Pith. sign in

REVIEW 4 major objections 5 minor 89 references

ACE shows that forcing an ensemble to disagree confidently on the right unlabeled inputs can recover multiple valid generalizations from completely correlated data.

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

2026-08-04 21:25 UTC pith:CEGUO3C5

load-bearing objection ACE is a genuine new method for the underspecification problem, and the paper is honest about where it breaks; the headline result is conditional on a user-supplied mix-rate lower bound, but it deserves serious peer review. the 4 major comments →

arxiv 2509.07955 v1 pith:CEGUO3C5 submitted 2025-09-09 cs.LG cs.AI

ACE and Diverse Generalization via Selective Disagreement

Oliver Daniels , Stuart Armstrong , Alexandre Maranh\~ao , Mahirah Fairuz Rahman , Benjamin M. Marlin , Rebecca Gorman This is my paper
classification cs.LG cs.AI
keywords spurious correlationsunderspecificationdiverse generalizationensemble learningself-trainingconcept extrapolationmix ratemeasurement tampering detection
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.

ACE addresses the case where spurious correlations are complete: on the labeled source distribution, every candidate concept is perfectly correlated with the label, so the correct generalization is underspecified. The paper trains an ensemble of classifiers that all fit the labeled data but are forced to disagree, confidently and selectively, on unlabeled examples from the target distribution. The central claim is that this selective-disagreement loss recovers multiple human-interpretable concepts, matches or outperforms existing diverse-generalization methods on complete-correlation benchmarks, and continues to perform on incomplete correlations. If correct, ACE gives practitioners a way to surface candidate generalizations from data that would otherwise support only one, with a label-free validation-loss criterion for choosing among them.

Core claim

A top-k pseudo-label loss on unlabeled target data turns an ensemble's default diversity into distinct concepts. For each disagreement pattern, ACE selects the k inputs the ensemble is most confident fit that pattern and trains each head to match it confidently. The hardened pseudo-labels push decision boundaries toward low-density regions and suppress artificial disagreement on aligned inputs. The paper proves the loss is proper whenever each group's mix-rate lower bound does not exceed its true frequency, while DivDis and D-BAT are proper at only one mix rate each. Across six complete-correlation benchmarks and a measurement-tampering task, the recovered hypotheses match meaningful human c

What carries the argument

The mechanism is the ACE top-k pseudo-label loss. On unlabeled target inputs, the ensemble computes, for each disagreement group, the probability that every head's predicted label matches that group's pattern; it then builds a mask selecting the top k_g = ⌊N_T r̄_g⌋ inputs per group (or, in the group-aggregated variant, the top k (input, group) pairs overall) and trains the heads to match the group's hardened labels on exactly those inputs. This bootstraps default ensemble diversity into confident, selective disagreement, implements low-density separation through an entropy-like penalty, and keeps joint training stable. The user-set mix-rate lower bound r̄_g is the knob that encodes prior kn

Load-bearing premise

ACE assumes that each disagreement group occurs in the target distribution at least as often as the user-set lower bound r̄_g, so that the top-k masks select inputs where the concepts genuinely disagree; if the bound overshoots the true rate, the pseudo-labels push the ensemble to manufacture disagreement on correlation-aligned examples and the method's advantage collapses.

What would settle it

Run ACE on a complete-correlation benchmark (e.g., CIFAR-MNIST or Waterbirds-CC) with a target distribution whose true mix rate is known to be far below the configured lower bound, say r=0.02 with r̄=0.5. The paper's proper-scoring analysis predicts a collapse in multi-label accuracy; if the ensemble still recovers both concepts at high accuracy, the claimed dependence on the lower bound is falsified.

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

If this is right

  • On the complete-correlation benchmarks, ACE matches or outperforms DivDis, D-BAT, and ERM when the configured mix-rate lower bound is at or below the true disagreement frequency; ACE alone remains reliable across a range of mix rates rather than at a single one.
  • On Waterbirds with incomplete spurious correlations, ACE 0.5 reaches ground-truth accuracy comparable to DivDis while achieving higher spurious-attribute and worst-group accuracy.
  • Total validation loss on held-out source and target subsets is a label-free model-selection signal; for ACE it correlates almost perfectly with test error, whereas source loss alone is anti-correlated.
  • In measurement tampering detection, ACE distinguishes real from fake positives without using untrusted measurements, beating DivDis and the two baseline detectors and coming within about one AUROC point of the top probing baseline on Diamond in a Vault.
  • Mix-rate scheduling and a group-aggregated loss extend ACE to imbalanced disagreement groups by gradually increasing the enforced disagreement level and pooling group probabilities before top-k selection.

Where Pith is reading between the lines

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

  • If the low-density-separation intuition holds beyond the tested benchmarks, ACE's selective top-k loss could be chained with representation-level domain adaptation to handle target shifts that make even correlation-aligned inputs hard; the paper lists domain adaptation as future work, and the loss itself does not address that shift.
  • The validation-loss plateau could become an automatic, scale-free recipe for choosing the mix-rate lower bound on new datasets if the ad-hoc 20th-percentile threshold is replaced by a change-point detector; the paper's own results show the plateau is visible but the threshold choice remains manual.
  • The paper defines the loss for F concepts and binary labels but evaluates only two-head models, so the multi-concept regime is an untested extension; the disagreement-group combinatorics grow as 2^F, and the top-k selection would need to share a fixed budget across exponentially many groups.
  • In alignment applications, ACE suggests a route to detecting measurement tampering that does not rely on an ensemble of correlated measurements: if one head genuinely isolates the intended concept, disagreement between heads can flag tampering even when no failed attempt leaves a visible trace.

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

4 major / 5 minor

Summary. The paper studies the problem of learning under complete spurious correlation, where multiple concepts are perfectly correlated on the source distribution but come apart on an unlabeled target distribution. The authors propose ACE, which trains an ensemble of classifiers to disagree confidently and selectively on the unlabeled target inputs most likely to belong to disagreement groups. The objective uses top-k pseudo-labels with user-specified lower bounds on group mix rates. The paper reports experiments across six complete-correlation benchmarks, an incomplete-correlation study on Waterbirds, a measurement-tampering-detection case study, ablations of mix-rate scheduling and group-loss aggregation, and an analysis of validation-loss-based model selection. The central empirical claim is that ACE matches or outperforms existing diverse-generalization methods (DivDis, D-BAT, ERM) when the mix-rate lower bound is valid and reasonably tight, while also performing competitively in the MTD setting.

Significance. If the claimed results hold, ACE is a meaningful contribution to the under-specified generalization literature. The paper's strengths include a multi-dataset evaluation with six datasets, multiple random seeds, careful ablations, and an interesting application to measurement tampering detection. The proper-scoring analysis, although idealized, provides a useful way to compare disagreement-based losses. The authors are also unusually explicit about limitations in §6.1, including sensitivity to the mix-rate lower bound and the ad hoc nature of the threshold in §4.3. However, the headline claims in the abstract are stated unconditionally, while the evidence is conditional on the user supplying correct lower bounds. The proposed model-selection procedure is not evaluated without oracle information, and the incomplete-correlation claim rests on a single dataset. These gaps are significant but, in my view, addressable within the scope of a revision.

major comments (4)
  1. [§4.1, Fig. 2; §6.1] The central claim that ACE 'matches or outperforms existing methods' on complete-spurious-correlation benchmarks is conditional on the user-supplied lower bound r̄_g being close to and not exceeding the true target group rates. Fig. 2 shows that ACE 0.5 degrades sharply at target mix rate 0.1, and ACE 0.1 degrades as the mix rate increases. Section 6.1 explicitly admits this sensitivity. The proposed mix-rate inference in §4.3 uses a 20th-percentile threshold 'chosen in an ad-hoc manor using oracle information about the mix rate and test performance.' Thus, the abstract's unconditional wording and the claim of 'principled unsupervised model selection' are not supported as stated. The revision should qualify the benchmark claim and provide a non-oracle evaluation of the selection procedure.
  2. [App. B.3, Prop. 3; Eq. (3)–(5)] The proper-scoring proof for ACE-TopK assumes inputs are sampled uniformly from a discrete set and that the true distribution q 'is able to perfectly sort instances.' This does not apply to the actual algorithm, where the masks in Eq. (4) are computed from the current imperfect network probabilities. If those masks include correlation-aligned instances, Eq. (3) trains both heads to confidently disagree on them, which is exactly the artificial-disagreement failure mode acknowledged in §6.1. As written, Prop. 3 establishes a property of the idealized population loss, not of the stochastic coordinate-wise optimization used in the experiments. The paper should either state this limitation explicitly or provide an analysis/experiment measuring mask precision under realistic training dynamics.
  3. [§4.2, Table 1] The abstract claims that ACE 'remain[s] robust to incomplete spurious correlations,' but the only incomplete-correlation experiment is Waterbirds (Table 1). No incomplete-correlation versions of CelebA, Multi-NLI, or other benchmarks are presented. A single dataset is a thin basis for a headline robustness claim. The revision should either add at least one additional incomplete-correlation benchmark or clearly scope the claim to Waterbirds.
  4. [§4.3 and App. G] The validation-loss model-selection criterion is computed with the same auxiliary loss and the same mix-rate lower bound used during training. Since this is essentially the training objective evaluated on held-out data, the strong correlation between validation loss and test error (Figs. 10–15) does not by itself establish a principled, independent selection signal. Moreover, for D-BAT the paper switches to source validation accuracy because the disagreement loss overfits. The comparison of model-selection behavior across methods is therefore uneven. The revision should clarify what independence claim is being made and, ideally, evaluate selection on datasets not used to choose the threshold.
minor comments (5)
  1. [§2.1] The sentence 'the disagreement groups are c1 = (0,1) and c2 = (0,1)' appears to contain a typo; the second group should presumably be (1,0).
  2. [Table 1 caption] The caption contains an incomplete sentence: 'Oracle test data is used to With the single label...' The intended wording should be corrected.
  3. [§5 and §6] The conclusion calls ACE 'state-of-the-art' on complete-spurious-correlation benchmarks, but the MTD case study in Fig. 4 shows that Probing for Evidence of Tamper remains the best method. The wording should distinguish the complete-correlation benchmark results from the MTD results to avoid overclaiming.
  4. [§3.3 and App. B.3] The group-aggregated loss and the per-group top-k loss are different algorithms, but the proper-scoring proof in App. B.3 only covers the per-group version. The relationship between the two variants and the theoretical claim should be clarified.
  5. [General] The paper states that code will be made available pending acceptance. Since reproducibility is a stated goal, including the code or a public anonymized repository in the revision would strengthen the contribution.

Circularity Check

0 steps flagged

No load-bearing circularity: ACE's self-training loss is the method, not a hidden re-use of the target; benchmark claims are checked against external labels, and the main self-referential elements (validation loss model selection, oracle-tuned 20th-percentile threshold) are disclosed limitations rather than forced reductions.

full rationale

The derivation chain is self-contained. ACE's target loss (Eq. 3) trains heads against top-k pseudo-labels selected from their own disagreement; this is the proposed bootstrapping algorithm, not a circular inference. The proper-scoring result (Prop. 3, App. B.3) is explicitly conditional on the user-supplied lower bound r̄_g not exceeding the true group mix rate r_g, and it is not used to prove that ACE recovers the intended concepts; it only characterizes the loss. The central benchmark claim is evaluated against external target labels (e.g., Fig. 2 uses oracle head assignment but external ground-truth and spurious labels for accuracy), so the main empirical assertion does not reduce to the training objective. The model-selection criterion in Sec. 3.3/App. G does use the same top-k validation loss as the training loss, which makes it self-referential as an unsupervised selection signal, but the paper separately checks that validation loss correlates with test error (Fig. 10), providing an external check. Similarly, the 20th-percentile threshold for mix-rate lower-bound inference (Sec. 4.3) is admittedly chosen 'in an ad-hoc manor using oracle information about the mix rate and test performance' and flagged in Sec. 6.1 as a limitation; it is not presented as a derived prediction. There is one self-citation ([4], Armstrong) used only for the 'model splintering' terminology, and it is not load-bearing. No step reduces a claimed result to its inputs by construction, so the paper is not significantly circular.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

ACE rests on a user-supplied lower bound on the frequency of concept disagreement in the target distribution, on the cluster assumption that human concepts occupy low-density regions, and on treating the target as unlabeled. No new physical entities are introduced; the main free parameters are the mix-rate bound, auxiliary loss weights, schedule, and validation threshold.

free parameters (5)
  • Global mix rate lower bound r̄ = 0.1 and 0.5 in experiments
    User-set assumption about minimum target disagreement frequency; drives top-k sizes and, if violated, breaks the proper-scoring property.
  • Group mix rate lower bounds r̄_g = r̄ / |G'| for balanced datasets, else k = ceil(N_T · r̄) aggregated
    Determines how many target instances per disagreement group are pseudo-labeled; central to the selective disagreement loss.
  • Auxiliary loss weights λ_g = per dataset/method table, e.g. TopK 1.0-8.0, DivDis 1.0-64.0, D-BAT 0.01-0.3
    Tuned using total validation loss and a default weight of 1; strongly affects the balance between source and target losses.
  • Mix-rate schedule T = linear ramp from 0 to r̄ over training epochs
    Incremental schedule for the mix-rate lower bound, borrowed from pseudo-labeling practice; improves low-mix-rate performance.
  • Validation-loss percentile threshold = 20th percentile
    Used for mix-rate lower-bound selection; the paper states this threshold was chosen in an ad-hoc manner using oracle information.
axioms (6)
  • domain assumption Source distribution has complete spurious correlation: ρ_pS(Ci,Cj) = 1 for all concept pairs.
    Section 2; this is the problem setting, not derived.
  • domain assumption Target distribution is available only as unlabeled inputs and contains some disagreement groups.
    Section 2; concept extrapolation assumes unlabeled pT(x) and no target labels.
  • ad hoc to paper Each disagreement group appears at a rate at least r̄_g, i.e. the user lower bound is valid.
    Section 3.4 and Appendix B.3; if violated, top-k masks select non-disagreement instances and ACE's loss is not proper.
  • domain assumption Human concepts are separated by low-density regions in representation space (cluster assumption / natural abstractions hypothesis).
    Section 3.4; used to argue that hardened pseudo-labels recover disentangled concepts rather than arbitrary boundaries.
  • standard math Proper-scoring analysis assumes inputs uniformly sampled from a discrete set and that the true distribution can perfectly sort top-k instances.
    Appendix B.3; the proof is simplified and depends on these conditions.
  • domain assumption In the measurement tampering application, the untrusted distribution can be treated as fully unlabeled despite untrusted measurements being available.
    Section 5; this is the mapping of MTD to concept extrapolation.

pith-pipeline@v1.3.0-alltime-deepseek · 18632 in / 9906 out tokens · 114468 ms · 2026-08-04T21:25:34.830107+00:00 · methodology

0 comments
read the original abstract

Deep neural networks are notoriously sensitive to spurious correlations - where a model learns a shortcut that fails out-of-distribution. Existing work on spurious correlations has often focused on incomplete correlations,leveraging access to labeled instances that break the correlation. But in cases where the spurious correlations are complete, the correct generalization is fundamentally \textit{underspecified}. To resolve this underspecification, we propose learning a set of concepts that are consistent with training data but make distinct predictions on a subset of novel unlabeled inputs. Using a self-training approach that encourages \textit{confident} and \textit{selective} disagreement, our method ACE matches or outperforms existing methods on a suite of complete-spurious correlation benchmarks, while remaining robust to incomplete spurious correlations. ACE is also more configurable than prior approaches, allowing for straight-forward encoding of prior knowledge and principled unsupervised model selection. In an early application to language-model alignment, we find that ACE achieves competitive performance on the measurement tampering detection benchmark \textit{without} access to untrusted measurements. While still subject to important limitations, ACE represents significant progress towards overcoming underspecification.

Figures

Figures reproduced from arXiv: 2509.07955 by Alexandre Maranh\~ao, Benjamin M. Marlin, Mahirah Fairuz Rahman, Oliver Daniels, Rebecca Gorman, Stuart Armstrong.

Figure 2
Figure 2. Figure 2: Aggregated Multi-Label Accuracy Across Mix Rates Box plots of multi-label (ground truth and spurious) accuracy gain over ERM of diverse generalization methods across mix rates, aver￾aged over 6 datasets with complete spurious correlations and 3 random seeds for each configuration. Heads are assigned to labels using oracle test data. ACE 0.1 outperforms all other methods at a mix rate of 0.1, but suffers an… view at source ↗
Figure 3
Figure 3. Figure 3: Validation Loss Density Estimate for CIFAR-MNIST Validation loss density estimate for CIFAR-MNIST as a function of mix rate lower bound and smallest group lower bound. Selecting the highest lower bound subject to being within the 20th percentile of the validation loss density estimate provides a good signal for mix rate lower bound. 4.3 Mix Rate Lower Bound Selection Thus far we have assumed fixed mix rate… view at source ↗
Figure 4
Figure 4. Figure 4: Measurement Tampering Detection Real vs Fake AUROC Mean and standard deviation of AUROC scores for detecting real versus fake positive outcomes across different methods, averaged over 8 random seeds. ACE outperforms DivDis, and both outperform baseline methods, but Probing for Evidence of Tamper remains state-of-the-art. show validation loss degradation provides a useful signal for lower bound selection, o… view at source ↗
Figure 5
Figure 5. Figure 5: Accuracy Metrics Across Mix Rates Mean (line) and standard deviation (shaded area) of group truth and spurious accuracy metrics across mix rates (taken over 3 seeds). Results largely align with the main results. Most notably, DivDis achieves higher ground-truth accuracy than spurious accuracy on Waterbirds and CelebA, even though both generalizations are equally valid and the spurious features are prefered… view at source ↗
Figure 6
Figure 6. Figure 6: Worst-Group Accuracy Across Mix Rates Mean (line) and standard deviation (shaded area) of worst-group accuracy across mix rates (taken over 3 seeds). Again results largely conform to the main body findings, though TopK 0.5 approximately pareto dominates all other methods on the realistic datasets (Waterbirds, CelebA, and MultiNLI) [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: (Scheduling Ablation) Aggregated Multi-Label Accuracy Across Mix Rates Box plots of multi-label accuracy gain over ERM, averaged over 6 datasets and 3 random seeds, for TopK 0.1 and 0.5 with and without mix rate scheduling Mix rate scheduling substantially improves the performance of TopK 0.5 on low mix rates, but provides only a modest improvement to TopK 0.1. strong signal for mix rate lower bound for th… view at source ↗
Figure 8
Figure 8. Figure 8: (Scheduling Ablation) Accuracy Metrics Across Mix Rates 24 [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: (Combined Group Loss Ablation) Accuracy Metrics Across Mix Rates The Per-group loss lower bounds perform better on class-balanced datasets (top row) while the group-aggregated loss performs better on class-in-balanced datasets (bottom row) 25 [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Correlation Coefficient Between Validation Loss and Test Error Violin plots of the correlation coefficient between validation loss and test error, averaged over 5 datasets and 3 random seeds for each method (we exclude toy grid because we do not use early stopping for this dataset). We find that for all methods, validation loss is a good proxy for test error when the implicit (or explicit) assumptions of … view at source ↗
Figure 11
Figure 11. Figure 11: Correlation Coefficient Between Source Validation Loss and Test Error Violin plots of the correlation coefficient between validation loss and test error, averaged over 5 datasets and 3 random seeds for each method (we exclude toy grid because we do not use early stopping for this dataset). We find that for D-BAT, source validation loss is a good proxy for test error when the mix rate, but negatively corre… view at source ↗
Figure 12
Figure 12. Figure 12: Correlation between Validation Loss and Multi-Label Test Error for ACE TopK [PITH_FULL_IMAGE:figures/full_fig_p027_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Correlation between Validation Loss and Multi-Label Test Accuracy for DivDis and [PITH_FULL_IMAGE:figures/full_fig_p028_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Correlation between Source Validation Loss and Multi-Label Test Error for ACE TopK [PITH_FULL_IMAGE:figures/full_fig_p029_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Correlation between Source Validation Loss and Multi-Label Test Error for DivDis and [PITH_FULL_IMAGE:figures/full_fig_p030_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Validation Loss Density Estimate [PITH_FULL_IMAGE:figures/full_fig_p031_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Test Accuracy Density Estimate 31 [PITH_FULL_IMAGE:figures/full_fig_p031_17.png] view at source ↗

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

89 extracted references · 55 canonical work pages · 1 internal anchor

  1. [1]

    Falk, and Ioannis Mitliagkas

    Isabela Albuquerque, João Monteiro, Tiago H. Falk, and Ioannis Mitliagkas. Adversarial target-invariant representation learning for domain generalization.CoRR, abs/1911.00804, 2019

    Pith/arXiv arXiv 1911

  2. [2]

    Concrete problems in ai safety, 2016

    Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mané. Concrete problems in ai safety, 2016

    2016

  3. [3]

    Invariant risk mini- mization, 2020

    Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk mini- mization, 2020. 9

    2020

  4. [4]

    Model splintering: moving from one imperfect model to another

    Stuart Armstrong. Model splintering: moving from one imperfect model to another. https://www.alignmentforum.org/posts/k54rgSg7GcjtXnMHX/ model-splintering-moving-from-one-imperfect-model-to-another, 2020

    2020

  5. [5]

    A theory of learning from different domains.Machine Learning, 79:151–175, 2010

    Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando C Pereira, and Jen- nifer Wortman Vaughan. A theory of learning from different domains.Machine Learning, 79:151–175, 2010

    2010

  6. [6]

    Analysis of representa- tions for domain adaptation

    Shai Ben-David, John Blitzer, Koby Crammer, and Fernando Pereira. Analysis of representa- tions for domain adaptation. InProceedings of the 20th International Conference on Neural Information Processing Systems, NIPS’06, page 137–144, Cambridge, MA, USA, 2006. MIT Press

    2006

  7. [7]

    Pythia: A suite for analyzing large language models across training and scaling, 2023

    Stella Biderman, Hailey Schoelkopf, Quentin Anthony, Herbie Bradley, Kyle O’Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, Aviya Skowron, Lintang Sutawika, and Oskar van der Wal. Pythia: A suite for analyzing large language models across training and scaling, 2023

    2023

  8. [8]

    Measuring progress on scalable oversight for large language models.arXiv preprint arXiv:2211.03540, 2022

    Samuel R Bowman, Jeeyoon Hyun, Ethan Perez, Edwin Chen, Craig Pettit, Scott Heiner, Kamil˙e Lukoši¯ut˙e, Amanda Askell, Andy Jones, Anna Chen, et al. Measuring progress on scalable oversight for large language models.arXiv preprint arXiv:2211.03540, 2022

    Pith/arXiv arXiv 2022

  9. [9]

    Weak-to-strong gener- alization: Eliciting strong capabilities with weak supervision.arXiv preprint arXiv:2312.09390, 2023

    Collin Burns, Pavel Izmailov, Jan Hendrik Kirchner, Bowen Baker, Leo Gao, Leopold Aschen- brenner, Yining Chen, Adrien Ecoffet, Manas Joglekar, Jan Leike, et al. Weak-to-strong gener- alization: Eliciting strong capabilities with weak supervision.arXiv preprint arXiv:2312.09390, 2023

    Pith/arXiv arXiv 2023

  10. [10]

    Learning imbalanced datasets with label-distribution-aware margin loss, 2019

    Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss, 2019

    2019

  11. [11]

    Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning

    Paola Cascante-Bonilla, Fuwen Tan, Yanjun Qi, and Vicente Ordonez. Curriculum labeling: Revisiting pseudo-labeling for semi-supervised learning. InAAAI Conference on Artificial Intelligence, 2020

    2020

  12. [12]

    Semi-supervised classification by low density separation

    Olivier Chapelle and Alexander Zien. Semi-supervised classification by low density separation. In Robert G. Cowell and Zoubin Ghahramani, editors,Proceedings of the Tenth International Workshop on Artificial Intelligence and Statistics, volume R5 ofProceedings of Machine Learning Research, pages 57–64. PMLR, 06–08 Jan 2005. Reissued by PMLR on 30 March 2021

    2005

  13. [13]

    Self-training avoids using spurious features under domain shift.Advances in Neural Information Processing Systems, 33:21061– 21071, 2020

    Yining Chen, Colin Wei, Ananya Kumar, and Tengyu Ma. Self-training avoids using spurious features under domain shift.Advances in Neural Information Processing Systems, 33:21061– 21071, 2020

    work page 2020

  14. [14]

    Invariant causal mechanisms through distribution matching, 2022

    Mathieu Chevalley, Charlotte Bunne, Andreas Krause, and Stefan Bauer. Invariant causal mechanisms through distribution matching, 2022

    work page 2022

  15. [15]

    Deep reinforcement learning from human preferences.Advances in neural information processing systems, 30, 2017

    Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences.Advances in neural information processing systems, 30, 2017

    2017

  16. [16]

    Reward model ensembles help mitigate overoptimization, 2024

    Thomas Coste, Usman Anwar, Robert Kirk, and David Krueger. Reward model ensembles help mitigate overoptimization, 2024

    work page 2024

  17. [17]

    Environment inference for invariant learning, 2021

    Elliot Creager, Jörn-Henrik Jacobsen, and Richard Zemel. Environment inference for invariant learning, 2021

    work page 2021

  18. [18]

    Class-balanced loss based on effective number of samples, 2019

    Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples, 2019

    work page 2019

  19. [19]

    Un- derspecification presents challenges for credibility in modern machine learning.Journal of Machine Learning Research, 23(226):1–61, 2022

    Alexander D’Amour, Katherine Heller, Dan Moldovan, Ben Adlam, Babak Alipanahi, Alex Beutel, Christina Chen, Jonathan Deaton, Jacob Eisenstein, Matthew D Hoffman, et al. Un- derspecification presents challenges for credibility in modern machine learning.Journal of Machine Learning Research, 23(226):1–61, 2022. 10

    work page 2022

  20. [20]

    The mnist database of handwritten digit images for machine learning research.IEEE Signal Processing Magazine, 29(6):141–142, 2012

    Li Deng. The mnist database of handwritten digit images for machine learning research.IEEE Signal Processing Magazine, 29(6):141–142, 2012

    2012

  21. [21]

    Bert: Pre-training of deep bidirectional transformers for language understanding, 2019

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding, 2019

    2019

  22. [22]

    Distributionally robust losses for latent covariate mixtures, 2022

    John Duchi, Tatsunori Hashimoto, and Hongseok Namkoong. Distributionally robust losses for latent covariate mixtures, 2022

    work page 2022

  23. [23]

    Self-ensembling for visual domain adaptation, 2018

    Geoffrey French, Michal Mackiewicz, and Mark Fisher. Self-ensembling for visual domain adaptation, 2018

    work page 2018

  24. [24]

    Domain-adversarial training of neural networks

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario March, and Victor Lempitsky. Domain-adversarial training of neural networks. Journal of machine learning research, 17(59):1–35, 2016

    2016

  25. [25]

    Wichmann

    Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A. Wichmann. Shortcut learning in deep neural networks.Nature Machine Intelligence, 2(11):665–673, November 2020

    2020

  26. [26]

    Wichmann, and Wieland Brendel

    Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A. Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness, 2022

    2022

  27. [27]

    Strictly proper scoring rules, prediction, and estimation

    Tilmann Gneiting and Adrian E Raftery. Strictly proper scoring rules, prediction, and estimation. Journal of the American statistical Association, 102(477):359–378, 2007

    2007

  28. [28]

    Semi-supervised learning by entropy minimization

    Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. In Proceedings of the 17th International Conference on Neural Information Processing Systems, NIPS’04, page 529–536, Cambridge, MA, USA, 2004. MIT Press

    work page 2004

  29. [29]

    Ai control: Improving safety despite intentional subversion, 2024

    Ryan Greenblatt, Buck Shlegeris, Kshitij Sachan, and Fabien Roger. Ai control: Improving safety despite intentional subversion, 2024

    work page 2024

  30. [30]

    Borgwardt, Malte J

    Arthur Gretton, Karsten M. Borgwardt, Malte J. Rasch, Bernhard Schölkopf, and Alexander Smola. A kernel two-sample test.Journal of Machine Learning Research, 13(25):723–773, 2012

    2012

  31. [31]

    In search of lost domain generalization, 2020

    Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization, 2020

    work page 2020

  32. [32]

    The unreasonable effectiveness of easy training data for hard tasks, 2024

    Peter Hase, Mohit Bansal, Peter Clark, and Sarah Wiegreffe. The unreasonable effectiveness of easy training data for hard tasks, 2024

    work page 2024

  33. [33]

    Deep residual learning for image recognition, 2015

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition, 2015

    2015

  34. [34]

    Benchmarking neural network robustness to common corruptions and perturbations, 2019

    Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations, 2019

    2019

  35. [35]

    On the founda- tions of shortcut learning

    Katherine Hermann, Hossein Mobahi, Thomas FEL, and Michael Curtis Mozer. On the founda- tions of shortcut learning. InThe Twelfth International Conference on Learning Representations, 2024

    work page 2024

  36. [36]

    Efros, and Trevor Darrell

    Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei A. Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation, 2017

    work page 2017

  37. [37]

    The platonic representation hypothesis, 2024

    Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola. The platonic representation hypothesis, 2024

    work page 2024

  38. [38]

    Simple data balancing achieves competitive worst-group-accuracy, 2022

    Badr Youbi Idrissi, Martin Arjovsky, Mohammad Pezeshki, and David Lopez-Paz. Simple data balancing achieves competitive worst-group-accuracy, 2022

    work page 2022

  39. [39]

    The class imbalance problem: Significance and strategies

    Nathalie Japkowicz. The class imbalance problem: Significance and strategies. InProc. of the Int’l Conf. on artificial intelligence, volume 56, pages 111–117, 2000. 11

    work page 2000

  40. [40]

    Last layer re-training is sufficient for robustness to spurious correlations, 2023

    Polina Kirichenko, Pavel Izmailov, and Andrew Gordon Wilson. Last layer re-training is sufficient for robustness to spurious correlations, 2023

    work page 2023

  41. [41]

    Neural network ensembles, cross validation and active learning

    Anders Krogh and Jesper Vedelsby. Neural network ensembles, cross validation and active learning. InProceedings of the 8th International Conference on Neural Information Processing Systems, NIPS’94, page 231–238, Cambridge, MA, USA, 1994. MIT Press

    work page 1994

  42. [42]

    Out-of-distribution generalization via risk extrapo- lation (rex), 2021

    David Krueger, Ethan Caballero, Joern-Henrik Jacobsen, Amy Zhang, Jonathan Binas, Dinghuai Zhang, Remi Le Priol, and Aaron Courville. Out-of-distribution generalization via risk extrapo- lation (rex), 2021

    work page 2021

  43. [43]

    Understanding self-training for gradual domain adaptation

    Ananya Kumar, Tengyu Ma, and Percy Liang. Understanding self-training for gradual domain adaptation. In Hal Daumé III and Aarti Singh, editors,Proceedings of the 37th International Conference on Machine Learning, volume 119 ofProceedings of Machine Learning Research, pages 5468–5479. PMLR, 13–18 Jul 2020

    work page 2020

  44. [44]

    Simple and scalable predictive uncertainty estimation using deep ensembles, 2017

    Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles, 2017

    work page 2017

  45. [45]

    A survey on out-of-distribution detection in nlp.arXiv preprint arXiv:2305.03236, 2023

    Hao Lang, Yinhe Zheng, Yixuan Li, Jian Sun, Fei Huang, and Yongbin Li. A survey on out-of-distribution detection in nlp.arXiv preprint arXiv:2305.03236, 2023

    Pith/arXiv arXiv 2023

  46. [46]

    Goal misgeneralization in deep reinforcement learning, 2023

    Lauro Langosco, Jack Koch, Lee Sharkey, Jacob Pfau, Laurent Orseau, and David Krueger. Goal misgeneralization in deep reinforcement learning, 2023

    work page 2023

  47. [47]

    Pseudo-label : The simple and efficient semi-supervised learning method for deep neural networks

    Dong-Hyun Lee. Pseudo-label : The simple and efficient semi-supervised learning method for deep neural networks. InICML 2013 Workshop: Challenges in Representation Learning (WREPL), 2013

    work page 2013

  48. [48]

    Learning debiased representation via disentangled feature augmentation, 2021

    Jungsoo Lee, Eungyeup Kim, Juyoung Lee, Jihyeon Lee, and Jaegul Choo. Learning debiased representation via disentangled feature augmentation, 2021

    work page 2021

  49. [49]

    Diversify and disambiguate: Learning from underspecified data, 2023

    Yoonho Lee, Huaxiu Yao, and Chelsea Finn. Diversify and disambiguate: Learning from underspecified data, 2023

    work page 2023

  50. [50]

    You are what you eat – ai alignment requires understanding how data shapes structure and generalisation, 2025

    Simon Pepin Lehalleur, Jesse Hoogland, Matthew Farrugia-Roberts, Susan Wei, Alexander Gi- etelink Oldenziel, George Wang, Liam Carroll, and Daniel Murfet. You are what you eat – ai alignment requires understanding how data shapes structure and generalisation, 2025

    work page 2025

  51. [51]

    Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C. Kot. Domain generalization with adversarial feature learning. In2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5400–5409, 2018

    work page 2018

  52. [52]

    Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn

    Evan Zheran Liu, Behzad Haghgoo, Annie S. Chen, Aditi Raghunathan, Pang Wei Koh, Shiori Sagawa, Percy Liang, and Chelsea Finn. Just train twice: Improving group robustness without training group information, 2021

    work page 2021

  53. [53]

    Deep learning face attributes in the wild

    Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaoou Tang. Deep learning face attributes in the wild. InProceedings of International Conference on Computer Vision (ICCV), December 2015

    2015

  54. [54]

    Learning transferable features with deep adaptation networks.CoRR, abs/1502.02791, 2015

    Mingsheng Long and Jianmin Wang. Learning transferable features with deep adaptation networks.CoRR, abs/1502.02791, 2015

    Pith/arXiv arXiv 2015

  55. [55]

    Mingsheng Long, Jianmin Wang, and Michael I. Jordan. Deep transfer learning with joint adaptation networks.CoRR, abs/1605.06636, 2016

    Pith/arXiv arXiv 2016

  56. [56]

    Domain adaptation: Learning bounds and algorithms.CoRR, abs/0902.3430, 2009

    Yishay Mansour, Mehryar Mohri, and Afshin Rostamizadeh. Domain adaptation: Learning bounds and algorithms.CoRR, abs/0902.3430, 2009

    Pith/arXiv arXiv 2009

  57. [57]

    Michaud, Yonatan Belinkov, David Bau, and Aaron Mueller

    Samuel Marks, Can Rager, Eric J. Michaud, Yonatan Belinkov, David Bau, and Aaron Mueller. Sparse feature circuits: Discovering and editing interpretable causal graphs in language models, 2024

    work page 2024

  58. [58]

    Learning from failure: Training debiased classifier from biased classifier, 2020

    Junhyun Nam, Hyuntak Cha, Sungsoo Ahn, Jaeho Lee, and Jinwoo Shin. Learning from failure: Training debiased classifier from biased classifier, 2020. 12

    work page 2020

  59. [59]

    Spread spurious attribute: Improving worst-group accuracy with spurious attribute estimation, 2022

    Junhyun Nam, Jaehyung Kim, Jaeho Lee, and Jinwoo Shin. Spread spurious attribute: Improving worst-group accuracy with spurious attribute estimation, 2022

    2022

  60. [60]

    Codegen: An open large language model for code with multi-turn program synthesis, 2023

    Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, and Caiming Xiong. Codegen: An open large language model for code with multi-turn program synthesis, 2023

    2023

  61. [61]

    Agree to disagree: Diversity through disagreement for better transferability, 2022

    Matteo Pagliardini, Martin Jaggi, François Fleuret, and Sai Praneeth Karimireddy. Agree to disagree: Diversity through disagreement for better transferability, 2022

    work page 2022

  62. [62]

    Ridge rider: Finding diverse solutions by following eigenvectors of the hessian, 2020

    Jack Parker-Holder, Luke Metz, Cinjon Resnick, Hengyuan Hu, Adam Lerer, Alistair Letcher, Alex Peysakhovich, Aldo Pacchiano, and Jakob Foerster. Ridge rider: Finding diverse solutions by following eigenvectors of the hessian, 2020

    work page 2020

  63. [63]

    Dice: Diversity in deep ensembles via conditional redundancy adversarial estimation, 2021

    Alexandre Rame and Matthieu Cord. Dice: Diversity in deep ensembles via conditional redundancy adversarial estimation, 2021

    work page 2021

  64. [64]

    why should i trust you?

    Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. "why should i trust you?": Explaining the predictions of any classifier, 2016

    2016

  65. [65]

    In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning, 2021

    Mamshad Nayeem Rizve, Kevin Duarte, Yogesh S Rawat, and Mubarak Shah. In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection framework for semi-supervised learning, 2021

    work page 2021

  66. [66]

    Coup probes: Catching catastrophes with probes trained off-policy.AI Alignment Forum, Nov 2023

    Fabien Roger. Coup probes: Catching catastrophes with probes trained off-policy.AI Alignment Forum, Nov 2023

    work page 2023

  67. [67]

    Benchmarks for detecting measurement tampering, 2023

    Fabien Roger, Ryan Greenblatt, Max Nadeau, Buck Shlegeris, and Nate Thomas. Benchmarks for detecting measurement tampering, 2023

    work page 2023

  68. [68]

    Hashimoto, and Percy Liang

    Shiori Sagawa, Pang Wei Koh, Tatsunori B. Hashimoto, and Percy Liang. Distributionally robust neural networks for group shifts: On the importance of regularization for worst-case generalization, 2020

    work page 2020

  69. [69]

    Facenet: A unified embedding for face recognition and clustering

    Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. In2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), page 815–823. IEEE, June 2015

    work page 2015

  70. [70]

    The pitfalls of simplicity bias in neural networks, 2020

    Harshay Shah, Kaustav Tamuly, Aditi Raghunathan, Prateek Jain, and Praneeth Netrapalli. The pitfalls of simplicity bias in neural networks, 2020

    work page 2020

  71. [71]

    Wasserstein distance guided representation learning for domain adaptation, 2018

    Jian Shen, Yanru Qu, Weinan Zhang, and Yong Yu. Wasserstein distance guided representation learning for domain adaptation, 2018

    work page 2018

  72. [72]

    A dirt-t approach to unsupervised domain adaptation.arXiv preprint arXiv:1802.08735, 2018

    Rui Shu, Hung H Bui, Hirokazu Narui, and Stefano Ermon. A dirt-t approach to unsupervised domain adaptation.arXiv preprint arXiv:1802.08735, 2018

    Pith/arXiv arXiv 2018

  73. [73]

    Diversity inducing information bottleneck in model ensembles, 2020

    Samarth Sinha, Homanga Bharadhwaj, Anirudh Goyal, Hugo Larochelle, Animesh Garg, and Florian Shkurti. Diversity inducing information bottleneck in model ensembles, 2020

    work page 2020

  74. [74]

    Return of Frustratingly Easy Domain Adaptation

    Baochen Sun, Jiashi Feng, and Kate Saenko. Return of frustratingly easy domain adaptation. CoRR, abs/1511.05547, 2015

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  75. [75]

    Deep coral: Correlation alignment for deep domain adaptation, 2016

    Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation, 2016

    work page 2016

  76. [76]

    Evading the simplicity bias: Training a diverse set of models discovers solutions with superior ood general- ization, 2022

    Damien Teney, Ehsan Abbasnejad, Simon Lucey, and Anton van den Hengel. Evading the simplicity bias: Training a diverse set of models discovers solutions with superior ood general- ization, 2022

    work page 2022

  77. [77]

    Overcoming simplicity bias in deep networks using a feature sieve

    Rishabh Tiwari and Pradeep Shenoy. Overcoming simplicity bias in deep networks using a feature sieve. InProceedings of the 40th International Conference on Machine Learning, ICML’23. JMLR.org, 2023. 13

    work page 2023

  78. [78]

    Adversarial discriminative domain adaptation, 2017

    Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation, 2017

    work page 2017

  79. [79]

    Camargo, and Ard A

    Guillermo Valle-Pérez, Chico Q. Camargo, and Ard A. Louis. Deep learning generalizes because the parameter-function map is biased towards simple functions, 2019

    work page 2019

  80. [80]

    Springer science & business media, 1999

    Vladimir Vapnik.The nature of statistical learning theory. Springer science & business media, 1999

    work page 1999

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.