggregation method with a real interpretability edge, but the headline comparisons rest on an unspecified preprocessing pipeline and a false Pareto-convergence claim. read the letter →

arxiv 2506.03399 v1 pith:5BSY4PO3 submitted 2025-06-03 cs.HC cs.AI

Sean Steinle This is my paper · ORCID
classification cs.HCcs.AI
keywords preferencesamplingtrustworthinessmulti-criteriadecision-makinglargelanguagemodelsDirichletdistributionweightedsumscalarizationParetooptimalitymodelselection
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

The paper introduces preference sampling, an aggregation method that turns a multi-dimensional trustworthiness evaluation of large language models into a single score and a single recommended model. The idea is to sample many random weightings of the evaluated characteristics from a Dirichlet distribution, apply each weighting to the models' scores, and count how often each model comes out on top. The paper claims that this scheme is consistently reductive — it always narrows the candidate set further than Pareto optimality, which in these datasets never removes more than half the candidates — and that it responds to user priors, whereas averaging gives the same ranking regardless of what a user values. If the claim holds, practitioners get an interpretable, preference-sensitive way to decide which of many well-scored models to actually deploy.

What carries the argument

The central object is the preference vector: a stochastic vector of weights over trustworthiness characteristics, sampled from the Dirichlet distribution. The Dirichlet $\alpha$ parameter carries the user's prior — a symmetric low-magnitude vector like $[1,\dots,1]$ expresses neutral low-confidence preferences, raising one component expresses a preference for that characteristic, and increasing total magnitude expresses greater confidence. Weighted-sum scalarization, $s = \sum_{c=1}^{C} w_c x_c$, turns each candidate model's multidimensional scores into a single scalar per preference sample, and the model with the highest scalar wins that sample. Aggregating over many samples yields a trustworthiness score per model: the share of sampled preferences for which that model is optimal. This mechanism simultaneously produces sparsity (only preference-real models get nonzero scores), user sensitivity ($\alpha$ changes the winner), and interpretability (scores are probabilities of being the most trustworthy).

What would settle it

Recompute preference sampling and Pareto optimality on the TrustLLM and DecodingTrust scores under two explicit normalization schemes (e.g., min-max versus z-score, and a consistent rule for flipping minimized objectives) and check whether the recommended model changes; if it does, the method's output is not well-defined. Likewise, finding any candidate set where Pareto optimality removes more than half the models would contradict the paper's claimed upper bound.

Watch

Extended reading notes

Core claim

Preference sampling is a weighted-sum scalarization method with a Bayesian prior over the weights. A preference vector is a stochastic vector over trustworthiness characteristics, drawn from the Dirichlet distribution with parameter $\alpha$; weighted-sum scalarization $s = \sum_{c=1}^{C} w_c x_c$ converts each model's multidimensional scores into one number, and the highest-scoring model wins that sample. Repeating this over many samples gives each model a trustworthiness score equal to the fraction of the preference space in which it is optimal. On TrustLLM and DecodingTrust, the paper finds that this method always reduces the candidate set more than Pareto optimality does (Pareto optimality leaves 11/21, 12/21, 21/21, or 7/8 candidates depending on configuration, while preference sampling identifies a winner and leaves only preference-real models with nonzero score). It also finds that varying $\alpha$ changes the recommended model — a robustness-focused prior selects GPT-4, while a privacy-focused prior selects Llama-2 — and that roughly half of Pareto-optimal models are never optimal for any sampled preference, so Pareto optimality is a weak standard for trustworthiness.

Load-bearing premise

The method assumes that all trustworthiness characteristics can be normalized onto a common scale and combined by weighted sums, with user knowledge captured by a Dirichlet prior, but the paper leaves the normalization unspecified, so the winning model is not well-defined until those choices are fixed.

Editorial extensions

If this is right

  • A user who cares more about privacy than robustness can encode that priority in the alpha parameter, and the recommended model changes accordingly; averaging cannot express this at all.
  • A user's confidence in a priority is also expressible: raising the alpha value for a preferred characteristic monotonically drives that characteristic's best model toward a near-100 percent trustworthiness score.
  • The resulting score has a concrete reading: it is the fraction of sampled preference vectors for which the model is the best among the candidates, which makes cross-ontology comparisons more meaningful than averaged percentages.
  • Preference sampling eliminates candidates that are Pareto-optimal but never optimal for a realistic preference, so it gives a stricter, more decision-relevant notion of trustworthiness than Pareto optimality.
  • Because the score is computed from many cheap weighted sums, the method is fast enough to serve as an interactive tool for exploring tradeoffs by rerunning with different alpha vectors.

Reading between the lines

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

  • The paper leaves open how TrustLLM and DecodingTrust scores are normalized before scalarization; a natural extension is to treat the normalization as part of the method and test whether the winner is stable across min-max, z-score, and rank-based scaling.
  • Because each model's score is a share of a fixed preference space, the aggregation is zero-sum: adding or removing a candidate reallocates shares, so voting-theoretic ideas such as Condorcet winners or agenda effects could predict when the preference-sampling winner is stable under candidate-set changes.
  • The same sampling scheme should transfer to any multi-dimensional evaluation with commensurable scores, such as vision-language model benchmarks, but only if characteristic scores can be placed on a common scale without distorting tradeoffs.
  • A direct mathematical consequence the paper does not spell out: with positive weights, weighted-sum scalarization can only select Pareto-optimal points, so preference sampling is a refinement of, not a replacement for, Pareto optimality.
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 introduces 'preference sampling,' a method that samples preference vectors from a Dirichlet distribution, applies weighted-sum scalarization to multi-dimensional trustworthiness scores, and defines each model's trustworthiness score as the fraction of sampled preference vectors for which it is the unique argmax. The authors claim two principal advantages over existing aggregation methods: (i) preference sampling is 'consistently reductive,' always selecting a single best model, whereas Pareto optimality rarely reduces the candidate set; and (ii) preference sampling is 'consistently sensitive to user priors' through the Dirichlet alpha parameter, whereas simple averaging is not. These claims are demonstrated on two LLM trustworthiness evaluation frameworks, TrustLLM and DecodingTrust, with a series of experiments varying the alpha vector to express robustness, privacy, fairness, and confidence levels. The paper also discusses the preference-real set, the set of models that are optimal for at least one sampled preference, and argues that Pareto optimality is a poor litmus test for trustworthiness because many Pareto-optimal models are never preference-real.

Significance. The proposed method is simple, computationally inexpensive, and offers a clear interpretation: a model's trustworthiness score is the probability, under the user's Dirichlet prior, that it is the most preferred model. If the empirical claims were fully supported, preference sampling would be a useful decision-support tool for multi-criteria model selection, and the paper's demonstration of user-sensitive aggregation would be a practical contribution to the humans-and-AI decision-making literature. The paper also benefits from presenting concrete experiments on two widely used LLM trustworthiness benchmarks. However, the current manuscript overstates its results: the headline '100% reduction' property is definitional rather than empirical, the claimed convergence of the preference-real set to the Pareto set is mathematically incorrect for positive weights, and the empirical comparisons rest on an unspecified preprocessing pipeline. Because these issues affect the central claims, the paper cannot be accepted in its present form.

major comments (4)
  1. [Section 3A, Section 2C] The preprocessing applied to TrustLLM and DecodingTrust scores is never specified. Section 3A states that 'preprocessing procedures like normalization or objective scaling may be necessary' but does not say which were used, and Section 2C notes that TrustLLM mixes maximized and minimized sub-characteristics and that DecodingTrust uses custom per-characteristic aggregation functions. Weighted-sum scalarization is not invariant to per-criterion affine rescalings, so the argmax model for a fixed preference vector can change with the normalization chosen. Consequently, all reported results that depend on the argmax (Tables V-VII, IX, X, XII, and the preference-real set analyses in Figures 5-6) are results about an unspecified preprocessing pipeline, not about preference sampling as defined. The authors must specify the exact normalization and objective-orientation transformation applied to each dataset, or the paper's central empirical claims are not reproducible.
  2. [Section 4A2] The statement that 'as samples tend towards infinity the preference real set converges to the Pareto set' is false for positive weights. A weighted sum with strictly positive weights can only select supported Pareto points; Pareto-optimal points that lie on non-convex portions of the frontier are never optimal for any positive weighted sum. The correct statement is that the preference-real set converges to the set of supported Pareto points. This error undermines the 'surprising' framing that roughly half of Pareto-optimal models were never chosen after 100,000,000 preference samples; for linear scalarization, the existence of unsupported Pareto points is expected. The authors should either correct the convergence claim or restrict the Pareto comparison to supported Pareto points.
  3. [Section 4A1, Abstract] The claim that preference sampling 'fully reduces the set of candidate models 100% of the time' is guaranteed by construction, not an empirical discovery. Since the Dirichlet distribution is continuous, sampled preference vectors are distinct almost surely, and the weighted-sum argmax is unique almost surely (unless two models tie on a low-dimensional subset, which has probability zero). The comparison with Pareto optimality's reduction ratio is therefore a comparison between a procedure that always returns one winner by definition and a procedure that does not. This would be a legitimate design choice, but the paper presents it as an empirical advantage without acknowledging its definitional nature. The authors should reframe this claim as a property of the method, not a finding from the experiments.
  4. [Section 4B2, Table X] The experiments on confidence (2-3-1 through 2-3-6) are not fully reported. For DecodingTrust, the alpha vector is listed as [1,1,1,1,1,1] for every experiment ID 2-3-1 through 2-3-6, which cannot be correct if confidence in a single characteristic is increasing. Either the table omits the relevant alpha components for DecodingTrust, or the experiments do not actually manipulate confidence in the way the text describes. This makes it impossible to verify the monotonicity claim in Section 4B2.
minor comments (6)
  1. [Section 3A] The text uses 'weighed sum scalarization' repeatedly; standard terminology is 'weighted sum scalarization.' Please correct throughout.
  2. [Section 5D] In the last paragraph of Section 5D, 'sset' should be 'set.'
  3. [Section 4A1, Table VIII] Table VIII reports ratios such as '11/21' and '7/8' but does not state the number of models in the candidate set for each configuration; the denominator is clear for TrustLLM (21) and DecodingTrust (8) but the ratio format could be misinterpreted as a percentage on first reading. Consider adding a caption note.
  4. [Section 3C1] Figures 3 and 4 are referenced in the text but the browser-facing version of the paper does not include the actual figures; if this is a submission artifact, the figures should be embedded. If not, the simulation study is not auditable.
  5. [Section 4A2, Figures 5-6] The axes labels for Figures 5 and 6 are not described in the text. The captions should state what is plotted (e.g., number of sampled preferences vs. share of preference space assigned to each model) so the reader can interpret the convergence behavior.
  6. [References] The reference to DecodingTrust [21] cites only the first author and 'et al.' with no year and no venue information; the full citation should be completed.

Circularity Check

2 steps flagged · score 6.0 of 10

The two headline advantages—'consistently reductive' and 'consistently sensitive to user priors'—are built into the method's definitions rather than discovered empirically.

  1. self definitional [Section III.A and Section IV.B.1]
    "we apply the sampled preference vectors to multi-dimensional model scores using weighed-sum scalarization and then choose the highest scoring model. ... a model's trustworthiness score corresponds to the percentage of preference samples where it was the optimal model out of all candidate models."

    The trustworthiness score is defined as the frequency with which a model is the argmax of a weighted sum over sampled preference vectors. Because the weights are drawn from a continuous Dirichlet distribution, ties in the weighted sum occur with probability zero, so every sample almost surely produces a single highest-scoring model. Thus the claim that preference sampling fully reduces the candidate set 100% of the time is entailed by the definition of the score and the argmax operation, not established by the TrustLLM and DecodingTrust experiments.

  2. self definitional [Introduction and Section III.B]
    "the alpha parameter of the Dirichlet distribution will dictate which candidate model is recommended. ... the Dirichlet distribution can be used to express more confidence by increasing the magnitude of the alpha vector or to indicate skew by increasing the relative weight of one of the values."

    Alpha is the sole parameter of the Dirichlet distribution from which preference vectors are sampled, and the method selects the weighted-sum argmax using those vectors. Therefore changing alpha changes the distribution of possible winners by construction. The assertion that preference sampling is 'sensitive to user priors' is a restatement of the method's parameterization, not an empirical result; averaging is 'intransigent' simply because it has no such parameter. The specific winning models under each prior are data-dependent, but the claimed advantage over averaging reduces to the design of the input.

full rationale

The central advertised findings reduce to the method's own definitions. The '100% reductive' claim is the argmax property of weighted-sum scalarization with continuous random weights, and the 'sensitive to user priors' claim is the presence of the alpha parameter in the Dirichlet sampler. These are not predictions tested against the data; they are true by construction. The remaining empirical content—that GPT-4 tends to win under robustness-focused priors and Llama-2 under privacy-focused priors—is not itself circular, and the paper contains no load-bearing self-citation chain. However, this empirical content is underdetermined because Section III.A only notes that 'preprocessing procedures like normalization or objective scaling may be necessary' without specifying what was applied to TrustLLM and DecodingTrust, so the exact numerical conclusions are not fully defined. Weighing the definitional character of the two headline claims against the partly independent empirical rankings, a score of 6 reflects partial circularity rather than complete equivalence.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on user-supplied or unspecified parameters (alpha, sample size, normalization) and on several unstated modeling assumptions about how trustworthiness and preferences behave. No new physical or conceptual entities are introduced; 'preference real' is a defined property, not an entity.

free parameters (3)
  • Dirichlet alpha vector = examples like [1,1,1,1,1,1], [1,1,1,1,2,1], [1,1,1,1,10,1]; also preference vectors P1, P2, P3
    The output trustworthiness scores depend completely on this user-specified parameter. The paper gives examples but no elicitation procedure or default recommendation, so the method's results are not fixed.
  • Number of preference samples = 10,000 to 100,000,000
    The paper varies sample size across experiments and does not report standard errors; the trustworthiness score is a Monte Carlo estimate whose precision depends on this number.
  • Normalization and objective scaling = unspecified
    Section 3A says preprocessing may be necessary for scores not in [0,1] or for minimized objectives, but the exact transformations applied to TrustLLM and DecodingTrust are never stated, so the scalarized scores are not uniquely determined.
assumptions (5)
  • domain assumption Trustworthiness is compositional and can be expressed as a weighted sum of characteristic scores.
    Section II adopts the compositional view without deriving it; the entire method is a weighted-sum composition.
  • domain assumption User preferences over characteristics are representable as a Dirichlet distribution over weight vectors.
    Section IIIB defines prior knowledge through the alpha parameter; no evidence or argument is given that real user preferences follow a Dirichlet distribution or that linear weights capture them.
  • domain assumption The model with the highest weighted-sum score is the best model for a given preference.
    Section 3A uses weighted-sum scalarization as the decision rule, which excludes non-linear, lexicographic, or satisficing preferences.
  • domain assumption The percentage of sampled preferences for which a model is optimal is a valid scalar trustworthiness score.
    Section III defines the trustworthiness score as this percentage; it is a modeling choice, not derived from decision theory.
  • ad hoc to paper Pareto optimal models that are never optimal for any sampled linear preference are not trustworthy.
    Section 4A2 rejects these models, but for linear scalarization their preference region is empty, so the paper's stated justification that only an exceptionally narrow set of preferences would choose them is not correct.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Sampling Preferences Yields Simple Trustworthiness Scores." pith.science (2026). https://pith.science/paper/5BSY4PO3

@misc{pith2026250603399,
  author       = {Pith},
  title        = {Pith review of: Sampling Preferences Yields Simple Trustworthiness Scores},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5BSY4PO3}},
  note         = {Machine review of arXiv:2506.03399}
}
read the original abstract

With the onset of large language models (LLMs), the performance of artificial intelligence (AI) models is becoming increasingly multi-dimensional. Accordingly, there have been several large, multi-dimensional evaluation frameworks put forward to evaluate LLMs. Though these frameworks are much more realistic than previous attempts which only used a single score like accuracy, multi-dimensional evaluations can complicate decision-making since there is no obvious way to select an optimal model. This work introduces preference sampling, a method to extract a scalar trustworthiness score from multi-dimensional evaluation results by considering the many characteristics of model performance which users value. We show that preference sampling improves upon alternate aggregation methods by using multi-dimensional trustworthiness evaluations of LLMs from TrustLLM and DecodingTrust. We find that preference sampling is consistently reductive, fully reducing the set of candidate models 100% of the time whereas Pareto optimality never reduces the set by more than 50%. Likewise, preference sampling is consistently sensitive to user priors-allowing users to specify the relative weighting and confidence of their preferences-whereas averaging scores is intransigent to the users' prior knowledge.

Figures

Figures reproduced from arXiv: 2506.03399 by the authors.

Figure 1
Figure 1. A hierarchical depiction of a portion of TrustLLM’s ontology (green components are maximized, red components are minimized) [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. A hierarchical view of DecodingTrust’s ontology [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Simulation data from uniform distribution [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Shares of TrustLLM preference space as samples increase [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 15 canonical work pages

  1. [1]

    Publications Office, 2019

    European Commission, Content Directorate-General for Communica- tions Networks, and Technology.Ethics guidelines for trustworthy AI. Publications Office, 2019

    work page 2019

  2. [2]

    Explaining models: an empirical study of how explanations impact fairness judgment

    Jonathan Dodge, Q Vera Liao, Yunfeng Zhang, Rachel KE Bellamy, and Casey Dugan. Explaining models: an empirical study of how explanations impact fairness judgment. InProceedings of the 24th international conference on intelligent user interfaces, pages 275–285, 2019

    work page 2019

  3. [3]

    Aaron Fisher, Cynthia Rudin, and Francesca Dominici. All models are wrong, but many are useful: Learning a variable’s importance by studying an entire class of prediction models simultaneously.Journal of Machine Learning Research, 20(177):1–81, 2019

    2019

  4. [4]

    Interpreting black-box models: a review on explainable artificial intelligence.Cognitive Computation, 16(1):45–74, 2024

    Vikas Hassija, Vinay Chamola, Atmesh Mahapatra, Abhinandan Singal, Divyansh Goel, Kaizhu Huang, Simone Scardapane, Indro Spinelli, Mufti Mahmud, and Amir Hussain. Interpreting black-box models: a review on explainable artificial intelligence.Cognitive Computation, 16(1):45–74, 2024

    2024

  5. [5]

    Vanlalruata Hnamte and Jamal Hussain. Dependable intrusion detection system using deep convolutional neural network: A novel framework and performance evaluation approach.Telematics and Informatics Reports, 11:100077, 2023

    work page 2023

  6. [6]

    Monas: Multi-objective neural architecture search using reinforcement learning.arXiv preprint arXiv:1806.10332, 2018

    Chi-Hung Hsu, Shu-Huan Chang, Jhao-Hong Liang, Hsin-Ping Chou, Chun-Hao Liu, Shih-Chieh Chang, Jia-Yu Pan, Yu-Ting Chen, Wei Wei, and Da-Cheng Juan. Monas: Multi-objective neural architecture search using reinforcement learning.arXiv preprint arXiv:1806.10332, 2018

    arXiv 2018

  7. [7]

    Yue Huang, Lichao Sun, Haoran Wang, Siyuan Wu, Qihui Zhang, Yuan Li, Chujie Gao, Yixin Huang, Wenhan Lyu, Yixuan Zhang, Xiner Li, Hanchi Sun, Zhengliang Liu, Yixin Liu, Yijue Wang, Zhikun Zhang, Bertie Vidgen, Bhavya Kailkhura, Caiming Xiong, Chaowei Xiao, Chunyuan Li, Eric P. Xing, Furong Huang, Hao Liu, Heng Ji, Hongyi Wang, Huan Zhang, Huaxiu Yao, Mano...

    work page 2024

  8. [8]

    Weighted sum model with partial preference information: Application to multi-objective optimization.European Journal of Operational Research, 260(2):665–679, 2017

    Sami Kaddani, Daniel Vanderpooten, Jean-Michel Vanpeperstraete, and Hassene Aissi. Weighted sum model with partial preference information: Application to multi-objective optimization.European Journal of Operational Research, 260(2):665–679, 2017

    work page 2017

Show all 23 references
  1. [9]

    Continuous multivariate distributions, Volume 1: Models and applica- tions, volume 334

    Samuel Kotz, Narayanaswamy Balakrishnan, and Norman L Johnson. Continuous multivariate distributions, Volume 1: Models and applica- tions, volume 334. John Wiley & Sons, 2019

    2019

  2. [10]

    It’s morphing time: Unleashing the potential of multiple llms via multi-objective optimization.arXiv preprint arXiv:2407.00487, 2024

    Bingdong Li, Zixiang Di, Yanting Yang, Hong Qian, Peng Yang, Hao Hao, Ke Tang, and Aimin Zhou. It’s morphing time: Unleashing the potential of multiple llms via multi-objective optimization.arXiv preprint arXiv:2407.00487, 2024

    2024 arXiv

  3. [11]

    Trustworthy ai: From principles to practices.ACM Computing Surveys, 55(9):1–46, 2023

    Bo Li, Peng Qi, Bo Liu, Shuai Di, Jingen Liu, Jiquan Pei, Jinfeng Yi, and Bowen Zhou. Trustworthy ai: From principles to practices.ACM Computing Surveys, 55(9):1–46, 2023

    2023

  4. [12]

    Trustworthy llms: A survey and guideline for evaluating large language models’ alignment.arXiv preprint arXiv:2308.05374, 2023

    Yang Liu, Yuanshun Yao, Jean-Francois Ton, Xiaoying Zhang, Ruocheng Guo, Hao Cheng, Yegor Klochkov, Muhammad Faaiz Taufiq, and Hang Li. Trustworthy llms: A survey and guideline for evaluating large language models’ alignment.arXiv preprint arXiv:2308.05374, 2023

    2023 arXiv

  5. [13]

    Multi-objective decision making for trustworthy ai

    Patrick Mannion, Fredrik Heintz, Thommen George Karimpanal, and Peter Vamplew. Multi-objective decision making for trustworthy ai. InProceedings of the Multi-Objective Decision Making (MODeM) Workshop, 2021

    2021

  6. [14]

    The weighted sum method for multi-objective optimization: new insights.Structural and multidisci- plinary optimization, 41:853–862, 2010

    R Timothy Marler and Jasbir S Arora. The weighted sum method for multi-objective optimization: new insights.Structural and multidisci- plinary optimization, 41:853–862, 2010

    2010

  7. [15]

    Dependable intrusion detection system for iot: A deep transfer learning based approach.IEEE Transactions on Industrial Informatics, 19(1):1006–1017, 2022

    Sk Tanzir Mehedi, Adnan Anwar, Ziaur Rahman, Kawsar Ahmed, and Rafiqul Islam. Dependable intrusion detection system for iot: A deep transfer learning based approach.IEEE Transactions on Industrial Informatics, 19(1):1006–1017, 2022

    2022

  8. [16]

    From anecdotal evidence to quantitative evaluation methods: A systematic review on evaluating explainable ai.ACM Computing Surveys, 55(13s):1–42, 2023

    Meike Nauta, Jan Trienes, Shreyasi Pathak, Elisa Nguyen, Michelle Peters, Yasmin Schmitt, J ¨org Schl ¨otterer, Maurice Van Keulen, and Christin Seifert. From anecdotal evidence to quantitative evaluation methods: A systematic review on evaluating explainable ai.ACM Computing ...

    2023

  9. [17]

    Fairness optimisation with multi-objective swarms for explain- able classifiers on data streams.Complex & Intelligent Systems, pages 1–14, 2024

    Diem Pham, Binh Tran, Su Nguyen, Damminda Alahakoon, and Mengjie Zhang. Fairness optimisation with multi-objective swarms for explain- able classifiers on data streams.Complex & Intelligent Systems, pages 1–14, 2024

    2024

  10. [18]

    Artificial intelligence risk management framework (ai rmf 1.0), 2023-01-26 05:01:00 2023

    Elham Tabassi. Artificial intelligence risk management framework (ai rmf 1.0), 2023-01-26 05:01:00 2023

    2023

  11. [19]

    Robustness may be at odds with accuracy

    Dimitris Tsipras, Shibani Santurkar, Logan Engstrom, Alexander Turner, and Aleksander Madry. Robustness may be at odds with accuracy. In Proceedings of the 7th International Conference on Learning Represen- tations (ICLR 2019), New Orleans, Louisiana, May 2019

    2019

  12. [20]

    Explainable artificial intelligence: a systematic review.arXiv preprint arXiv:2006.00093, 2020

    Giulia Vilone and Luca Longo. Explainable artificial intelligence: a systematic review.arXiv preprint arXiv:2006.00093, 2020

    2006 arXiv

  13. [21]

    Decodingtrust: A comprehensive assessment of trustworthiness in gpt models

    Boxin Wang, Weixin Chen, Hengzhi Pei, Chulin Xie, Mintong Kang, Chenhui Zhang, Chejian Xu, Zidi Xiong, Ritik Dutta, Rylan Schaeffer, et al. Decodingtrust: A comprehensive assessment of trustworthiness in gpt models. 2023

    2023

  14. [22]

    Arithmetic control of LLMs for diverse user preferences: Directional preference alignment with multi- objective rewards

    Haoxiang Wang, Yong Lin, Wei Xiong, Rui Yang, Shizhe Diao, Shuang Qiu, Han Zhao, and Tong Zhang. Arithmetic control of LLMs for diverse user preferences: Directional preference alignment with multi- objective rewards. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors,P...

    2024

  15. [23]

    To be robust or to be fair: Towards fairness in adversarial training

    Han Xu, Xiaorui Liu, Yaxin Li, Anil Jain, and Jiliang Tang. To be robust or to be fair: Towards fairness in adversarial training. InInternational conference on machine learning, pages 11492–11501. PMLR, 2021. VII. LICENSES Images from DecodingTrust shared unchanged in accor- d...

    2021

Pith tools

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