Pith. sign in

REVIEW 1 major objections 61 references

PecMan jointly models subgroup reliability, decision allocation, and collaborative prediction to improve fairness-accuracy trade-offs in medical image analysis.

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

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

T0 review · grok-4.3

2026-07-01 08:24 UTC pith:K4U2ZSUG

load-bearing objection PecMan combines fairness and human-AI allocation in one framework for medical imaging but the abstract leaves the actual results and methods uncheckable. the 1 major comments →

arxiv 2604.26991 v2 pith:K4U2ZSUG submitted 2026-04-28 cs.LG cs.AI

People-Centred Medical Image Analysis via Fairness-Aware Human-AI Cooperation

Zheng Zhang , Milad Masroor , Cuong Nguyen , Tahir Hassan , Yuanhong Chen , David Rosewarne , Kevin Wells , Thanh-Toan Do
show 1 more author
Gustavo Carneiro
This is my paper
classification cs.LG cs.AI
keywords medical image analysisfairnesshuman-AI cooperationsubgroup performancelearning to deferbenchmark
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Medical imaging models show performance gaps across patient subgroups, which affects how decisions should be split between automated systems and human experts under limited resources. PecMan tackles this by training subgroup-specialised predictors and adding a gating mechanism that routes each case to an AI model, a human, or both, without needing sensitive attributes at test time. The framework also supplies a new benchmark called FairHAI and a theoretical analysis of selection regret and fairness-coverage trade-offs. A sympathetic reader would care because prior work handled fairness and human-AI cooperation in isolation, while real medical systems must manage both at once. Experiments on multiple datasets show the joint approach yields better overall balances than separate methods.

Core claim

PecMan achieves consistently improved trade-offs among predictive accuracy, subgroup equity, and human involvement by combining subgroup-specialised predictors with a gating and consolidation mechanism that dynamically assigns cases to automated models, human experts, or their combination.

What carries the argument

Subgroup-specialised predictors together with a gating and consolidation mechanism that performs input-dependent allocation without sensitive attributes at test time.

Load-bearing premise

Subgroup-specialised predictors can be trained effectively and the gating mechanism can allocate decisions accurately based on input features without access to sensitive attributes during testing.

What would settle it

Running the same experiments on additional medical imaging datasets and finding that PecMan produces no better accuracy-equity-human-involvement trade-offs than methods addressing fairness or cooperation separately would falsify the central claim.

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

If this is right

  • Subgroup-specialised predictors raise reliability for different patient groups under the same allocation rules.
  • Input-dependent allocation without test-time sensitive attributes supports practical deployment where such attributes are unavailable.
  • The multi-agent gating analysis supplies bounds on selection regret and characterises fairness-coverage trade-offs.
  • The FairHAI benchmark supplies a concrete way to measure joint performance on accuracy, equity, and human workload.

Where Pith is reading between the lines

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

  • If the gating works reliably from input features alone, deployed systems may avoid collecting sensitive attributes at inference time.
  • The same joint modeling of reliability and allocation could be tested in other high-stakes classification settings that involve both subgroup disparities and expert review.
  • The selection-regret analysis offers a template for bounding performance in other multi-agent routing problems.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

1 major / 0 minor

Summary. The paper introduces PecMan, a framework for fairness-aware human-AI cooperative classification in medical image analysis. It jointly models subgroup-dependent reliability, decision allocation, and collaborative prediction using subgroup-specialised predictors and a gating mechanism that assigns cases to AI, humans, or combination without sensitive attributes at test time. The work also presents the FairHAI benchmark for evaluating trade-offs between accuracy, equity, and human involvement, along with a theoretical analysis of selection regret and fairness-coverage trade-offs. Experiments on multiple medical imaging datasets are claimed to show improved trade-offs over methods addressing fairness or human-AI cooperation in isolation.

Significance. If the results hold, this work would be significant for integrating fairness into human-AI decision making in medical domains. The theoretical analysis of multi-agent gating via selection regret and the characterisation of fairness-coverage trade-offs under input-dependent allocation represent a strength, providing analytical grounding beyond empirical claims. The introduction of the FairHAI benchmark could facilitate future research on these trade-offs.

major comments (1)
  1. [Abstract] The abstract states the framework and claims improved trade-offs but provides no details on mathematical formulations, experimental setups, statistical tests, or data handling, preventing assessment of whether the data supports the claims.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their review and for acknowledging the potential significance of our work, particularly the theoretical analysis and the FairHAI benchmark. Below we address the single major comment point by point.

read point-by-point responses

  1. Referee: [Abstract] The abstract states the framework and claims improved trade-offs but provides no details on mathematical formulations, experimental setups, statistical tests, or data handling, preventing assessment of whether the data supports the claims.

    Authors: Abstracts are intentionally concise to provide a high-level overview within strict length limits and are not intended to contain full mathematical formulations, experimental protocols, or statistical details; these are provided in the body of the manuscript (mathematical model in Section 3, experimental setups and data handling in Section 4, statistical tests and results in Section 5). The referee's own summary demonstrates a clear understanding of the framework, benchmark, theoretical contributions, and experimental claims, indicating that the full manuscript permits proper assessment. We are nevertheless willing to revise the abstract to incorporate one or two additional sentences highlighting the key elements of the FairHAI benchmark and the selection-regret analysis if the editor deems it beneficial. revision: partial

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper introduces PecMan as a joint framework for fairness-aware allocation, the FairHAI benchmark, and a theoretical analysis of selection regret under input-dependent gating. No equations, fitted parameters, or self-citations are presented that reduce the central claims (improved accuracy-equity-human-involvement trade-offs) to inputs by construction. The empirical results across datasets and the regret analysis are described as independent of any self-definitional loops or renamed known results, making the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available; no specific free parameters, axioms, or invented entities can be identified from the provided text.

pith-pipeline@v0.9.1-grok · 5750 in / 1177 out tokens · 57574 ms · 2026-07-01T08:24:17.586973+00:00 · methodology

0 comments
read the original abstract

Machine learning models for medical image analysis often exhibit subgroup-dependent performance, which impacts how decisions should be allocated between automated systems and human experts under limited resources. Prior work on AI fairness and human-AI cooperation, including learning to defer (L2D) and learning to complement (L2C), typically addresses these problems in isolation. We propose People-Centred Medical Image Analysis (PecMan), a framework for fairness-aware human-AI co-operative classification that jointly models subgroup-dependent reliability, decision allocation, and collaborative prediction. PecMan combines subgroup-specialised predictors with a gating and consolidation mechanism that dynamically assigns cases to automated models, human experts, or their combination, without requiring sensitive attributes at test time. We also introduce the FairHAI benchmark for evaluating trade-offs between predictive accuracy, subgroup equity, and human involvement. In addition, we provide a theoretical analysis of multi-agent gating via selection regret and characterise fairness-coverage trade-offs under input-dependent allocation. Experiments across multiple medical imaging datasets demonstrate that PecMan achieves consistently improved trade-offs compared to methods that address fairness or human-AI cooperation separately.

Figures

Figures reproduced from arXiv: 2604.26991 by Cuong Nguyen, David Rosewarne, Gustavo Carneiro, Kevin Wells, Milad Masroor, Tahir Hassan, Thanh-Toan Do, Yuanhong Chen, Zheng Zhang.

Figure 1
Figure 1. Figure 1: PecMan: A unified framework that jointly optimises fairness and human-AI collaboration. A gating mechanism selects the appropriate cohort-specific AI model and determines whether clinician input is needed, ensuring high accuracy, balanced group performance, and adherence to workload constraints. Although AI fairness, L2D, and L2C all aim to improve AI-assisted med￾ical decision-making, they have traditiona… view at source ↗
Figure 2
Figure 2. Figure 2: Step 0 – Backbone Training: PecMan initialises its backbone model using the FIS loss [82], which jointly optimises overall classification accuracy and fairness across patient groups, which in this case represent the sensitive attribute sex with values “male” and “female” view at source ↗
Figure 3
Figure 3. Figure 3: Step 1 – Group-specific Model Training: This step focuses on training classifiers tailored to individual patient cohorts, enabling fairness-aware performance across demographic groups. weights are defined as follows: s I (x, y, B) = exp(ℓBCE(hϕ (fθ(x)), y)) P (˜x,M˜ ,y, ˜ a˜)∈B exp(ℓBCE(hϕ (fθ(˜x)), y˜)), s G(a, B) = exp (DOT(L(B),La(B))) P j∈A exp (DOT(L(B),Lj (B))), (3) where DOT(L(B),La(B)) is the optim… view at source ↗
Figure 4
Figure 4. Figure 4: Step 2 – L2D+L2C Unbiased Training: PecMan trains the gating and consolidator models using the FIS loss, enabling unbiased decision-making that combines L2D and L2C strategies. 3.2.1. Step 0: Backbone Model Training - view at source ↗
Figure 5
Figure 5. Figure 5: The AUC vs coverage (top row) and ES-AUC vs. coverage (bottom row) of com view at source ↗
Figure 6
Figure 6. Figure 6: Performance analysis of PecMan on the testing samples of HAM10000. (a) The view at source ↗
Figure 7
Figure 7. Figure 7: The cohort-specific AUC (a,b), overall AUC (C), and ES-AUC vs. coverage view at source ↗
Figure 8
Figure 8. Figure 8: Training time of PecMan and competing methods on HAM10000 dataset. view at source ↗
Figure 9
Figure 9. Figure 9: Inference time of PecMan and competing methods on HAM10000 dataset. view at source ↗

Review history (2 revisions) →

discussion (0)

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

Reference graph

Works this paper leans on

61 extracted references

  1. [1]

    These developments are particularly impactful in ra- diology, where the volume of medical imaging far exceeds the capacity of the available workforce [2]

    Introduction The rapid advancements in medical image analysis have led to highly ac- curate data-centric medical AI models that are increasingly being integrated into clinical practice [1]. These developments are particularly impactful in ra- diology, where the volume of medical imaging far exceeds the capacity of the available workforce [2]. By reliably ...

  2. [2]

    For clinicians and healthcare providers, we advocate for AI as an assistive tool that enhances human-centred decision-making rather than replacing it [3]

    Literature Review This paper goes beyond classification accuracy to address the needs of key stakeholders in medical AI (i.e., clinicians, healthcare settings, patients, and regulatory agencies) by aligning system design with their workflows and pri- orities to foster trust, usability, and clinical effectiveness. For clinicians and healthcare providers, w...

  3. [3]

    Pre-processingapproaches(e.g., SMOTE[30])re-balancedatasetsprior to training

  4. [4]

    Post-processing methods (e.g., equalised odds [31]) adjust outputs after training

  5. [5]

    Among these, in-processing methods tend to show superior results, so we focus on these methods below

    In-processing methods directly modify the training to reduce bias while maintaining performance [24, 32, 33, 34]. Among these, in-processing methods tend to show superior results, so we focus on these methods below. For instance, adversarial training [24, 32] has been widely used to minimise a model’s ability to predict sensitive at- tributes, though ofte...

  6. [6]

    AI defers decisions to human experts

  7. [7]

    learning to complement

    AI makes a joint decision with human experts. Below, we explain the main HAI-CC approaches being studied in the litera- ture, relating them to these three options. �������� �� ����� ������L2D methods primarily focus on options (1) and (2), where either the AI or the human makes the final decision. These ap- proaches aim to jointly learn a predictor and a ...

  8. [8]

    E. J. Topol, High-performance medicine: The convergence of human and artificial intelligence, Nature medicine 25 (1) (2019) 44–56

    2019

  9. [9]

    M. M. Abuzaid, W. Elshami, H. Tekin, B. Issa, Assessment of the will- ingness of radiologists and radiographers to accept the integration of artificial intelligence into radiology practice, Academic Radiology 29 (1) (2022) 87–94

    2022

  10. [10]

    Derevianko, S

    A. Derevianko, S. F. M. Pizzoli, F. Pesapane, A. Rotili, D. Monzani, R. Grasso, E. Cassano, G. Pravettoni, The use of artificial intelligence (ai) in the radiology field: What is the state of doctor–patient commu- nication in cancer diagnosis?, Cancers 15 (2) (2023) 470

    2023

  11. [11]

    Zhang, Y

    S. Zhang, Y. Li, W. Liu, Q. Chu, Y. Chen, A decade of review in global regulation and research of artificial intelligence medical devices (2015- 2025), Frontiers in Medicine 12 (2025) 1630408

    2015

  12. [12]

    Jones, J

    C. Jones, J. Thornton, J. C. Wyatt, Artificial intelligence and clinical decision support: clinicians’ perspectives on trust, trustworthiness, and liability, Medical law review 31 (4) (2023) 501–520

    2023

  13. [13]

    Kumah, Artificial intelligence in healthcare and its implications for patient centered care, Discover Public Health 22 (1) (2025) 524

    E. Kumah, Artificial intelligence in healthcare and its implications for patient centered care, Discover Public Health 22 (1) (2025) 524

    2025

  14. [14]

    S. S. Jain, S. Goto, J. L. Hall, S. S. Khan, C. A. MacRae, C. Ofori, C. Pegus, M. Pencina, E. D. Peterson, L. H. Schwamm, et al., Pragmatic approaches to the evaluation and monitoring of artificial intelligence in health care: A science advisory from the american heart association, Circulation 152 (23) (2025) e433–e442

    2025

  15. [15]

    E. U. Alum, O. P.-C. Ugwu, Artificial intelligence in personalized medicine: transforming diagnosis and treatment, Discover Applied Sci- ences 7 (3) (2025) 193

    2025

  16. [16]

    L. A. Celi, J. Cellini, M.-L. Charpignon, E. C. Dee, F. Dernoncourt, R. Eber, W. G. Mitchell, L. Moukheiber, J. Schirmer, J. Situ, et al., Sources of bias in artificial intelligence that perpetuate healthcare dis- parities - a global review, PLOS Digital Health 1 (3) (2022) e0000022. 22

    2022

  17. [17]

    Oakden-Rayner, J

    L. Oakden-Rayner, J. Dunnmon, G. Carneiro, C. Ré, Hidden stratifica- tion causes clinically meaningful failures in machine learning for medical imaging, in: ACM CHIL, 2020, pp. 151–159

    2020

  18. [18]

    M. A. Ricci Lara, R. Echeveste, E. Ferrante, Addressing fairness in artificialintelligenceformedicalimaging, NatureCommunications13(1) (2022) 4581

    2022

  19. [19]

    Madras, T

    D. Madras, T. Pitassi, R. Zemel, Predict responsibly: Improving fairness and accuracy by learning to defer, in: NeurIPS, Vol. 31, 2018

    2018

  20. [20]

    Wilder, E

    B. Wilder, E. Horvitz, E. Kamar, Learning to complement humans, in: International Joint Conference on Artificial Intelligence, 2021

    2021

  21. [21]

    Y. Zong, Y. Yang, T. Hospedales, MEDFAIR: Benchmarking fairness for medical imaging, in: ICLR, 2023

    2023

  22. [22]

    Iqbal, M

    T. Iqbal, M. Masud, B. Amin, C. Feely, M. Faherty, T. Jones, M. Tier- ney, A. Shahzad, P. Vazquez, Towards integration of artificial intelli- gence into medical devices as a real-time recommender system for per- sonalised healthcare: State-of-the-art and future prospects, Health Sci- ences Review (2024)

    2024

  23. [23]

    Quadrianto, V

    N. Quadrianto, V. Sharmanska, O. Thomas, Discovering fair represen- tations in the data domain, in: CVPR, 2019, pp. 8227–8236

    2019

  24. [24]

    Zhang, J

    Y. Zhang, J. Sang, Towards accuracy-fairness paradox: Adversarial example-based data augmentation for visual debiasing, in: ACM Multi- media, 2020

    2020

  25. [25]

    V. V. Ramaswamy, S. S. Kim, O. Russakovsky, Fair attribute classifica- tion through latent space de-biasing, in: CVPR, 2021, pp. 9301–9310

    2021

  26. [26]

    S. Park, J. Lee, P. Lee, S. Hwang, D. Kim, H. Byun, Fair contrastive learning for facial attribute classification, in: CVPR, 2022, pp. 10389– 10398

    2022

  27. [27]

    Y. Roh, K. Lee, S. Whang, C. Suh, Fr-train: A mutual information- based approach to fair and robust training, in: ICML, PMLR, 2020, pp. 8147–8157. 23

    2020

  28. [28]

    M. B. Zafar, I. Valera, M. G. Rogriguez, K. P. Gummadi, Fairness con- straints: Mechanisms for fair classification, in: AISTATS, PMLR, 2017, pp. 962–970

    2017

  29. [29]

    B. H. Zhang, B. Lemoine, M. Mitchell, Mitigating unwanted biases with adversarial learning, in: AIES, 2018, pp. 335–340

    2018

  30. [30]

    Z. Wang, X. Dong, H. Xue, Z. Zhang, W. Chiu, T. Wei, K. Ren, Fairness-aware adversarial perturbation towards bias mitigation for de- ployed deep models, in: CVPR, 2022, pp. 10379–10388

    2022

  31. [31]

    M. P. Kim, A. Ghorbani, J. Zou, Multiaccuracy: Black-box post- processing for fairness in classification, in: AIES, 2019, pp. 247–254

    2019

  32. [32]

    Herington, M

    J. Herington, M. D. McCradden, K. Creel, R. Boellaard, E. C. Jones, A. K. Jha, A. Rahmim, P. J. Scott, J. J. Sunderland, R. L. Wahl, et al., Ethical considerations for artificial intelligence in medical imaging: de- ployment and governance, Journal of Nuclear Medicine 64 (10) (2023) 1509–1515

    2023

  33. [33]

    Obermeyer, B

    Z. Obermeyer, B. Powers, C. Vogeli, S. Mullainathan, Dissecting racial bias in an algorithm used to manage the health of populations, Science 366 (6464) (2019)

    2019

  34. [34]

    A. J. Larrazabal, N. Nieto, V. Peterson, D. H. Milone, E. Ferrante, Gen- der imbalance in medical imaging datasets produces biased classifiers for computer-aided diagnosis, National Academy of Sciences 117 (23) (2020) 12592–12594

    2020

  35. [35]

    V. C. Nitesh, SMOTE: Synthetic minority over-sampling technique, Journal of Artificial Intelligence Research 16 (1) (2002) 321

    2002

  36. [36]

    Pleiss, M

    G. Pleiss, M. Raghavan, F. Wu, J. Kleinberg, K. Q. Weinberger, On fairness and calibration, in: NeurIPS, Vol. 30, 2017

    2017

  37. [37]

    B. Kim, H. Kim, K. Kim, S. Kim, J. Kim, Learning not to learn: Train- ing deep neural networks with biased data, in: CVPR, 2019, pp. 9012– 9020

    2019

  38. [38]

    Madras, E

    D. Madras, E. Creager, T. Pitassi, R. Zemel, Learning adversarially fair and transferable representations, in: ICML, PMLR, 2018, pp. 3384– 3393. 24

    2018

  39. [39]

    Keswani, M

    V. Keswani, M. Lease, K. Kenthapadi, Towards unbiased and accurate deferral to multiple experts, in: AIES, 2021, pp. 154–165

    2021

  40. [40]

    Narasimhan, W

    H. Narasimhan, W. Jitkrittum, A. K. Menon, A. Rawat, S. Kumar, Post-hoc estimators for learning to defer to an expert, in: NeurIPS, Vol. 35, 2022

    2022

  41. [41]

    A. Mao, C. Mohri, M. Mohri, Y. Zhong, Two-stage learning to defer with multiple experts, in: NeurIPS, 2023

    2023

  42. [42]

    Zhang, C

    Z. Zhang, C. Nguyen, K. Wells, T.-T. Do, D. Rosewarne, G. Carneiro, Coverage-constrained human-ai cooperation with multiple experts, in: AAAI, 2026

    2026

  43. [43]

    Cortes, G

    C. Cortes, G. DeSalvo, M. Mohri, Learning with rejection, in: ALT, Springer, 2016

    2016

  44. [44]

    Charoenphakdee, Z

    N. Charoenphakdee, Z. Cui, Y. Zhang, M. Sugiyama, Classification with rejection based on cost-sensitive classification, in: ICML, PMLR, 2021, pp. 1507–1517

    2021

  45. [45]

    Raghu, K

    M. Raghu, K. Blumer, G. Corrado, J. Kleinberg, Z. Obermeyer, S. Mul- lainathan, The algorithmic automation problem: Prediction, triage, and human effort, in: Machine Learning for Health Symposium, 2018

    2018

  46. [46]

    Okati, A

    N. Okati, A. De, M. Rodriguez, Differentiable learning under triage 34 (2021) 9140–9151

    2021

  47. [47]

    Mozannar, D

    H. Mozannar, D. Sontag, Consistent estimators for learning to defer to an expert, in: ICML, PMLR, 2020, pp. 7076–7087

    2020

  48. [48]

    Verma, E

    R. Verma, E. Nalisnick, Calibrated learning to defer with one-vs-all classifiers, in: ICML, PMLR, 2022, pp. 22184–22202

    2022

  49. [49]

    Mozannar, H

    H. Mozannar, H. Lang, D. Wei, P. Sattigeri, S. Das, D. Sontag, Who should predict? Exact algorithms for learning to defer to humans, in: AISTATS, PMLR, 2023, pp. 10520–10545

    2023

  50. [50]

    Charusaie, H

    M.-A. Charusaie, H. Mozannar, D. Sontag, S. Samadi, Sample efficient learningofpredictorsthatcomplementhumans, in: ICML,PMLR,2022, pp. 2972–3005. 27

    2022

  51. [51]

    Y. Cao, H. Mozannar, L. Feng, H. Wei, B. An, In defense of soft- max parametrization for calibrated and consistent learning to defer, in: NeurIPS, Vol. 36, 2024

    2024

  52. [52]

    Straitouri, L

    E. Straitouri, L. Wang, N. Okati, M. G. Rodriguez, Improving expert predictions with conformal prediction, in: ICML, PMLR, 2023, pp. 32633–32653

    2023

  53. [53]

    S. Liu, Y. Cao, Q. Zhang, L. Feng, B. An, Mitigating underfitting in learning to defer with consistent losses, in: AISTATS, 2024

    2024

  54. [54]

    Mozannar, A

    H. Mozannar, A. Satyanarayan, D. Sontag, Teaching humans when to defer to a classifier via exemplars, in: AAAI Conference on Artificial Intelligence, Vol. 36, 2022, pp. 5323–5331

    2022

  55. [55]

    Verma, D

    R. Verma, D. Barrejón, E. Nalisnick, On the calibration of learning to defer to multiple experts, in: ICML Workshop on HMCT, 2022

    2022

  56. [56]

    Verma, D

    R. Verma, D. Barrejon, E. Nalisnick, Learning to defer to multiple ex- perts: Consistent surrogate losses, confidence calibration, and conformal ensembles, in: AISTATS, PMLR, 2023, pp. 11415–11434

    2023

  57. [57]

    Babbar, U

    V. Babbar, U. Bhatt, A. Weller, On the utility of prediction sets in human-AI teams, in: International Joint Conference on Artificial Intel- ligence, 2022

    2022

  58. [58]

    A. Mao, M. Mohri, Y. Zhong, Principled approaches for learning to defer with multiple experts, in: International Symposium on Artificial Intelligence and Mathematics, 2024

    2024

  59. [59]

    Hemmer, L

    P. Hemmer, L. Thede, M. Vössing, J. Jakubik, N. Kühl, Learning to defer with limited expert predictions, in: AAAI Conference on Artificial Intelligence, Vol. 37, 2023, pp. 6002–6011

    2023

  60. [60]

    Tailor, A

    D. Tailor, A. Patra, R. Verma, P. Manggala, E. Nalisnick, Learning to defer to a population: A meta-learning approach, in: AISTATS, 2024

    2024

  61. [61]

    Leitão, P

    D. Leitão, P. Saleiro, M. A. Figueiredo, P. Bizarro, Human-AI collabora- tion in decision-making: Beyond learning to defer, in: ICML Workshop on Human-Machine Collaboration and Teaming, 2022. 28

    2022