Pith. sign in

REVIEW 5 major objections 6 minor 41 references

DeCoDe: Defer-and-Complement Decision-Making via Decoupled Concept Bottleneck Models

T0 review · 5 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 DeCoDe is a concept-driven human-AI collaboration system that chooses per instance among AI-only, AI+human, and defer-to-human, and reports accuracy gains over binary deferral baselines.

desk verdict Concept-gated three-mode deferral is new and worth examining, but the empirical claims are inflated by an unrealistic simulated human and an underspecified fusion module. read the letter →

arxiv 2505.19220 v1 pith:JTWLMC2S submitted 2025-05-25 cs.AI cs.CY

Chengbo He , Bochao Zou , Junliang Xing , Jiansheng Chen , Yuanchun Shi , Huimin Ma This is my paper · ORCID
classification cs.AIcs.CY
keywords human-AIcollaborationlearningtodeferdefer-and-complementconceptbottleneckmodelinterpretablestrategyselectionsurrogatelosshuman-in-the-loopaccuracy-costtradeoff
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

DeCoDe is a proposed system for human-AI collaboration that replaces the usual binary choice between 'the AI predicts' and 'defer to a human' with three per-instance modes: the AI decides alone, the AI and a human decide together, or the case is handed entirely to a human. The mode is chosen by a gating network that reads human-interpretable concept representations, so the routing decision is semantic rather than a hidden confidence score. Because there are no ground-truth routing labels, the authors train the gating network with a surrogate loss that turns the relative correctness of the AI and a simulated human expert into soft pseudo-labels, and they add a cost term that discourages unnecessary human involvement. On the CUB, Derm7pt, and CelebA datasets under simulated human label noise from 10% to 50%, the paper reports that the three-mode system outperforms AI-only, human-only, and binary learning-to-defer baselines on accuracy-versus-human-cost curves. The practical point is that knowing when a human and the AI should decide together is a useful capability that binary deferral methods do not provide.

What carries the argument

The load-bearing mechanism is the concept-guided strategy selector, a gating network $g_\phi$ that maps the explicit concept vector $c_{\mathrm{exp}}$ onto a probability simplex over three strategies, together with the collaboration module $h_\psi$, which fuses the AI prediction and the expert label according to the selected strategy. The decoupled concept bottleneck supplies the interpretable input: the explicit concept vector $c_{\mathrm{exp}}$ captures human-annotated semantics, while the implicit latent vector $c_{\mathrm{imp}}$ carries task-relevant information outside the concept set, and prediction is $f(c_{\mathrm{exp}}) + \tilde{f}(c_{\mathrm{imp}})$ with a Jensen-Shannon divergence term $L_{JS}$ that keeps the interpretable prediction aligned with the full prediction. The surrogate loss $L_{\mathrm{DeCoDe}}$ is what makes training possible without routing labels: it supervises the strategy distribution against pseudo-labels derived from whether the AI and the simulated human are correct, and adds a cost term $\lambda(g_\phi^{(2)}(x) + g_\phi^{(3)}(x))$ that penalizes human involvement. This combination is what lets the method select per-instance strategies, explain its selection in concept terms, and resist unreliable human input.

What would settle it

Rerun the same accuracy-versus-coverage comparisons with a simulator in which the human error probability rises on instances the AI also tends to misclassify, or collect real expert labels for a held-out subset of CUB, Derm7pt, or CelebA and compare DeCoDe's curve against the binary deferral baselines; the central claim fails if the three-mode curve stops dominating the baselines in the mid-to-high coverage range.

Watch

Extended reading notes

Core claim

The central claim of the paper is that per-instance routing in human-AI collaboration should be a three-way decision over transparent semantic states, not a binary confidence-based handoff. DeCoDe builds this on a decoupled concept bottleneck model: a CNN produces an explicit concept vector $c_{\mathrm{exp}}$ and an implicit latent vector $c_{\mathrm{imp}}$, the final prediction combines both, and a strategy selector $g_\phi$ maps $c_{\mathrm{exp}}$ to a distribution over AI-only, AI+Human, and Defer-to-Human. The selector and a collaboration module $h_\psi$ are optimized jointly by the surrogate loss $L_{\mathrm{DeCoDe}} = L_{\mathrm{strategy}}(r(x), q(x)) + \alpha L_{\mathrm{CE}}(\hat{y}(x), y)$, where the pseudo-label $q(x)$ is derived from a heuristic rule: choose AI-only when the AI is correct, defer when a human is correct and the AI is not, and otherwise collaborate. The paper reports that this design improves accuracy-versus-cost tradeoffs over binary deferral baselines on CUB, Derm7pt, and CelebA at simulated noise levels of 0.1, 0.3, and 0.5, and that concept-based gating is more stable than image-based gating as human participation rises.

Load-bearing premise

The experiments simulate human experts as labelers who flip each answer with a fixed probability regardless of the image and regardless of what the AI predicted, and the reported gains depend on that independence.

Editorial extensions

If this is right

  • Under the tested noise levels, DeCoDe's three-mode routing achieves higher accuracy than AI-only, human-only, and binary deferral baselines at the same human participation ratios.
  • Because routing is a function of explicit concepts, each defer or collaborate decision comes with a semantic trace that a user can inspect and potentially correct.
  • The cost term gives operators a single knob to trade accuracy against human workload, so the same system can be tuned for cheaper or more careful operation.
  • Concept-level intervention and backward rectification let users change both the final label and the routing logic by editing concepts rather than retraining.
  • The full three-mode model is reported to degrade less than a defer-only variant when simulated human annotation quality worsens, since it can choose complementarity instead of blind deferral.

Reading between the lines

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

  • The concept interface suggests a natural next step: use automatically discovered concepts from a vision-language model so the routing mechanism can be applied where manual concept annotations do not exist.
  • The current surrogate loss needs both AI and human labels on the same training instances; an active-learning variant that requests human labels only for instances where routing is uncertain would test whether the gains survive much sparser human feedback.
  • The three-mode selector should generalize to multiple experts with different reliabilities and costs, since each expert can be represented as an additional human-query option with its own cost weight.
  • The authors' noise simulation makes human errors independent of instance difficulty, which likely overstates complementarity; a direct test with real expert annotations would reveal how much of the reported gain is an artifact of that independence.
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

5 major / 6 minor

Summary. DeCoDe proposes a concept-bottleneck-based framework for human-AI collaboration in which a gating network chooses among three strategies: AI-only, AI+Human fusion, and defer-to-human. The gating network operates on explicit concept representations, and a surrogate loss with correctness-based pseudo-labels is used to train routing without ground-truth deferral labels. Experiments on CUB, Derm7pt, and CelebA simulate expert annotations via label noise at 10-50% and compare with L2D baselines using accuracy-coverage curves. The paper also demonstrates concept intervention and a comparison of concept-based versus image-based gating. The central claim is that DeCoDe significantly outperforms AI-only, human-only, and traditional deferral baselines while retaining interpretability.

Significance. If the claims held, this would be a useful contribution: concept-grounded gating offers transparency, the three-mode decision (including complementarity) goes beyond binary L2D, and the surrogate-loss formulation avoids ground-truth routing labels. The paper's strengths include a clean architectural idea, evaluation across three datasets, and a concrete interpretability mechanism via intervention. However, the empirical support is not yet convincing due to the favorable human simulation and underspecified fusion/routing training details; these issues must be resolved before the comparative claims can be accepted. The paper does not provide code or trained models, which further limits reproducibility.

major comments (5)
  1. [Section 4.1, Figure 3] The simulated expert in Section 4.1 is generated by instance-independent label noise: each label is flipped with probability p regardless of the input or the AI's prediction. This creates a conditional independence structure between human and AI errors given the instance, so the expected accuracy of deferring any instance is exactly 1-p, and the complementarity that DeCoDe exploits is guaranteed by construction. Real expert errors are correlated with AI errors on difficult or atypical instances, so the reported gains in Figure 3 may shrink or disappear once this independence is relaxed. Please add experiments with difficulty-correlated or AI-correlated noise (e.g., error probability that depends on instance hardness or the AI's predicted confidence) and, ideally, a small-scale study with real human annotations, and report how the relative performance changes.
  2. [Section 3.2, Eq. (4)] The collaboration module hψ is never specified. Section 3.2 defines only the input selection p(·) in Eq. (4); it does not state how the selected vector is mapped to the final prediction, what hψ is (architecture, parameters, loss), or how Lco in Eq. (5) is implemented. Eq. (9) then introduces LDeCoDe with Lstrategy and LCE, but the relationship between Lco and LDeCoDe is unclear. Without this information, the AI+Human fusion mechanism is not reproducible, and the advantage of the complement mode over simple ensemble averaging cannot be assessed. Please provide the full definition of hψ, the explicit form of Lco, and an ablation showing the fusion contribution.
  3. [Section 3.3, Eq. (8)] The soft pseudo-label variant relies on a 'behavior-driven penalty' ℓ(s)(x), but ℓ(s) is not defined in the main text, and the referenced Appendix A is not included in the manuscript as provided. Since this soft-label loss is part of the claimed novel surrogate loss, the learning objective is not well-defined as submitted. Please either include the appendix or define ℓ(s) in the main text, and state how the heuristic losses are computed from model and expert predictions.
  4. [Figure 3, Section 4.2] The comparative claim is supported by single accuracy-coverage curves without error bars, number of seeds, or significance tests. The abstract's wording 'significantly outperforms' is therefore unsupported. Please report mean and standard deviation over multiple random seeds and perform paired significance tests for at least the headline comparisons, such as DeCoDe versus L2D-REAL and DeCoDe versus DeCoDe_defer_human.
  5. [Section 4.1, Figure 3] The procedure for generating the accuracy-coverage curves is not described. It is unclear how the human participation ratio is varied (e.g., by thresholding the gating probability, by tuning the cost penalty λ, or by a separate rejector for the L2D baselines) and whether all methods use the same protocol. This makes the comparisons in Figure 3 difficult to interpret. Please specify the curve-generation procedure for DeCoDe and each baseline, including how the cost trade-off is realized.
minor comments (6)
  1. [Figure 2] The word 'Prediciton' is misspelled and should read 'Prediction'.
  2. [Figure 5(b)] The color scale for 'Avg Concept Logit' is not labeled; please add a colorbar or otherwise explain the units and aggregation.
  3. [Eq. (3)] The Jensen-Shannon divergence is applied to sums of logits f(g(x)) + f̃(˜g(x)) without an explicit softmax or normalization; please clarify the exact quantities being compared.
  4. [References] Several references contain formatting errors, such as 'URLarXivpreprintarXiv:' without a separator in Kazhdan et al. and Ragkousis and Parbhoo, and a long garbled author string in the Frazer et al. entry.
  5. [Section 3.1] The claim that the Jensen-Shannon constraint in Eq. (3) prevents the model from ignoring the explicit concept signal is not justified; please provide an analytic argument or an empirical demonstration.
  6. [Conclusion] The Conclusion acknowledges that DeCoDe relies on manually annotated concepts; the paper would be strengthened by discussing how the gating network and surrogate loss transfer when concept annotations are unavailable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the strategy selector is trained on oracle-derived pseudo-labels and evaluated on held-out data, and the simulated-human limitation is a validity concern, not a circular derivation.

full rationale

The paper's derivation chain is self-contained and does not reduce to its own inputs. The strategy selector is trained with pseudo-labels defined by a rule-based comparison of AI and expert correctness (Eqs. 7-9), which is standard supervised learning to defer: the pseudo-labels are computed from ground-truth labels and expert labels, not from the selector's own output. The final evaluation in Section 4.2 is performed on held-out test data, so the reported accuracy-coverage curves are real generalization results rather than fitted predictions. The decoupled concept bottleneck backbone is adopted from external prior work (Zhang et al., 2024a), not from the authors' own prior claims, and the only self-citation (He et al., 2024) is a non-load-bearing related-work citation. The simulated expert with instance-independent label noise is a limitation on external validity, because real human errors are likely correlated with AI errors, but this is an experimental assumption and not a circular step: the experiments do demonstrate the claimed behavior under the stated simulation. No equation or argument in the paper defines a predicted quantity in terms of itself, and no fitted parameter is relabeled as a prediction. Therefore no specific circular step can be identified, and the appropriate score is 0.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on two modeling assumptions (concept availability and noise-simulated experts) plus a design rule for pseudo-labels. The free parameters λ and α control the trade-offs but their values are not reported. No new entities are postulated.

free parameters (2)
  • λ (human-effort cost weight) = not reported
    Appears in Eq. 5 to trade off accuracy against human involvement; selection procedure and values are omitted.
  • α (classification loss weight) = not reported
    In Eq. 9, balances strategy supervision against final classification loss; no value or tuning details are given.
assumptions (4)
  • domain assumption Human-annotated concepts are available for every training input and capture task-relevant semantics.
    DeCoDe's strategy selector and label predictor operate on explicit concept vectors c_exp; without concept labels the framework cannot be trained. The authors acknowledge this limitation in Section 6.
  • domain assumption Expert behavior is well modeled by input-independent label noise.
    All human baselines and training targets are generated by flipping labels at fixed rates (Section 4.1), which assumes human mistakes are uncorrelated with instance difficulty and with AI errors. This is load-bearing for the complementarity results.
  • domain assumption The decoupled representation (explicit plus implicit) from DCBM is sufficient for classification after JS alignment.
    The paper adopts DCBM's decomposition from Zhang et al. [2024a] without re-deriving it; if the implicit channel leaks or the JS constraint fails, concept-based gating loses its semantic grounding.
  • ad hoc to paper Correctness-based pseudo-labels provide a valid supervision signal for the gating network.
    The paper introduces the rule in Eqs. 7-8 without theoretical justification; if the heuristic is a poor proxy for optimal routing, the gating network will learn suboptimal strategies.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DeCoDe: Defer-and-Complement Decision-Making via Decoupled Concept Bottleneck Models." pith.science (2026). https://pith.science/paper/JTWLMC2S

@misc{pith2026250519220,
  author       = {Pith},
  title        = {Pith review of: DeCoDe: Defer-and-Complement Decision-Making via Decoupled Concept Bottleneck Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JTWLMC2S}},
  note         = {Machine review of arXiv:2505.19220}
}
read the original abstract

In human-AI collaboration, a central challenge is deciding whether the AI should handle a task, be deferred to a human expert, or be addressed through collaborative effort. Existing Learning to Defer approaches typically make binary choices between AI and humans, neglecting their complementary strengths. They also lack interpretability, a critical property in high-stakes scenarios where users must understand and, if necessary, correct the model's reasoning. To overcome these limitations, we propose Defer-and-Complement Decision-Making via Decoupled Concept Bottleneck Models (DeCoDe), a concept-driven framework for human-AI collaboration. DeCoDe makes strategy decisions based on human-interpretable concept representations, enhancing transparency throughout the decision process. It supports three flexible modes: autonomous AI prediction, deferral to humans, and human-AI collaborative complementarity, selected via a gating network that takes concept-level inputs and is trained using a novel surrogate loss that balances accuracy and human effort. This approach enables instance-specific, interpretable, and adaptive human-AI collaboration. Experiments on real-world datasets demonstrate that DeCoDe significantly outperforms AI-only, human-only, and traditional deferral baselines, while maintaining strong robustness and interpretability even under noisy expert annotations.

Figures

Figures reproduced from arXiv: 2505.19220 by the authors.

Figure 1
Figure 1. Comparison between traditional Learning to Defer (left) and our proposed De￾CoDe framework (right). While conventional deferral methods make binary AI–human decisions without providing explanations or enabling collaboration, DeCoDe leverages interpretable concept representations to support flexible strategy selection—autonomous AI, human-only, and human-AI complementarity—guided by a concept-driven gating network. W… view at source ↗
Figure 2
Figure 2. Overall architecture of the DeCoDe framework. The model is built upon a concept bottleneck structure, where explicit concepts are extracted from input images to form intermediate representations. These representations are used for downstream task prediction and strategy selection, enabling the model to adaptively choose among three decision modes: autonomous AI prediction, deferral to humans, and AI-human collaborat… view at source ↗
Figure 3
Figure 3. System accuracy as a function of human participation ratio for [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Effect of concept-level intervention on downstream predictions. The figure illustrates how manually modifying intermediate concepts can alter the model’s final prediction. The left shows the input image, the middle shows predicted and intervened concept values, and the…
Figure 5
Figure 5. Figure 5: Analysis of concept-based strategy selection in [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 32 canonical work pages

  1. [1]

    Does the whole exceed its parts? the effect of ai explanations on complementary team performance

    Gagan Bansal, Tongshuang Wu, Joyce Zhou, Raymond Fok, Besmira Nushi, Ece Kamar, Marco Tulio Ribeiro, and Daniel Weld. Does the whole exceed its parts? the effect of ai explanations on complementary team performance. In Proceedings of the 2021 CHI conference on human factors in computing systems, pages 1--16, 2021

    2021

  2. [2]

    Artificial intelligence for breast cancer screening in mammography (ai-stream): preliminary analysis of a prospective multicenter cohort study

    Yun-Woo Chang, Jung Kyu Ryu, Jin Kyung An, Nami Choi, Young Mi Park, Kyung Hee Ko, and Kyunghwa Han. Artificial intelligence for breast cancer screening in mammography (ai-stream): preliminary analysis of a prospective multicenter cohort study. Nature Communications, 16 0 (1): 0 2248, 2025

    work page 2025

  3. [3]

    When are two lists better than one?: Benefits and harms in joint decision-making

    Kate Donahue, Sreenivas Gollapudi, and Kostas Kollias. When are two lists better than one?: Benefits and harms in joint decision-making. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 10030--10038, 2024

    work page 2024

  4. [4]

    Enhancing the reliability and accuracy of ai-enabled diagnosis via complementarity-driven deferral to clinicians

    Krishnamurthy Dvijotham, Jim Winkens, Melih Barsbey, Sumedh Ghaisas, Robert Stanforth, Nick Pawlowski, Patricia Strachan, Zahra Ahmed, Shekoofeh Azizi, Yoram Bachrach, et al. Enhancing the reliability and accuracy of ai-enabled diagnosis via complementarity-driven deferral to clinicians. Nature Medicine, 29 0 (7): 0 1814--1820, 2023

    work page 2023

  5. [5]

    9 Hill Brendan 4 Karthik Ravishankar 2 Kunicki Katrina 4 Li Shuai 9 Makalic Enes 9 12 Nguyen Tuong L

    Helen ML Frazer, Carlos A Pe \ n a-Solorzano, Chun Fung Kwok, Michael S Elliott, Yuanhong Chen, Chong Wang, BRAIx Team Al-Qershi Osamah 9 Fox Samantha K. 9 Hill Brendan 4 Karthik Ravishankar 2 Kunicki Katrina 4 Li Shuai 9 Makalic Enes 9 12 Nguyen Tuong L. 9 Ralalage Prabhathi Basnayake 1 13 Schmidt Daniel 9 12 Weideman Prue C. 9, Jocelyn F Lippey, John L ...

    work page 2024

  6. [6]

    A taxonomy for human-llm interaction modes: An initial exploration

    Jie Gao, Simret Araya Gebreegziabher, Kenny Tsu Wei Choo, Toby Jia-Jun Li, Simon Tangi Perrault, and Thomas W Malone. A taxonomy for human-llm interaction modes: An initial exploration. In Extended Abstracts of the CHI Conference on Human Factors in Computing Systems, pages 1--11, 2024

    work page 2024

  7. [7]

    How human--ai feedback loops alter human perceptual, emotional and social judgements

    Moshe Glickman and Tali Sharot. How human--ai feedback loops alter human perceptual, emotional and social judgements. Nature Human Behaviour, 9 0 (2): 0 345--359, 2025

    work page 2025

  8. [8]

    Optimam mammography image database: a large-scale resource of mammography images and clinical data

    Mark D Halling-Brown, Lucy M Warren, Dominic Ward, Emma Lewis, Alistair Mackenzie, Matthew G Wallis, Louise S Wilkinson, Rosalind M Given-Wilson, Rita McAvinchey, and Kenneth C Young. Optimam mammography image database: a large-scale resource of mammography images and clinical data. Radiology: Artificial Intelligence, 3 0 (1): 0 e200103, 2020

    work page 2020

Show all 41 references
  1. [9]

    Addressing leakage in concept bottleneck models

    Marton Havasi, Sonali Parbhoo, and Finale Doshi-Velez. Addressing leakage in concept bottleneck models. In Advances in Neural Information Processing Systems, pages 23386--23397, 2022

    2022

  2. [10]

    Enhancing llm reasoning with multi-path collaborative reactive and reflection agents, 2024

    Chengbo He, Bochao Zou, Xin Li, Jiansheng Chen, Junliang Xing, and Huimin Ma. Enhancing llm reasoning with multi-path collaborative reactive and reflection agents, 2024. URL https://arxiv.org/abs/2501.00430

    2024 arXiv

  3. [11]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770--778, 2016

    2016

  4. [12]

    u hl, Michael V \

    Patrick Hemmer, Max Schemmer, Niklas K \"u hl, Michael V \"o ssing, and Gerhard Satzger. Complementarity in human-ai collaboration: Concept, sources, and evidence, 2024. URL https://arxiv.org/abs/2404.00029

    2024 arXiv

  5. [13]

    Artificial intelligence, human intelligence and hybrid intelligence based on mutual augmentation

    Mohammad Hossein Jarrahi, Christoph Lutz, and Gemma Newlands. Artificial intelligence, human intelligence and hybrid intelligence based on mutual augmentation. Big Data & Society, 9 0 (2): 0 20539517221142824, 2022

    2022

  6. [14]

    Who needs explanation and when? juggling explainable ai and user epistemic uncertainty

    Jinglu Jiang, Surinder Kahai, and Ming Yang. Who needs explanation and when? juggling explainable ai and user epistemic uncertainty. International Journal of Human-Computer Studies, 165: 0 102839, 2022

    2022

  7. [15]

    Seven-point checklist and skin lesion classification using multitask multimodal neural nets

    Jeremy Kawahara, Sara Daneshvar, Giuseppe Argenziano, and Ghassan Hamarneh. Seven-point checklist and skin lesion classification using multitask multimodal neural nets. IEEE journal of biomedical and health informatics, 23 0 (2): 0 538--546, 2018

    2018

  8. [16]

    Is disentanglement all you need? comparing concept-based & disentanglement approaches, 2021

    Dmitry Kazhdan, Botty Dimanov, Helena Andres Terre, Mateja Jamnik, Pietro Li \`o , and Adrian Weller. Is disentanglement all you need? comparing concept-based & disentanglement approaches, 2021. URL arXiv preprint arXiv:2104.06917

    2021 arXiv

  9. [17]

    Concept bottleneck models

    Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. Concept bottleneck models. In International conference on machine learning, pages 5338--5348. PMLR, 2020

    2020

  10. [18]

    Deep learning face attributes in the wild

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730--3738, 2015

    2015

  11. [19]

    Predict responsibly: improving fairness and accuracy by learning to defer

    David Madras, Toni Pitassi, and Richard Zemel. Predict responsibly: improving fairness and accuracy by learning to defer. In Advances in neural information processing systems, volume 31, 2018

    2018

  12. [20]

    A multimodal and fully automated system for prediction of pathological complete response to neoadjuvant chemotherapy in breast cancer

    Ning Mao, Yi Dai, Heng Zhou, Fan Lin, Tiantian Zheng, Ziyin Li, Ping Yang, Feng Zhao, Qin Li, Weiwei Wang, et al. A multimodal and fully automated system for prediction of pathological complete response to neoadjuvant chemotherapy in breast cancer. Science Advances, 11 0 (18):...

    2025

  13. [21]

    Using AI uncertainty quantification to improve human decision-making

    Laura Marusich, Jonathan Bakdash, Yan Zhou, and Murat Kantarcioglu. Using AI uncertainty quantification to improve human decision-making. In Forty-first International Conference on Machine Learning, 2024. URL https://openreview.net/forum?id=fowZNENcVJ

    2024

  14. [22]

    Consistent estimators for learning to defer to an expert

    Hussein Mozannar and David Sontag. Consistent estimators for learning to defer to an expert. In International conference on machine learning, pages 7076--7087. PMLR, 2020

    2020

  15. [23]

    Who should predict? exact algorithms for learning to defer to humans

    Hussein Mozannar, Hunter Lang, Dennis Wei, Prasanna Sattigeri, Subhro Das, and David Sontag. Who should predict? exact algorithms for learning to defer to humans. In International conference on artificial intelligence and statistics, pages 10520--10545. PMLR, 2023 a

    2023

  16. [24]

    Effective human-ai teams via learned natural language rules and onboarding

    Hussein Mozannar, Jimin Lee, Dennis Wei, Prasanna Sattigeri, Subhro Das, and David Sontag. Effective human-ai teams via learned natural language rules and onboarding. In Advances in Neural Information Processing Systems, pages 30466--30498, 2023 b

    2023

  17. [25]

    Post-hoc estimators for learning to defer to an expert

    Harikrishna Narasimhan, Wittawat Jitkrittum, Aditya K Menon, Ankit Rawat, and Sanjiv Kumar. Post-hoc estimators for learning to defer to an expert. In Advances in Neural Information Processing Systems, volume 35, pages 29292--29304, 2022

    2022

  18. [26]

    Nguyen, Thanh-Toan Do, and Gustavo Carneiro

    Cuong C. Nguyen, Thanh-Toan Do, and Gustavo Carneiro. Probabilistic learning to defer: Handling missing expert annotations and controlling workload distribution. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=z...

    2025

  19. [27]

    Influencing human--ai interaction by priming beliefs about ai can increase perceived trustworthiness, empathy and effectiveness

    Pat Pataranutaporn, Ruby Liu, Ed Finn, and Pattie Maes. Influencing human--ai interaction by priming beliefs about ai can increase perceived trustworthiness, empathy and effectiveness. Nature Machine Intelligence, 5 0 (10): 0 1076--1086, 2023

    2023

  20. [28]

    The algorithmic automation problem: Prediction, triage, and human effort, 2019

    Maithra Raghu, Katy Blumer, Greg Corrado, Jon Kleinberg, Ziad Obermeyer, and Sendhil Mullainathan. The algorithmic automation problem: Prediction, triage, and human effort, 2019. URL https://arxiv.org/abs/1903.12220

    2019 arXiv

  21. [29]

    Tree-based leakage inspection and control in concept bottleneck models, 2024

    Angelos Ragkousis and Sonali Parbhoo. Tree-based leakage inspection and control in concept bottleneck models, 2024. URL arXiv preprint arXiv:2410.06352

    2024 arXiv

  22. [30]

    Human-algorithmic interaction using a large language model-augmented artificial intelligence clinical decision support system

    Niroop Channa Rajashekar, Yeo Eun Shin, Yuan Pu, Sunny Chung, Kisung You, Mauro Giuffre, Colleen E Chan, Theo Saarinen, Allen Hsiao, Jasjeet Sekhon, et al. Human-algorithmic interaction using a large language model-augmented artificial intelligence clinical decision support sy...

    2024

  23. [31]

    Rethinking the inception architecture for computer vision

    Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818--2826, 2016

    2016

  24. [32]

    Calibrated learning to defer with one-vs-all classifiers

    Rajeev Verma and Eric Nalisnick. Calibrated learning to defer with one-vs-all classifiers. In International Conference on Machine Learning, pages 22184--22202. PMLR, 2022

    2022

  25. [33]

    Learning to defer to multiple experts: Consistent surrogate losses, confidence calibration, and conformal ensembles

    Rajeev Verma, Daniel Barrej \'o n, and Eric Nalisnick. Learning to defer to multiple experts: Consistent surrogate losses, confidence calibration, and conformal ensembles. In International Conference on Artificial Intelligence and Statistics, pages 11415--11434. PMLR, 2023

    2023

  26. [34]

    Exploiting human-ai dependence for learning to defer

    Zixi Wei, Yuzhou Cao, and Lei Feng. Exploiting human-ai dependence for learning to defer. In Forty-first International Conference on Machine Learning, 2024

    2024

  27. [35]

    Caltech-ucsd birds 200

    Peter Welinder, Steve Branson, Takeshi Mita, Catherine Wah, Florian Schroff, Serge Belongie, and Pietro Perona. Caltech-ucsd birds 200. Technical report, California Institute of Technology, 2010

    2010

  28. [36]

    Learning to complement humans

    Bryan Wilder, Eric Horvitz, and Ece Kamar. Learning to complement humans. In Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence, pages 1526--1533, 2021

    2021

  29. [37]

    Yu Wu, Yansong Li, Zeyu Dong, Nitya Sathyavageeswaran, and Anand D. Sarwate. Learning to help in multi-class settings. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=NCgTbt2j1F

    2025

  30. [38]

    Post-hoc concept bottleneck models, 2022

    Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models, 2022. URL arXiv preprint arXiv:2205.15480

    2022 arXiv

  31. [39]

    The decoupling concept bottleneck model

    Rui Zhang, Xingbo Du, Junchi Yan, and Shihua Zhang. The decoupling concept bottleneck model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024 a

    2024

  32. [40]

    Learning to complement and to defer to multiple users

    Zheng Zhang, Wenjie Ai, Kevin Wells, David Rosewarne, Thanh-Toan Do, and Gustavo Carneiro. Learning to complement and to defer to multiple users. In European Conference on Computer Vision, pages 144--162. Springer, 2024 b

    2024

  33. [41]

    Coverage-constrained human-ai cooperation with multiple experts, 2024 c

    Zheng Zhang, Cuong Nguyen, Kevin Wells, Thanh-Toan Do, David Rosewarne, and Gustavo Carneiro. Coverage-constrained human-ai cooperation with multiple experts, 2024 c . URL https://arxiv.org/abs/2411.11976

    2024 arXiv

Pith tools

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