Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Relating Misfit to Gain in Weak-to-Strong Generalization Beyond the Squared Loss

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read The paper proves a general misfit–gain law for weak-to-strong learning: the student's gain over the weak teacher is at least their misfit for any Bregman-divergence loss, and for nonconvex classes this holds up to a vanishing error when…

desk verdict A real extension of the misfit-gain framework to classification, with a genuinely new k-convex recipe—but the main theorem's proof has a gap that needs repair before the advertised O(sqrt(c/k)) bound is claimed. read the letter →

arxiv 2501.19105 v2 pith:L7AVAQQH submitted 2025-01-31 cs.LG math.PR

Abhijeet Mulgund , Chirag Pabbaraju This is my paper · ORCID
classification cs.LGmath.PR MSC 68Q3262B10
keywords weak-to-stronggeneralizationBregmandivergencecross-entropyKLmisfit-gaininequalityconvexcombinationsJensengapclassification
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

This paper extends a quantitative law of weak-to-strong generalization from squared-loss regression to any loss that is a Bregman divergence, which includes the cross-entropy used in classification. The law states that when the strong model class is convex and the target is realizable, the strong student's expected loss on the true target is at most the weak teacher's expected loss minus the divergence between student and teacher outputs: gain is at least misfit. Because practical strong classes are usually not convex, the paper proves the same inequality for a student that is a convex combination of $k$ strong functions, with an extra error term $O(\sqrt{c/k})$ that vanishes as $k$ grows. It also proposes a concrete recipe, minimizing the reversed KL divergence from the strong output to the weak labels over $k$ logistic heads, and reports that loss gain tracks KL misfit on NLP and vision benchmarks. The paper's own experiments show the guarantee depends on the target being representable in the strong class, which fails on real data and is restored only by redefining the target as the best strong model.

What carries the argument

The machinery is the Bregman divergence $D_\psi$ generated by a strictly convex $C^1$ function $\psi$, together with the generalized Pythagorean inequality for forward Bregman projections onto a closed convex set: for $z$ and any $x$ in the convex set $W$, $D_\psi(x,z) \geq D_\psi(x,P_W(z)) + D_\psi(P_W(z),z)$. Applied to the expected divergence $\mathbb{E}[D_\psi(\cdot, f_w \circ h_w)]$ over the convex class $F \circ h_s$, this is what turns near-minimal misfit on weak labels into a lower bound on loss gain. For the nonconvex case, the load-bearing device is the Jensen approximation gap of the generator, here the negative Shannon entropy, which lets a $k$-fold convex combination approximate any point of the convex hull in Bregman cost with error $O(\sum_i \rho_i/k)$, plus a Pinsker-type log-ratio bound; combining these gives the $O(\sqrt{c/k})$ slack in Theorem 4.3.

What would settle it

Direct test on the theorem: build a synthetic classification problem where realizability and the regularization condition hold, optimize the reversed-KL objective over $\operatorname{co}_k F$ to near-global optimality, and estimate all three expectations in Inequality (3) on fresh data; a violation larger than the $O(\sqrt{c/k})$ term would refute Theorem 4.3. A complementary test for the practical claim: on a real dataset where a strong model trained on true labels still has nontrivial loss, compute the same three terms for the true target; the authors report that the inequality fails in this setting, so any claimed bound for the true target must confront this failure.

Watch

Extended reading notes

Core claim

The central discovery is a misfit-gain inequality that holds beyond squared loss. Theorem 4.1 shows that if the strong class $F$ is convex and the target $g$ lies in $F \circ h_s$, then any student $f_s$ that nearly minimizes the expected Bregman divergence to the weak labels satisfies $\mathbb{E}[D_\psi(g(X), f_s(h_s(X)))] \leq \mathbb{E}[D_\psi(g(X), f_w(h_w(X)))] - \mathbb{E}[D_\psi(f_s(h_s(X)), f_w(h_w(X)))] + \epsilon$. Because binary cross-entropy is KL divergence plus an entropy constant, Corollary 4.2 converts this into a classification guarantee: training with the KL objective in the reverse direction provably reduces expected cross-entropy by at least the KL misfit. For nonconvex $F$, Theorem 4.3 proves the same inequality holds up to $O(\sqrt{c/k})$ when the student is a convex combination of $k$ functions from $F$, under a mild regularization condition; the slack comes from a Jensen-gap bound on the negative Shannon entropy generator.

Load-bearing premise

The load-bearing premise is realizability: the true target function must be exactly representable by some function in the strong model class, and the paper's own experiments (Section 5.5) show this fails on real datasets; the misfit–gain law is restored only when the target is redefined as the best strong model.

Editorial extensions

If this is right

  • For any regression or classification loss that is a Bregman divergence, a convex strong class yields a provable gain of at least the student–teacher misfit; the earlier squared-loss result is the special case $\psi(x)=x^2/2$.
  • For classification, minimizing the KL divergence with the strong output in the first argument and the weak labels in the second, over a convex combination of $k$ logistic heads, provably lowers cross-entropy on the true target once $k$ is large enough.
  • The quantities in the inequality are expectations that can be estimated from hold-out data, so the bound can be used to predict how much gain to expect from a given weak teacher before training the strong model.
  • No assumption is placed on the weak model; the requirements fall on the strong class, convexity or a mild regularization condition, and on realizability of the target.
  • Empirically, $k=100$ is enough for gain and misfit to track each other closely on NLP and vision benchmarks, and increasing $k$ shrinks the difference as the error term predicts.

Reading between the lines

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

  • Editorial inference: since the paper's own Section 5.5 shows realizability fails on real data, the practically relevant statement is about the best achievable strong model rather than the ground-truth target; a version of Theorem 4.3 stated for unrealizable targets would be needed to certify gains on actual benchmarks.
  • Editorial inference: the reversed-KL recipe is a projection in the probability simplex rather than in logit space; testing whether optimizing in the dual logit space with the same $k$-combination changes how quickly gain approaches misfit would separate the geometry of the loss from the optimizer's behavior.
  • Editorial inference: the Jensen-gap argument is specialized to KL in the nonconvex theorem; extending Lemma A.4 to other Bregman generators, such as the Itakura-Saito divergence, would show whether the $O(\sqrt{c/k})$ rate is a property of cross-entropy or an artifact of this proof technique.
  • Editorial inference: no accuracy-level theorem is claimed, but the experiments consistently show accuracy moving with loss; a testable extension is that the misfit–gain gap predicts accuracy improvement whenever the loss–accuracy curve is locally linear.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper extends the misfit-gain characterization of weak-to-strong generalization from squared-loss regression to losses induced by Bregman divergences, with a focus on classification via the KL/cross-entropy loss. For a convex strong class, Theorem 4.1 establishes that a near-minimizer of the Bregman misfit to the weak model achieves a target loss no larger than the weak model's loss minus the misfit, up to an arbitrarily small error. Corollary 4.2 specializes this to cross-entropy. For non-convex strong classes, Theorem 4.3 instead considers convex combinations of k functions from the strong class and claims the same gain-misfit inequality up to an O(sqrt(c/k)) error. The paper reports synthetic and real-data experiments supporting the qualitative and, to a degree, quantitative relationship, while Section 5.5 acknowledges that the realizability assumption fails on real datasets and only recovers quantitative agreement by evaluating losses against the best strong model rather than the ground-truth target.

Significance. If all claims were fully established, this would be a valuable conceptual generalization of the squared-loss theory in Charikar et al. (2024): it identifies the Bregman Pythagorean inequality as the core mechanism, extends the characterization to cross-entropy, and provides a concrete algorithmic recipe for weak-to-strong training with non-convex strong classes. The multi-class formulation and the Jensen-gap machinery are natural and potentially reusable. However, the paper's headline non-convex result, Theorem 4.3, rests on a proof step in Appendix A.2.2 that does not follow from the stated assumptions, so the advertised O(sqrt(c/k)) guarantee is not currently proven. The authors also deserve credit for explicitly testing and reporting the failure of realizability on real data, and for separating that issue from the structural inequality rather than hiding it.

major comments (4)
  1. [Section 4, Theorem 4.1 and Fact 3.4] The statement of Theorem 4.1 assumes only that F is a convex set of functions, but the proof invokes Fact 3.4 to obtain a unique Bregman projection onto F ∘ h_s. That fact requires the projected set to be closed as well as convex. Without closedness, existence of the projection and the equality between inf and min used in the proof are not automatic. The authors later acknowledge that uniqueness of the exact minimizer requires closedness, but the theorem statement and proof should either add a closedness hypothesis or prove attainment under the stated assumptions.
  2. [Appendix A.2.2, proof of Theorem A.2 after Lemma A.7] The central bound on |E[DKL(g||g_proj)] - E[DKL(g||f_s)]| is not justified. The proof uses the display E[ sqrt(2 DKL(f_s||g_proj)) / min_i(f_s,i, g_proj,i) ] <= E[ max_i rho_i * sqrt(c/(4k) max_i rho_i + eps) ]. This requires a pointwise upper bound on DKL(f_s||g_proj), but only the expectation E[DKL(f_s||g_proj)] is bounded, via the Jensen-gap argument. It also treats 1/min_i as bounded by max_i rho_i, whereas the stated regularization assumption controls only first moments of sup_{f in F} 1/f_i(X), not pointwise values. A correct argument would need a pointwise bound on M(X) := max_i sup_f 1/f_i(X), or a second-moment condition, and would then produce a factor such as E[M(X)^2]^{1/2} rather than the stated quantity. As written, the O(sqrt(c/k)) error term in Theorem 4.3 is not established.
  3. [Appendix A.3] The claimed generalization to infinite-support input distributions inherits the same defect. Assumption 2 in Theorem A.2 controls first moments of sup 1/f_i(X); the repaired proof described above would require E[M(X)^2] < infinity, which is not implied by the stated condition. Either the infinite-support claim should be made conditional on an additional second-moment or boundedness assumption, or a separate argument must be supplied.
  4. [Section 5.5 and Theorem 4.3, Assumption 1] The authors verify that realizability, Assumption 1 of Theorem 4.3, does not hold in their real-data experiments, and they recover quantitative agreement only by evaluating the weak and strong losses with respect to the best strong model rather than the ground-truth target. This should be stated more prominently as a limitation: Figures 2 and 3a-c do not test the quantitative inequality (3) for the actual target function, and the abstract's claim of experimental support should be qualified accordingly.
minor comments (5)
  1. [Section 4, proof sketch of Corollary 4.2] The text says 'Theorem 4.2' when referring to the corollary being proved; the same mislabeling appears in the paragraph after Corollary 4.2 ('Theorem 4.2 still applies to coF').
  2. [Appendix A.2.2, Lemma A.5] The notation 'a := (supp∈S(1/pi))^c_{i=1}' should use 'sup' rather than 'supp', and the definition of a_i as the coordinatewise supremum should be written explicitly.
  3. [Appendix A.2.2, Theorem A.2 statement] The Regularization assumption displays max( E[ sup_{f in F} 1/f(X) ], E[ sup_{f in F} 1/f(X)) ] ), where the second term should presumably be E[ sup 1/(1-f(X)) ] for binary classification; the parenthesis error makes the two displayed terms identical.
  4. [Throughout, equations (10), (11), and surrounding text] The square-root symbol is rendered as a placeholder '...' in several places (e.g., O(... c/k), O(... c/k)); the intended expression O(sqrt(c/k)) should be typeset correctly.
  5. [Appendix A.2.2, final paragraph] The word 'optmial' should be 'optimal'.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation found: the misfit-gain inequality is derived from independent convex-geometric bounds, not from its own conclusion.

full rationale

The central result, Theorem 4.3 / Theorem A.2, is derived from the Bregman Pythagorean inequality (Fact 3.4), a Jensen-gap bound for negative Shannon entropy (Lemma A.5), and a Pinsker-based Lipschitz bound (Lemma A.7). None of these inputs contains the target inequality as an assumption, and the target inequality is a consequence of these ingredients rather than an identity imposed by definition. No parameters are fitted to data in the derivation, and no empirical 'prediction' is a renamed fit: the gain (reduction in target cross-entropy) and the misfit (reverse KL divergence between the strong and weak outputs) are separately measured and then compared. The realizability assumption is stated explicitly and checked against data; Section 5.5 reports that it fails on real datasets and that the authors recover agreement only by redefining the reference target to be the best strong model. This is an honest limitation of the experiment, not a derivation that builds the conclusion into its premises. The only self-citation is Charikar et al. [12], which shares an author; it is used as a baseline and as a special case, not to justify the new theorem. The proof issue raised in Appendix A.2.2, concerning the pointwise use of an expectation bound, is a potential correctness gap rather than circularity: the inequality may be unproven, but it is not equivalent by construction to the theorem's inputs. Therefore no significant circularity is present.

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

The central theorem relies on standard Bregman geometry, a realizability assumption on the target, a regularity condition on the function class, and a finite-support simplification. No new physical or conceptual entities are introduced. The proof of Theorem 4.3 also implicitly relies on a pointwise control of inverse probabilities that is stronger than the stated expectation bound.

assumptions (5)
  • standard math Generalized Pythagorean inequality for Bregman divergences over closed convex sets (Fact 3.4).
    Load-bearing for Theorem 4.1; it provides the geometric inequality that turns Bregman projection into a misfit-gain bound.
  • domain assumption Realizability: the target g is in the strong class, g = f* composed with hs (Assumption 1 of Theorems 4.1 and 4.3).
    Without this, the projection of the weak model onto the strong class need not relate to the target, and the inequality has no proven analogue.
  • domain assumption Regularization: E[sup_{f in F} 1/f_i(X)] < infinity for all classes i (Assumption 2 of Theorem 4.3).
    Used to bound the Jensen gap and the inverse-probability term; the proof appears to require a pointwise version of this bound, not just the expectation.
  • domain assumption Input distributions have finite support (Section 3.1).
    Simplifies function spaces to finite dimension and guarantees a minimum probability mass pmin > 0, which is used in the proof of Theorem 4.1.
  • domain assumption Convexity of the strong class for Theorem 4.1.
    Needed for existence and uniqueness of the Bregman projection onto F composed with hs.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Relating Misfit to Gain in Weak-to-Strong Generalization Beyond the Squared Loss." pith.science (2026). https://pith.science/paper/L7AVAQQH

@misc{pith2026250119105,
  author       = {Pith},
  title        = {Pith review of: Relating Misfit to Gain in Weak-to-Strong Generalization Beyond the Squared Loss},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L7AVAQQH}},
  note         = {Machine review of arXiv:2501.19105}
}
abstract

The paradigm of weak-to-strong generalization constitutes the training of a strong AI model on data labeled by a weak AI model, with the goal that the strong model nevertheless outperforms its weak supervisor on the target task of interest. For the setting of real-valued regression with the squared loss, recent work quantitatively characterizes the gain in performance of the strong model over the weak model in terms of the misfit between the strong and weak model. We generalize such a characterization to learning tasks whose loss functions correspond to arbitrary Bregman divergences when the strong class is convex. This extends the misfit-based characterization of performance gain in weak-to-strong generalization to classification tasks, as the cross-entropy loss can be expressed in terms of a Bregman divergence. In most practical scenarios, however, the strong model class may not be convex. We therefore weaken this assumption and study weak-to-strong generalization for convex combinations of $k$ strong models in the strong class, in the concrete setting of classification. This allows us to obtain a similar misfit-based characterization of performance gain, upto an additional error term that vanishes as $k$ gets large. Our theoretical findings are supported by thorough experiments on synthetic as well as real-world datasets.

Figures

Figures reproduced from arXiv: 2501.19105 by the authors.

Figure 1
Figure 1. Synthetic data experiments. The Gain and Misfit closely track each other. For [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Weak model (gpt2) is trained once on true labels and thereafter fixed. Each strong model is trained on the weak labels (on a held-out set separate from that which weak model was trained on) for 10 random initializations of the k-convex combination of logistic regression heads; we plot the average test loss/misfit across these 10 runs, along with the standard deviations as the error bars. weights in the objective to … view at source ↗
Figure 3
Figure 3. (a), (b) Weak model (AlexNet) is trained once on true labels and thereafter fixed; we report numbers [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: For all datasets except ImageNet, we observe that the difference between misfit and gain decreases as [PITH_FULL_IMAGE:figures/full_fig_p022_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. On Weak-to-Strong Generalization and f-Divergence

    cs.LG 2025-06 conditional novelty 6.0 of 10

    Replacing cross-entropy with f-divergence losses in weak-to-strong generalization gives modest accuracy gains and improved label-noise tolerance, though the paper's theoretical equivalence result is constructed after ...

  2. On the Mechanisms of Weak-to-Strong Generalization: A Theoretical Perspective

    stat.ML 2025-05 conditional novelty 6.0 of 10

    In high-dimensional linear and one-step feature-learning models, a regularized student can outperform its teacher by fixing under-regularization, using better regularization structure, or retaining pretrained hard features.

Reference graph

Works this paper leans on

64 extracted references · 31 canonical work pages · cited by 2 Pith papers

  1. [1]

    Some new estimates of the ‘jensen gap’

    Shoshana Abramovich and Lars-Erik Persson. Some new estimates of the ‘jensen gap’. Journal of Inequalities and Applications , 2016(1), February 2016. ISSN 1029-242X. doi: 10.1186/s13660-016-0985-4. URL http://dx.doi.org/10.1186/s13660-016-0985-4

    doi:10.1186/s13660-016-0985-4 2016

  2. [2]

    Information geometry and its applications , volume 194

    Shun-ichi Amari. Information geometry and its applications , volume 194. Springer, 2016

    2016

  3. [3]

    Springer International Publishing, 2017

    Nihat Ay, J¨ urgen Jost, Hˆ ong Vˆ an Lˆ e, and Lorenz Schwachh¨ ofer.Information Geometry . Springer International Publishing, 2017. ISBN 9783319564784. doi: 10.1007/978-3-319-56478-4. URL http: //dx.doi.org/10.1007/978-3-319-56478-4

    doi:10.1007/978-3-319-56478-4 2017

  4. [4]

    Training a helpful and harmless assistant with reinforcement learning from human feedback

    Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862 , 2022

    arXiv 2022

  5. [5]

    Construction of best bregman approximations in reflexive banach spaces

    Heinz Bauschke and Patrick Combettes. Construction of best bregman approximations in reflexive banach spaces. Proceedings of the American Mathematical Society , 131(12):3757–3766, April 2003. ISSN 1088-6826. doi: 10.1090/s0002-9939-03-07050-3. URL http://dx.doi.org/10.1090/S0002-9939-03-07050-3

    work page doi:10.1090/s0002-9939-03-07050-3 2003

  6. [6]

    Bauschke, Jonathan M

    Heinz H. Bauschke, Jonathan M. Borwein, and Patrick L. Combettes. Bregman monotone optimization algo- rithms. SIAM Journal on Control and Optimization , 42(2):596–636, 2003. doi: 10.1137/S0363012902407120. URL https://doi.org/10.1137/S0363012902407120

    work page doi:10.1137/s0363012902407120 2003

  7. [7]

    Combining labeled and unlabeled data with co-training

    Avrim Blum and Tom Mitchell. Combining labeled and unlabeled data with co-training. In Proceedings of the eleventh annual conference on Computational learning theory , pages 92–100, 1998

    work page 1998

  8. [8]

    The relaxation method of finding the common point of convex sets and its application to the solution of problems in convex programming

    Lev M Bregman. The relaxation method of finding the common point of convex sets and its application to the solution of problems in convex programming. USSR computational mathematics and mathematical physics, 7(3):200–217, 1967

    work page 1967

Show all 64 references
  1. [9]

    Weak-to-strong generalization: Eliciting strong capabilities with weak supervision, 2023

    Collin Burns, Pavel Izmailov, Jan Hendrik Kirchner, Bowen Baker, Leo Gao, Leopold Aschenbrenner, Yining Chen, Adrien Ecoffet, Manas Joglekar, Jan Leike, Ilya Sutskever, and Jeff Wu. Weak-to-strong generalization: Eliciting strong capabilities with weak supervision, 2023. URL h...

    2023 arXiv

  2. [10]

    Redundant axioms in the definition of bregman functions

    Dan Butnariu, Charles Byrne, and Yair Censor. Redundant axioms in the definition of bregman functions. Journal of Convex Analysis , 10(1):245–254, 2003. URL https://www.heldermann-verlag.de/jca/jca10/ jca0313.pdf

    2003

  3. [11]

    Emerging properties in self-supervised vision transformers

    Mathilde Caron, Hugo Touvron, Ishan Misra, Herv´ e J´ egou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision , pages 9650–9660, 2021

    2021

  4. [12]

    Quantifying the gain in weak-to-strong generalization, 2024

    Moses Charikar, Chirag Pabbaraju, and Kirankumar Shiragur. Quantifying the gain in weak-to-strong generalization, 2024. URL https://arxiv.org/abs/2405.15116

    2024 arXiv

  5. [13]

    Convergence analysis of a proximal-like minimization algorithm using bregman functions

    Gong Chen and Marc Teboulle. Convergence analysis of a proximal-like minimization algorithm using bregman functions. SIAM Journal on Optimization , 3(3):538–543, 1993. doi: 10.1137/0803026. URL https://doi.org/10.1137/0803026

    1993 doi

  6. [14]

    Boolq: Exploring the surprising difficulty of natural yes/no questions

    Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions. arXiv preprint arXiv:1905.10044, 2019

    1905 arXiv

  7. [15]

    Schapire, and Yoram Singer

    Michael Collins, Robert E. Schapire, and Yoram Singer. Logistic regression, adaboost and bregman distances. Machine Learning, 48(1/3):253–285, 2002. ISSN 0885-6125. doi: 10.1023/a:1013912006537. URL http://dx.doi.org/10.1023/A:1013912006537

    2002 doi

  8. [16]

    Pac generalization bounds for co-training

    Sanjoy Dasgupta, Michael Littman, and David McAllester. Pac generalization bounds for co-training. Advances in neural information processing systems , 14, 2001

    2001

  9. [17]

    Inderjit S. Dhillon. Learning with bregman divergences. https://www.cs.utexas.edu/~inderjit/Talks/ bregtut.pdf, 2007. 11

    2007

  10. [18]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020

    2010 arXiv

  11. [19]

    Weak-to-strong generalization

    Adrien Ecoffet, Manas Joglekar, Jeffrey Wu, Jan Hendrik Kirchner, and Pavel Izmailov. Weak-to-strong generalization. https://github.com/openai/weak-to-strong, 2023

    2023

  12. [20]

    Ekeland and R

    I. Ekeland and R. Temam. Convex Analysis and Variational Problems . Classics in Applied Mathematics. Society for Industrial and Applied Mathematics, 1999. ISBN 9781611971088. URL https://books.google. com/books?id=OD7nXNEptZcC

    1999

  13. [21]

    Topics in convex optimisation

    Hamza Fawzi. Topics in convex optimisation. https://www.damtp.cam.ac.uk/user/hf323/L22-III-OPT/ lecture11.pdf, 2022

    2022

  14. [22]

    Red teaming language models to reduce harms: Methods, scaling behaviors, and lessons learned

    Deep Ganguli, Liane Lovitt, Jackson Kernion, Amanda Askell, Yuntao Bai, Saurav Kadavath, Ben Mann, Ethan Perez, Nicholas Schiefer, Kamal Ndousse, et al. Red teaming language models to reduce harms: Methods, scaling behaviors, and lessons learned. arXiv preprint arXiv:2209.07858 , 2022

    2022 arXiv

  15. [23]

    Roitberg

    Xiang Gao, Meera Sitharam, and Adrian E. Roitberg. Bounds on the jensen gap, and implications for mean-concentrated distributions, 2020. URL https://arxiv.org/abs/1712.05267

    2020 arXiv

  16. [24]

    Vision super- alignment: Weak-to-strong generalization for vision foundation models

    Jianyuan Guo, Hanting Chen, Chengcheng Wang, Kai Han, Chang Xu, and Yunhe Wang. Vision super- alignment: Weak-to-strong generalization for vision foundation models. arXiv preprint arXiv:2402.03749 , 2024

    2024 arXiv

  17. [25]

    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

  18. [26]

    Distilling the knowledge in a neural network

    Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531 , 2015

    2015 arXiv

  19. [27]

    Bounding jensen’s gap: Elementary approaches

    https://math.stackexchange.com/users/955195/small deviation. Bounding jensen’s gap: Elementary approaches. Mathematics Stack Exchange. URL https://math.stackexchange.com/q/4722509. URL:https://math.stackexchange.com/q/4722509 (version: 2023-06-20)

    2023

  20. [28]

    Cosmos qa: Machine reading comprehension with contextual commonsense reasoning

    Lifu Huang, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Cosmos qa: Machine reading comprehension with contextual commonsense reasoning. arXiv preprint arXiv:1909.00277 , 2019

    1909 arXiv

  21. [29]

    High- dimensional analysis of knowledge distillation: Weak-to-strong generalization and scaling laws

    M Emrullah Ildiz, Halil Alperen Gozeten, Ege Onur Taga, Marco Mondelli, and Samet Oymak. High- dimensional analysis of knowledge distillation: Weak-to-strong generalization and scaling laws. arXiv preprint arXiv:2410.18837, 2024

    arXiv

  22. [30]

    Aligner: Achieving efficient alignment through weak-to-strong correction

    Jiaming Ji, Boyuan Chen, Hantao Lou, Donghai Hong, Borong Zhang, Xuehai Pan, Juntao Dai, and Yaodong Yang. Aligner: Achieving efficient alignment through weak-to-strong correction. arXiv preprint arXiv:2402.02416, 2024

    2024 arXiv

  23. [31]

    The jensen’s gap and comparison of f-divergences

    Andrey Nikolaevich Konenkov. The jensen’s gap and comparison of f-divergences. Preprint, 2024. URL http://dx.doi.org/10.2139/ssrn.4965167

    2024 doi

  24. [32]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009

    2009

  25. [33]

    Imagenet classification with deep convolutional neural networks

    Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems , 25, 2012

    2012

  26. [34]

    On information and sufficiency

    Solomon Kullback and Richard A Leibler. On information and sufficiency. The annals of mathematical statistics, 22(1):79–86, 1951

    1951

  27. [35]

    Theoretical analysis of weak-to-strong generalization, 2024

    Hunter Lang, David Sontag, and Aravindan Vijayaraghavan. Theoretical analysis of weak-to-strong generalization, 2024. URL https://arxiv.org/abs/2405.16043

    2024 arXiv

  28. [36]

    Co-supervised learning: Improving weak-to-strong generalization with hierarchical mixture of experts

    Yuejiang Liu and Alexandre Alahi. Co-supervised learning: Improving weak-to-strong generalization with hierarchical mixture of experts. arXiv preprint arXiv:2402.15505 , 2024

    2024 arXiv

  29. [37]

    Hidden factors and hidden topics: understanding rating dimensions with review text

    Julian McAuley and Jure Leskovec. Hidden factors and hidden topics: understanding rating dimensions with review text. In Proceedings of the 7th ACM conference on Recommender systems , pages 165–172, 2013

    2013

  30. [38]

    Self-distillation amplifies regularization in hilbert space

    Hossein Mobahi, Mehrdad Farajtabar, and Peter Bartlett. Self-distillation amplifies regularization in hilbert space. Advances in Neural Information Processing Systems , 33:3351–3361, 2020. 12

    2020

  31. [39]

    On student-teacher deviations in distillation: does it pay to disobey? Advances in Neural Information Processing Systems, 36:5961–6000, 2023

    Vaishnavh Nagarajan, Aditya K Menon, Srinadh Bhojanapalli, Hossein Mobahi, and Sanjiv Kumar. On student-teacher deviations in distillation: does it pay to disobey? Advances in Neural Information Processing Systems, 36:5961–6000, 2023

    2023

  32. [40]

    Introducing Superalignment

    OpenAI. Introducing Superalignment. https://openai.com/blog/introducing-superalignment, 2023

    2023

  33. [41]

    A modern introduction to online learning, 2023

    Francesco Orabona. A modern introduction to online learning, 2023. URL https://arxiv.org/abs/1912. 13213

    2023

  34. [42]

    Understanding the gains from repeated self-distillation

    Divyansh Pareek, Simon S Du, and Sewoong Oh. Understanding the gains from repeated self-distillation. arXiv preprint arXiv:2407.04600 , 2024

    2024 arXiv

  35. [43]

    Language models are unsupervised multitask learners

    Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019

    2019

  36. [44]

    Re-examination of bregman functions and new properties of their divergences

    Daniel Reem, Simeon Reich, and Alvaro De Pierro. Re-examination of bregman functions and new properties of their divergences. Optimization, 68(1):279–348, November 2018. ISSN 1029-4945. doi: 10.1080/02331934.2018.1543295. URL http://dx.doi.org/10.1080/02331934.2018.1543295

    2018

  37. [45]

    Rockafellar

    R.T. Rockafellar. Convex Analysis. Princeton Landmarks in Mathematics and Physics. Princeton University Press, 1997. ISBN 9780691015866. URL https://books.google.com/books?id=1TiOka9bx3sC

    1997

  38. [46]

    Berg, and Li Fei-Fei

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJC...

    2015 doi

  39. [47]

    Weak-to-strong generalization through the data-centric lens

    Changho Shin, John Cooper, and Frederic Sala. Weak-to-strong generalization through the data-centric lens. arXiv preprint arXiv:2412.03881 , 2024

    2024 arXiv

  40. [48]

    A statistical framework for weak-to-strong generalization, 2024

    Seamus Somerstep, Felipe Maia Polo, Moulinath Banerjee, Ya’acov Ritov, Mikhail Yurochkin, and Yuekai Sun. A statistical framework for weak-to-strong generalization, 2024. URL https://arxiv.org/abs/2405.16236

    2024 arXiv

  41. [49]

    Easy- to-hard generalization: Scalable alignment beyond human supervision

    Zhiqing Sun, Longhui Yu, Yikang Shen, Weiyang Liu, Yiming Yang, Sean Welleck, and Chuang Gan. Easy- to-hard generalization: Scalable alignment beyond human supervision. arXiv preprint arXiv:2403.09472 , 2024

    2024 arXiv

  42. [50]

    Determination of bounds for the jensen gap and its applications

    Hidayat Ullah, Muhammad Adil Khan, and Tareq Saeed. Determination of bounds for the jensen gap and its applications. Mathematics, 9(23), 2021. ISSN 2227-7390. doi: 10.3390/math9233132. URL https://www.mdpi.com/2227-7390/9/23/3132

    2021 doi

  43. [51]

    Theoretical foundation of co-training and disagreement-based algorithms

    Wei Wang and Zhi-Hua Zhou. Theoretical foundation of co-training and disagreement-based algorithms. arXiv preprint arXiv:1708.04403 , 2017

    2017 arXiv

  44. [52]

    Theoretical analysis of self-training with deep networks on unlabeled data

    Colin Wei, Kendrick Shen, Yining Chen, and Tengyu Ma. Theoretical analysis of self-training with deep networks on unlabeled data. arXiv preprint arXiv:2010.03622 , 2020

    2010 arXiv

  45. [53]

    Crowdsourcing multiple choice science questions

    Johannes Welbl, Nelson F Liu, and Matt Gardner. Crowdsourcing multiple choice science questions. arXiv preprint arXiv:1707.06209, 2017

    2017 arXiv

  46. [54]

    Provable weak-to-strong generalization via benign overfitting

    David X Wu and Anant Sahai. Provable weak-to-strong generalization via benign overfitting. arXiv preprint arXiv:2410.04638, 2024

    2024 arXiv

  47. [55]

    Weak-to-strong reasoning

    Yuqing Yang, Yan Ma, and Pengfei Liu. Weak-to-strong reasoning. arXiv preprint arXiv:2407.13647 , 2024

    2024 arXiv

  48. [56]

    How does disagreement help generalization against label corruption? In International conference on machine learning , pages 7164–7173

    Xingrui Yu, Bo Han, Jiangchao Yao, Gang Niu, Ivor Tsang, and Masashi Sugiyama. How does disagreement help generalization against label corruption? In International conference on machine learning , pages 7164–7173. PMLR, 2019

    2019

  49. [57]

    Zeevi and Ronny Meir

    Assaf J. Zeevi and Ronny Meir. Density estimation through convex combinations of densities: Approx- imation and estimation bounds. Neural Networks , 10(1):99–109, 1997. ISSN 0893-6080. doi: https: //doi.org/10.1016/S0893-6080(96)00037-8. URL https://www.sciencedirect.com/scien...

    1997 doi

  50. [58]

    Transcendence: Generative models can outperform the experts that train them

    Edwin Zhang, Vincent Zhu, Naomi Saphra, Anat Kleiman, Benjamin L Edelman, Milind Tambe, Sham M Kakade, and Eran Malach. Transcendence: Generative models can outperform the experts that train them. arXiv preprint arXiv:2406.11741 , 2024. 13 A Appendix A.1 Common Bregman Diverge...

    2024 arXiv

  51. [59]

    g = f∗ ◦ hs,

    (Realizability) ∃f∗ ∈ Fs.t. g = f∗ ◦ hs,

  52. [60]

    (Convexity) F is a convex set of functions,

  53. [61]

    (Sequential Consistency) For y ∈ Uψ fixed, if Dψ(xn, y) → 0, then xn → y. Then for any ϵ >0, there exists δ >0 such that for all fs ∈ Fthat satisfy EX Dψ (fs(hs(X)), fw(hw(X))) ≤ inf f ∈F EX [Dψ (f (hs(X)), fw(hw(X)))] + δ, we have EX Dψ (g(X), fs(hs(X))) ≤ EX [Dψ (g(X), fw(hw...

  54. [63]

    (9) Proof

    (Convexity) F is a convex set of functions, then for any ϵ >0, there exists δ >0 so that for all fs ∈ Fsuch that EX [DKL(fs(hs(X))∥fw(hw(X)))] ≤ inf f ∈F EX [DKL(f (hs(X))∥fw(hw(X)))] + δ, we have EX [XE(g(X)∥fs(hs(X)))] ≤ EX [XE(g(X)∥fw(hw(X)))] − EX [DKL(fs(hs(X))∥fw(hw(X)))...

  55. [64]

    (Realizability) ∃f∗ ∈ Fso that g = f∗ ◦ hs,

  56. [65]

    functional

    (Regularization) F is s.t. max i∈[c] EX supf ∈F 1/fi(X) < ∞. Then for any k ∈ N, there exists δ >0 so for all fs ∈ cokF such that EX [DKL(fs(hs(X))∥fw(hw(X)))] ≤ inf f ∈cokF EX [DKL(f (hs(X))∥fw(hw(X)))] + δ, we have EX [XE(g(X)∥fs(hs(X)))] ≤ EX [XE(g(X)∥fw(hw(X)))] − EX [DKL(...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.