The reviewed record of science sign in
Pith

arxiv: 2605.28165 · v1 · pith:BZSRIMRM · submitted 2026-05-27 · cs.LG

Unification and Optimization of Robust Supervised Learning

Jonas Hanselle , Valentin Margraf , Clemens Damke , Eyke H\"ullermeier This is my paper

Reviewed by Pith2026-06-29 14:01 UTCgrok-4.3pith:BZSRIMRMopen to challenge →

classification cs.LG
keywords robust supervised learningunified frameworkhyperparameter optimizationdistribution shiftlabel noisevicinal risk minimizationMixup
0
0 comments X

The pith

Robust learning methods for different failure modes unify into a design space where joint hyperparameter optimization matches the best single baseline across benchmarks.

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

The paper organizes a broad class of robust alternatives to empirical risk minimization along three design axes and decomposes them into four sequential stages: reference distribution enrichment, input-space perturbation, label-space perturbation, and sample-level aggregation. Each stage admits a choice among pessimistic, neutral, or optimistic stances. This decomposition produces a unified design space in which the configuration choices become hyperparameters that can be optimized jointly. Across tabular, image, and reward modeling benchmarks the resulting procedure competes with the strongest single-method baseline in each setting, supplying a default that does not require advance identification of the dominant failure mode.

Core claim

Existing robust methods can be decomposed into the stages of reference distribution enrichment, input-space perturbation, label-space perturbation, and sample-level aggregation, with each stage admitting pessimistic, neutral, or optimistic choices; joint hyperparameter optimization over this space yields performance competitive with the best single-method baseline on tabular, image, and reward modeling tasks.

What carries the argument

Decomposition of robust learning into sequential stages (reference distribution enrichment, input-space perturbation, label-space perturbation, sample-level aggregation) each with a stance (pessimistic, neutral, or optimistic) that turns the choices into jointly optimizable hyperparameters.

If this is right

  • A single optimization run can serve as a default when the practitioner does not know which failure mode dominates.
  • Robustness components can be mixed across stages rather than selected as an isolated technique.
  • The design space makes systematic comparison of stance and perturbation choices feasible without developing new methods from scratch.

Where Pith is reading between the lines

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

  • The staged view may make it easier to diagnose which stage contributes most to robustness on a given dataset.
  • The same decomposition could be tested on settings beyond supervised learning to see whether the same stage-and-stance structure appears.

Load-bearing premise

That the decomposition into these four stages and three stances captures the essential properties of existing robust methods without significant loss.

What would settle it

A benchmark suite in which the jointly optimized procedure underperforms the strongest specialized robust method by a clear margin on multiple tasks when the dominant failure mode is known in advance.

Figures

Figures reproduced from arXiv: 2605.28165 by Clemens Damke, Eyke H\"ullermeier, Jonas Hanselle, Valentin Margraf.

Figure 1
Figure 1. Figure 1: Two-moons classification with a gap in class 0 and a shifted test distribution (triangles). [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Our three-axis framework (i) enriches the empirical distribution [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Unified robust training procedure. A mini-batch passes through four stages: distributional [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Average normalized score per modality. Per-method composite across all (dataset, metric) in the OOD setting; higher is better. Joint (red) leads on Tabular and Reward Learning and is in the top group on Image. Baselines in grey. Shapley value analysis. Using the Faithful Shapley Interaction Index [33, 44], we analyze three representative components, each acting on a different part of the training procedure… view at source ↗
Figure 5
Figure 5. Figure 5: HPO budget on HH-RLHF: running-best OOD-validation performance vs. trial count. Joint [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Shapley interaction analysis on OOD test performance for reward learning. Each bar is a [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Exemplary demonstration of Mixup, KL-DFO and W-DRO in comparison to standard [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Training and validation learning curves for ERM, DRO, and DFO. ERM exhibits classical [PITH_FULL_IMAGE:figures/full_fig_p023_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Single-axis baselines: distribution of best-trial hyperparameter magnitudes per method, [PITH_FULL_IMAGE:figures/full_fig_p024_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Single-axis baselines: hyperparameter values selected by HPO, one row per (dataset, [PITH_FULL_IMAGE:figures/full_fig_p025_10.png] view at source ↗
read the original abstract

The literature has proposed various robust alternatives to empirical risk minimisation to address failure modes such as distribution shift, label noise and finite-sample degeneracies. Examples include distributionally robust optimization, label smoothing, vicinal risk minimization, and Mixup. However, such approaches are typically developed in isolation, forcing practitioners to commit a priori to a single failure mode even when the dominant mode for the task is unclear. To address this, we organize a broad class of existing methods along three common design axes and derive a tractable training procedure that decomposes robust learning into sequential stages (reference distribution enrichment, input-space perturbation, label-space perturbation, and sample-level aggregation), each with a choice of stance (pessimistic, neutral, or optimistic). This results in a unified design space in which joint hyperparameter optimization can compose and configure robustness strategies suited to the task at hand. Across tabular, image, and reward modeling benchmarks, joint hyperparameter optimization is competitive with the best single-method baseline in each setting, offering a reliable default for practitioners who do not know a priori which failure mode dominates their task.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The paper claims to unify a broad class of robust supervised learning methods (DRO, label smoothing, vicinal risk minimization, Mixup, etc.) by organizing them along three common design axes and deriving a tractable procedure that decomposes robust learning into four sequential stages—reference distribution enrichment, input-space perturbation, label-space perturbation, and sample-level aggregation—each admitting a pessimistic, neutral, or optimistic stance. Joint hyperparameter optimization over the resulting design space is then shown to be competitive with the best single-method baseline on tabular, image, and reward-modeling benchmarks, providing a default strategy when the dominant failure mode is unknown a priori.

Significance. If the staged decomposition faithfully reproduces the essential properties of the original methods, the work supplies a practical, task-adaptive alternative to committing to one robust technique in advance. The multi-domain empirical evaluation is a concrete strength, as is the explicit construction of a composable space that permits joint optimization rather than isolated method selection.

major comments (1)
  1. [Methodology (decomposition into stages)] The strongest empirical claim (joint HPO competitive with best single-method baselines) depends on the decomposition into the four sequential stages preserving the properties of the source methods. The manuscript must demonstrate, for at least one non-trivial example such as DRO or Mixup, that the staged procedure is equivalent (or explicitly note the approximation) to the original formulation; otherwise the benchmark comparisons rest on an unverified assumption that the composed procedures match the baselines they are measured against.
minor comments (2)
  1. [Section 3] Clarify the precise definition of the three design axes and how each stance choice maps onto existing hyperparameters; a small table or diagram would aid reproducibility.
  2. [Experiments] Report the number of random seeds, standard deviations, and any statistical tests supporting the claim that joint HPO is “competitive” with the per-setting best baseline.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on the decomposition's fidelity. We address the major comment below and agree that additional clarification is warranted.

read point-by-point responses

  1. Referee: [Methodology (decomposition into stages)] The strongest empirical claim (joint HPO competitive with best single-method baselines) depends on the decomposition into the four sequential stages preserving the properties of the source methods. The manuscript must demonstrate, for at least one non-trivial example such as DRO or Mixup, that the staged procedure is equivalent (or explicitly note the approximation) to the original formulation; otherwise the benchmark comparisons rest on an unverified assumption that the composed procedures match the baselines they are measured against.

    Authors: We agree that the empirical claims rest on the staged decomposition preserving key properties of the source methods. The manuscript derives the four stages directly from the common design axes identified across the literature (reference distribution, input perturbation, label perturbation, aggregation), with each stance (pessimistic/neutral/optimistic) chosen to recover standard formulations. In the revision we will add an explicit subsection (likely Section 3.3) that (i) maps DRO to reference-distribution enrichment under the pessimistic stance and (ii) maps Mixup to input-space perturbation under the neutral stance, showing either exact recovery of the original objective or the precise approximation introduced by sequential decomposition. This will be accompanied by a short proof sketch or counter-example where equivalence does not hold, ensuring the joint-HPO baselines are not compared against an unverified surrogate. revision: yes

Circularity Check

0 steps flagged

No circularity: unification is reorganization of existing methods

full rationale

The paper organizes existing robust methods (DRO, Mixup, label smoothing, etc.) along three design axes and decomposes them into sequential stages (reference distribution enrichment, input-space perturbation, label-space perturbation, sample-level aggregation) with stance choices. This produces a design space for joint HPO, but the decomposition is an explicit modeling choice rather than a reduction of any quantity to itself by definition or fitted parameter. No equations or claims reduce by construction to the paper's own inputs, and no load-bearing self-citations are invoked to justify uniqueness or the central procedure. The empirical claim (joint HPO competitive with single-method baselines) rests on benchmark comparisons that are external to the framework definition itself.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that existing methods map cleanly onto the three design axes and four stages; no new entities are postulated and no numerical parameters are fitted outside the joint optimization process itself.

free parameters (1)
  • stance choices per stage
    Pessimistic/neutral/optimistic stance for each of the four stages is selected via hyperparameter optimization rather than fixed in advance.
axioms (1)
  • domain assumption Existing robust methods can be organized along three common design axes and decomposed into the four sequential stages without loss of their core properties
    This organizing principle is invoked to derive the unified training procedure described in the abstract.

pith-pipeline@v0.9.1-grok · 5726 in / 1320 out tokens · 56978 ms · 2026-06-29T14:01:46.031862+00:00 · methodology

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

50 extracted references · 4 canonical work pages · 3 internal anchors

  1. [1]

    Akiba, T., Sano, S., Yanase, T., Ohta, T., and Koyama, M. (2019). Optuna: A next-generation hyperparameter optimization framework. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining

    2019

  2. [2]

    Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback

    Bai, Y ., Jones, A., Ndousse, K., Askell, A., Chen, A., DasSarma, N., Drain, D., Fort, S., Ganguli, D., Henighan, T., Joseph, N., Kadavath, S., Kernion, J., Conerly, T., Showk, S. E., Elhage, N., Hatfield-Dodds, Z., Hernandez, D., Hume, T., Johnston, S., Kravec, S., Lovitt, L., Nanda, N., Olsson, C., Amodei, D., Brown, T. B., Clark, J., McCandlish, S., Ol...

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  3. [3]

    Ben-Tal, A., Den Hertog, D., De Waegenaere, A., Melenberg, B., and Rennen, G. (2013). Robust Solutions of Optimization Problems Affected by Uncertain Probabilities.Management Science, 59(2):341–357

    2013

  4. [4]

    Blanchet, J., Li, J., Lin, S., and Zhang, X. (2025). Distributionally robust optimization and robust statistics.Statistical Science, 40(3):351–377

    2025

  5. [5]

    Bradley, R. A. and Terry, M. E. (1952). Rank Analysis of Incomplete Block Designs: I. The Method of Paired Comparisons.Biometrika, 39(3/4):324

    1952

  6. [6]

    Brier, G. W. (1950). Verification of forecasts expressed in terms of probability.Monthly Weather Review, 78(1):1–3

    1950

  7. [7]

    Carratino, L., Cissé, M., Jenatton, R., and Vert, J. (2022). On mixup regularization.J. Mach. Learn. Res., 23:325:1–325:31

    2022

  8. [8]

    Chapelle, O., Weston, J., Bottou, L., and Vapnik, V . (2000). Vicinal Risk Minimization. In Advances in Neural Information Processing Systems, volume 13. MIT Press

    2000

  9. [9]

    Choquet, G. (1954). Theory of capacities.Annales de l’Institut Fourier, 5:131–295

    1954

  10. [10]

    and Grote, T

    Freiesleben, T. and Grote, T. (2023). Beyond generalization: a theory of robustness in machine learning.Synthese, 202(4):109

    2023

  11. [11]

    Gao, R., Chen, X., and Kleywegt, A. J. (2024). Wasserstein distributionally robust optimization and variation regularization.Oper. Res

    2024

  12. [12]

    Gardner, J., Popovic, Z., and Schmidt, L. (2023). Benchmarking distribution shift in tabular data with tableshift.Advances in Neural Information Processing Systems

    2023

  13. [13]

    (2023).Bayesian Optimization

    Garnett, R. (2023).Bayesian Optimization. Cambridge University Press. Available for free at https://bayesoptbook.com/

    2023

  14. [14]

    J., Shlens, J., and Szegedy, C

    Goodfellow, I. J., Shlens, J., and Szegedy, C. (2015). Explaining and harnessing adversarial examples. In3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings

    2015

  15. [15]

    and Lopez-Paz, D

    Gulrajani, I. and Lopez-Paz, D. (2021). In search of lost domain generalization. In9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021

    2021

  16. [16]

    He, P., Gao, J., and Chen, W. (2023). Debertav3: Improving deberta using electra-style pre- training with gradient-disentangled embedding sharing. InThe Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net

    2023

  17. [17]

    D., Zoph, B., Gilmer, J., and Lakshminarayanan, B

    Hendrycks, D., Mu, N., Cubuk, E. D., Zoph, B., Gilmer, J., and Lakshminarayanan, B. (2020). Augmix: A simple method to improve robustness and uncertainty under data shift. InInternational Conference on Learning Representations

    2020

  18. [18]

    Hu, J., Mukherjee, D., and Paschalidis, I. C. (2025). Dro-augment framework: Robustness by synergizing wasserstein distributionally robust optimization and data augmentation.CoRR, abs/2506.17874. 11

    work page arXiv 2025

  19. [19]

    Hu, W., Niu, G., Sato, I., and Sugiyama, M. (2018). Does distributionally robust supervised learning give robust classifiers? InProceedings of the 35th International Conference on Machine Learning, pages 2029–2037

    2018

  20. [20]

    and Hong, L

    Hu, Z. and Hong, L. J. (2013). Kullback-Leibler divergence constrained distributionally robust optimization.Available at Optimization Online, 1(2):9

    2013

  21. [21]

    Jeong, D., Aggarwal, S., Robinson, J., Kumar, N., Spearot, A., and Park, D. S. (2023). Ex- haustive or exhausting? evidence on respondent fatigue in long surveys.Journal of Development Economics, 161:102992

    2023

  22. [22]

    and Xie, W

    Jiang, N. and Xie, W. (2024). Distributionally Favorable Optimization: A Framework for Data- Driven Decision-Making with Endogenous Outliers.SIAM Journal on Optimization, 34(1):419– 458

    2024

  23. [23]

    Kaufmann, T., Weng, P., Bengs, V ., and Hüllermeier, E. (2025). A survey of reinforcement learning from human feedback.Trans. Mach. Learn. Res

    2025

  24. [24]

    Kingma, D. P. and Ba, J. (2015). Adam: A method for stochastic optimization. In Bengio, Y . and LeCun, Y ., editors,3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings

    2015

  25. [25]

    Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. InAdvances in Neural Information Processing Systems 25: 26th Annual Conference on Neural Information Processing Systems 2012. Proceedings of a meeting held December 3-6, 2012, Lake Tahoe, Nevada, United States, pages 1106–1114

    2012

  26. [26]

    M., Nguyen, V

    Kuhn, D., Esfahani, P. M., Nguyen, V . A., and Shafieezadeh-Abadeh, S. (2019). Wasserstein Distributionally Robust Optimization: Theory and Applications in Machine Learning. In Netessine, S., Shier, D., and Greenberg, H. J., editors,Operations Research & Management Science in the Age of Analytics, pages 130–166. INFORMS

    2019

  27. [27]

    C., and Sidford, A

    Levy, D., Carmon, Y ., Duchi, J. C., and Sidford, A. (2020). Large-scale methods for distribu- tionally robust optimization.Advances in neural information processing systems, 33:8847–8860

    2020

  28. [28]

    Li, T., Beirami, A., Sanjabi, M., and Smith, V . (2021). Tilted empirical risk minimization. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net

    2021

  29. [29]

    and Hüllermeier, E

    Lienen, J. and Hüllermeier, E. (2021). From label smoothing to label relaxation. InThirty-Fifth AAAI Conference on Artificial Intelligence, AAAI 2021, Thirty-Third Conference on Innovative Applications of Artificial Intelligence, IAAI 2021, The Eleventh Symposium on Educational Advances in Artificial Intelligence, EAAI 2021, Virtual Event, February 2-9, 2...

    2021

  30. [30]

    Liu, Z., Luo, P., Wang, X., and Tang, X. (2015). Deep learning face attributes in the wild. In 2015 IEEE International Conference on Computer Vision, ICCV 2015, Santiago, Chile, December 7-13, 2015, pages 3730–3738. IEEE Computer Society

    2015

  31. [31]

    Madry, A., Makelov, A., Schmidt, L., Tsipras, D., and Vladu, A. (2018). Towards deep learning models resistant to adversarial attacks. InInternational Conference on Learning Representations

    2018

  32. [32]

    Müller, R., Kornblith, S., and Hinton, G. E. (2019). When does label smoothing help? In Advances in Neural Information Processing Systems, volume 32

    2019

  33. [33]

    Muschalik, M., Baniecki, H., Fumagalli, F., Kolpaczki, P., Hammer, B., and Hüllermeier, E. (2024). shapiq: Shapley interactions for machine learning. InAdvances in Neural Information Processing Systems 38: Annual Conference on Neural Information Processing Systems 2024, NeurIPS 2024, Vancouver, BC, Canada, December 10 - 15, 2024

    2024

  34. [34]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I. (2021). Learning transferable visual models from natural language supervision. InProceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, Proceedings ...

    2021

  35. [35]

    J., and Hassani, H

    Robey, A., Chamon, L., Pappas, G. J., and Hassani, H. (2022). Probabilistically robust learning: Balancing average and worst-case performance. In Chaudhuri, K., Jegelka, S., Song, L., Szepesvari, C., Niu, G., and Sabato, S., editors,Proceedings of the 39th International Conference on Machine Learning, volume 162 ofProceedings of Machine Learning Research,...

    2022

  36. [36]

    Robey, A., Chamon, L. F. O., Pappas, G. J., Hassani, H., and Ribeiro, A. (2021). Adversarial robustness with semi-infinite constrained learning. InAdvances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 6198–6215

    2021

  37. [37]

    Rockafellar, R. T. and Uryasev, S. (2000). Optimization of conditional value-at-risk.The Journal of Risk, 2(3):21–41

    2000

  38. [38]

    W., Hashimoto, T

    Sagawa, S., Koh, P. W., Hashimoto, T. B., and Liang, P. (2020). Distributionally robust neural networks. InInternational Conference on Learning Representations

    2020

  39. [39]

    L., Bouabid, S., and Muandet, K

    Singh, A., Chau, S. L., Bouabid, S., and Muandet, K. (2024). Domain generalisation via imprecise learning. InProceedings of the 41st International Conference on Machine Learning, volume 235 ofICML’24, pages 45544–45570, Vienna, Austria. JMLR.org

    2024

  40. [40]

    Sinha, A., Namkoong, H., and Duchi, J. (2018). Certifying some distributional robustness with principled adversarial training. InInternational Conference on Learning Representations

    2018

  41. [41]

    and Bottou, L

    Slowik, A. and Bottou, L. (2022). On distributionally robust optimization and data rebalancing. In Camps-Valls, G., Ruiz, F. J. R., and Valera, I., editors,International Conference on Artificial Intelligence and Statistics, AISTATS 2022, 28-30 March 2022, Virtual Event, Proceedings of Machine Learning Research, pages 1283–1297. PMLR

    2022

  42. [42]

    Stiennon, N., Ouyang, L., Wu, J., Ziegler, D., Lowe, R., V oss, C., Radford, A., Amodei, D., and Christiano, P. F. (2020). Learning to summarize with human feedback. InAdvances in Neural Information Processing Systems, volume 33, pages 3008–3021. Curran Associates, Inc

    2020

  43. [43]

    Szegedy, C., Vanhoucke, V ., Ioffe, S., Shlens, J., and Wojna, Z. (2016). Rethinking the inception architecture for computer vision. In2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 2818–2826. IEEE Computer Society

    2016

  44. [44]

    Tsai, C., Yeh, C., and Ravikumar, P. (2023). Faith-shap: The faithful shapley interaction index. J. Mach. Learn. Res., 24:94:1–94:42

    2023

  45. [45]

    Learning Credal Ensembles via Distributionally Robust Optimization

    Wang, K., Faza, G. A., Cuzzolin, F., Chau, S. L., Moens, D., and Hallez, H. (2026). Learning Credal Ensembles via Distributionally Robust Optimization.arXiv:2602.08470 [cs.LG]

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  46. [46]

    Wang, Y ., Zou, D., Yi, J., Bailey, J., Ma, X., and Gu, Q. (2020). Improving adversarial robustness requires revisiting misclassified examples. In8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020

    2020

  47. [47]

    J., Yoo, Y ., and Choe, J

    Yun, S., Han, D., Chun, S., Oh, S. J., Yoo, Y ., and Choe, J. (2019). Cutmix: Regularization strategy to train strong classifiers with localizable features. In2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 6022–6031. IEEE

    2019

  48. [48]

    N., and Lopez-Paz, D

    Zhang, H., Cissé, M., Dauphin, Y . N., and Lopez-Paz, D. (2018). Mixup: Beyond empirical risk minimization. In6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Conference Track Proceedings. OpenReview.net

    2018

  49. [49]

    P., Ghaoui, L

    Zhang, H., Yu, Y ., Jiao, J., Xing, E. P., Ghaoui, L. E., and Jordan, M. I. (2019). Theoretically principled trade-off between robustness and accuracy. InProceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, Proceedings of Machine Learning Research, pages 7472–7482. PMLR

    2019

  50. [50]

    Fine-Tuning Language Models from Human Preferences

    Ziegler, D. M., Stiennon, N., Wu, J., Brown, T. B., Radford, A., Amodei, D., Christiano, P. F., and Irving, G. (2019). Fine-tuning language models from human preferences.CoRR, abs/1909.08593. 13 A Extended Background In supervised learning, we consider an input space X , an output space Y and we assume there exists an unknown distribution P ⋆ ∈ P(X × Y) ,...

    work page internal anchor Pith review Pith/arXiv arXiv 2019