Pith. sign in

REVIEW 3 major objections 4 minor 36 references

Second-Order Bounds for [0,1]-Valued Regression via Betting Loss

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a worst-case clipped betting loss yields variance-adaptive second-order generalization bounds for [0,1]-valued regression, with the leading error term scaling as $E_x[\sigma_x^2]$ rather than a worst-case proxy.

desk verdict Genuinely new betting-loss idea for variance-adaptive [0,1]-valued regression, but the main proof has a repairable algebra error in the central step and the estimator is computationally intractable as stated. read the letter →

arxiv 2507.12584 v1 pith:JX2YA5WC submitted 2025-07-16 cs.LG

Yinan Li , Kwang-Sung Jun This is my paper
classification cs.LG
keywords [01]-valuedregressionsecond-orderboundsvariance-adaptivebettinglosslogfirst-orderheteroscedasticnoisegeneralization
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 is trying to establish that for [0,1]-valued regression, a loss function alone can deliver a second-order generalization bound: the prediction error of the empirical minimizer scales with the expected conditional variance $E_x[\sigma_x^2]$ rather than with a worst-case Bernoulli proxy such as $E_x[f^*(x)(1-f^*(x))]$. The proposed loss, called the betting loss, is a worst-case clipped log-betting objective, and the paper proves a high-probability bound for its minimizer that adapts to heteroscedastic noise without being given the variances. The paper also proves that the log-loss minimizer already yields a first-order bound that improves the standard one by scaling with $E_x[f^*(x)(1-f^*(x))]$. If correct, this means variance adaptivity is statistically free: no variance model or variance oracle is needed to obtain noise-dependent rates.

What carries the argument

The central object is the betting loss $L(f)$, a worst-case clipped version of the log-betting loss: it measures how much a candidate $f$ can be hurt when each label residual $y-f_x$ is multiplied by a clipped perturbation $\phi(h_x-f_x)$ in $[-c,c]$, maximized over all alternative hypotheses $h\in F$, perturbation size $\phi\le n/4$, and clipping $c\le 1/4$. The argument works by (i) using betting-based martingale concentration to control $L(f)-L(f^*)$ uniformly over $F$ and over a discretized grid of $(\phi,c)$, and (ii) expanding the negative log-betting factor around the conditional mean to expose $\sigma_x^2$ in the leading term, with the clipping and boundedness of $f$ making the remainder manageable. For infinite classes, the step from finite $F$ to the linear class is carried by a Lipschitz property of $L$ and standard covering numbers.

What would settle it

Take a concrete finite class, for example $F=\{f^*,g\}$ with $g$ differing from $f^*$ on a measurable set of positive probability, generate data from a known heteroscedastic model with $E_x[\sigma_x^2]$ small, and compute the betting-loss minimizer exactly by enumerating the finitely many candidates; if on repeated trials the empirical MAE exceeds the theorem's high-probability upper bound more often than $\delta$ (with any fixed positive constants), the claimed bound fails. For the computational side, if this finite min-max already lacks a closed-form argmin, the paper has not supplied an implementable estimator.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims the following. For a finite class $F$ containing the true regression function $f^*$, define $L(f) = \max_{h\in F}\max_{\phi\in[0,n/4]}\max_{c\in[0,1/4]} \frac1n \sum_{(x,y)\in D_n} \ln\bigl(1 + (y-f_x)\,[\phi(h_x-f_x)]_{[-c,c]}\bigr)$. With probability at least $1-\delta$, every $f\in F$ satisfies $E_x|f_x-f^*_x| \lesssim \sqrt{E_x[\sigma_x^2] \cdot \frac{1}{n}\ln(|F|n/\delta)} + \frac{1}{n}\ln(|F|n/\delta)$, plus terms involving the excess betting loss $L(f)-L(f^*)$, so the empirical minimizer $\hat f$ of $L$ inherits a variance-adaptive rate with no variance knowledge. The same techniques, specialized to the linear class $\{x\mapsto \theta^\top x + 1/2 : \|\theta\|_2\le 1/2\}$, give $E_x|\hat f_x - f^*_x| \lesssim \sqrt{E_x[\sigma_x^2]\, d/n \ln(n/\delta)} + d/n \ln(n/\delta)$, which the paper notes matches minimax rates up to logarithmic factors.

Load-bearing premise

The load-bearing premise is that the min-max estimator $\hat f = \arg\min_{f\in F}\max_{h\in F}\max_{\phi\in[0,n/4]}\max_{c\in[0,1/4]} \frac1n \sum_{(x,y)\in D_n} \ln(1 + (y-f_x)[\phi(h_x-f_x)]_{[-c,c]})$ exists and can actually be computed; the paper gives no efficient algorithm for this non-convex problem, and for infinite classes it only argues via a covering argument without constructing the minimizer.

Editorial extensions

If this is right

  • The betting-loss minimizer $\hat f$ satisfies, without variance knowledge, $E_x|\hat f_x - f^*_x| = O(\sqrt{E_x[\sigma_x^2] \ln(|F| n/\delta)/n} + \ln(|F|n/\delta)/n)$, so low-noise regions dominate the rate.
  • This bound is never worse in leading order than first-order bounds based on $E_x[f^*(x)(1-f^*(x))]$, and can be much smaller when conditional variance is small.
  • For the linear class in $d$ dimensions, the bound becomes $O(\sqrt{E_x[\sigma_x^2]\, d \ln(n/\delta)/n} + d\ln(n/\delta)/n)$, matching the minimax error rate up to logarithmic factors.
  • The log-loss minimizer alone already yields the improved first-order rate $O(\sqrt{E_x[f^*(x)(1-f^*(x))]\ln(|F|/\delta)/n} + \ln(|F|/\delta)/n)$.
  • Variance adaptivity is achieved purely through the loss function, so no separate variance oracle, variance function class, or distribution model is needed in the statistical analysis.

Reading between the lines

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

  • The paper leaves open whether the non-convex min-max defining $\hat f$ can be solved efficiently; if it cannot, the practical value of the result depends on finding convex surrogates or relaxations of the betting loss.
  • Because the betting loss is a worst-case perturbation, it may transfer to contextual bandits as a regression oracle, in the same spirit that log loss produces first-order regret bounds; the paper hints at this but does not prove it.
  • The bound is adaptive to the average conditional variance but not per-input; a natural extension would be a localized bound depending on $\sigma_x^2$ at each $x$, which the current techniques do not provide.
  • A numerical study comparing the betting-loss minimizer with squared-loss and log-loss minimizers on synthetic heteroscedastic data would test whether the theoretical rate advantage shows up in moderate sample sizes.
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

3 major / 4 minor

Summary. This paper studies [0,1]-valued regression under realizability. It first claims a first-order bound for the log-loss minimizer that scales with E[f*(x)(1-f*(x))], improving on the immediate implication of Foster and Krishnamurthy (2021). Its main contribution is a betting loss L(f) = max over h in F, phi in [0,phi], c in [0,1/4] of (1/n) sum ln(1 + (y - f_x)[phi(h_x - f_x)]_{[-c,c]}), with phi = n/4, and an estimator that minimizes this worst-case loss. Theorem 3 (Appendix Theorem 12) claims a uniform high-probability bound on E|f_x - f*_x| of the form sqrt{(25/12) E sigma^2 ( (2/n) ln(48|F| phi n^2/delta) + (L(f)-L(f*)) )} + (6/n) ln(...) + (5/2)(L(f)-L(f*)), from which the empirical minimizer inherits a variance-adaptive rate O(sqrt(E sigma^2 ln(|F| n/delta)/n)). Corollary 4 extends the result to linear classes by a covering-number argument.

Significance. If the central theorem is correct, the result is significant: it would show that a single loss and estimator can be variance-adaptive without receiving or modeling conditional variances, improving on first-order bounds and matching the minimax rate for linear classes up to logarithmic factors. The betting-loss construction is a fresh idea in this regression setting, and the paper is proof-heavy and self-contained, with explicit constants and a concrete linear-class instantiation. The first-order result on the log loss is also useful. However, the verification is currently blocked by an algebra/notation error in the proof of the central theorem and by an ambiguous definition used inside that proof; these are fixable, but until corrected the main claim is not fully established.

major comments (3)
  1. [Appendix B, Theorem 12 (step after Eq. (8))] The deduction in the c0 <= 1/4 case does not parse as written and does not establish the displayed constant. Let A = E[1/2 |Delta_x|^2 T] and B = E[5/3 sigma_x^2 |Delta_x|^2 T^2] with T = (1/Delta*) wedge (1/|Delta_x|). With c = c0 = A/B, inequality (8) gives (4/5) A^2/B <= R, hence A^2 <= (25/12) E sigma_x^2 R. Since A = (1/2) E[|Delta_x|^2 T], this yields [E|Delta_x|^2 T]^2 <= (25/3) E sigma_x^2 R, not (25/12) E sigma_x^2 R. Moreover the step from this to E[|Delta_x| 1{|Delta_x| >= Delta*}] <= sqrt(25/12 E sigma_x^2 R) requires the additional bound E[|Delta_x| 1{|Delta_x| >= Delta*}] <= 2A, which gives a factor sqrt(25/3), not sqrt(25/12). The theorem statement displays 25/12 while Corollary 14 uses 25/3, so the paper is internally inconsistent on exactly this constant. The variance-adaptive rate may survive, but the exact theorem as stated is not verified.
  2. [Appendix B, Definition 6 and Lemma 7] The displayed definition of U_x is missing parentheses and has ambiguous signs: it reads max{(-f*_x) - Delta_{h,x,phi,c}/(1+Delta_x Delta_{h,x,phi,c}), (1-f*_x) - Delta_{h,x,phi,c}/(1+Delta_x Delta_{h,x,phi,c})}, but the proof of Lemma 7 asserts that for Delta_{f*,x,phi,c} >= 0 this maximum equals f*_x Delta_{f*,x,phi,c}/(1+Delta_x Delta_{f*,x,phi,c}). That equality does not follow from the printed expression, and the analogous assertion in the negative case is also not transparent. Since Lemma 10 and Lemma 11 use U_x <= 1/4 to replace the second-order term by (4/3) sigma_x^2 Delta^2, this ambiguous definition sits directly in the load-bearing chain for Theorem 12 and must be restated and reproved cleanly.
  3. [Appendix A, Theorem 5 (proof of the first-order bound)] The proof contains a false equality: after Jensen's inequality the text obtains E_x[ f*_x sqrt(hat{f}_x/f*_x) + (1-f*_x) sqrt((1-hat{f}_x)/(1-f*_x)) ] and then writes "= E_x[ f*_x hat{f}_x + (1-f*_x)(1-hat{f}_x) ]". These two quantities are not equal; the first is E_x[ sqrt(f*_x hat{f}_x) + sqrt((1-f*_x)(1-hat{f}_x)) ]. Without replacing the linear terms by square-root terms, the subsequent identification with the Hellinger distance D^2(f*_x, hat{f}_x) in Eq. (4) is invalid. This is a stated theorem in the paper, and although it is preliminary to the central second-order result, the proof must be corrected.
minor comments (4)
  1. [Algorithm 1 and Corollary 4] The estimator is defined as the exact argmin of a nonconvex min-max objective; for the infinite linear class no computational procedure or existence argument is given. The conclusion acknowledges computational challenges, but the body should state near Algorithm 1 that the guarantees are statistical statements about the exact minimizer, not about a practical algorithm.
  2. [Section 3, notation] The symbol phi is overloaded: phi := n/4 is a fixed parameter, while the same symbol is used as the variable in max_{phi in [0,phi]}. Rename the fixed parameter, e.g., phi-bar, to avoid confusion in Theorem 12 and the appendix.
  3. [Throughout] There are several typos and presentation issues: "derivesecond-order", "Lipchitz", an incomplete reference "(Neu)" without year/venue, and the unparenthesized formula for U_x in Definition 6. These should be cleaned up.
  4. [Appendix B, Lemma 9] The block-counting step is terse. With epsilon = 1/(4n) the discretization has phi/(4 epsilon^2) = n^3 blocks, and the final bound ln(8 phi n^2/delta) = ln(2 n^3/delta) uses 1/4 <= ln 2. This is correct but should be spelled out, since it is easy to misread as a missing factor.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the variance-adaptive second-order bound follows from a self-contained betting-martingale concentration argument, the estimator never receives or fits the variance, and the same-author citations are non-load-bearing; the Appendix algebra gap around Eq. (8) is a correctness risk, not a circular step.

full rationale

Walking the claimed derivation chain, the central result (Theorem 3, restated as Theorem 12 in Appendix B) is a high-probability uniform bound on E|fx − f*x| in terms of the population variance Eσ²x and the empirical excess betting loss L(f) − L(f*), and no displayed step reduces that bound to the loss's definition or to a fitted input. Algorithm 1's objective L(f) = max_h max_ϕ max_c (1/n) Σ ln(1 + (y − fx)[ϕ(hx − fx)][−c,c]) is data-dependent only, over the fixed grids ϕ ∈ [0, n/4] and c ∈ [0, 1/4]; the estimator never receives σ²x and no variance parameter is fitted, so the bound's Eσ²x term cannot be a disguised fit. The variance enters through Lemma 11's expansion 1/(1+x) = 1 − x + x²/(1+x) with E[η|x] = 0 and E[η²|x] = σ²x, a genuine second-moment computation, and the excess term for the empirical minimizer is ≤ 0 by the definition of argmin, the standard ERM step, so the fast variance-adaptive rate is derived rather than definitional. The same-author citations (Orabona and Jun 2024; Jun and Kim 2024) are inspiration or related work only: the Appendix re-derives the betting martingale bounds from Markov's inequality with an explicit ε-discretization of [0, ϕ] × [0, 1/4], and the log-loss bound is credited to Foster and Krishnamurthy (2021), so no load-bearing premise imports an unverified same-author result. What is genuinely flagged is missing support, not circularity: in Appendix B the c0/Δ* substitution step around Eq. (8) does not parse as printed (direct substitution of c0 into Eq. (8) does not yield the displayed squared inequality, and the final constant depends on this transition), Definition 6's Ux has a parenthesization slip, Lemma 7 proves Ux ≤ 1/4 only for h = f* (the case later used in the proof), and the paper itself concedes the estimator's computability is open in the conclusion. These are correctness and implementation caveats located exactly where the variance-adaptive claim is established, but they are not equivalences between the result and its own inputs.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central bound rests on realizability, i.i.d. data, concentration tools, and the assumption that the min-max estimator exists. No new physical or mathematical entities are introduced.

free parameters (2)
  • perturbation scale phi = n/4
    Fixed in Algorithm 1; chosen by hand, not learned from data. The theorem's proof uses this value as the upper bound of the max.
  • clipping threshold bound = 1/4
    The algorithm maximizes over c in [0,1/4]; the range is a design choice that keeps the log argument bounded in [3/4,5/4].
assumptions (5)
  • domain assumption i.i.d. data assumption
    Stated in Section 2: (x_t,y_t) ~ D i.i.d.
  • domain assumption realizability: f* in F
    Assumed in Section 2. The theorem bounds E|f-f*|; without realizability, the target is not in the class.
  • standard math standard concentration inequalities (Markov, union bound)
    Used throughout the appendix proofs.
  • domain assumption finite hypothesis class for Theorem 3; covering numbers for corollary
    The main theorem's proof uses |F|; the linear corollary uses metric entropy.
  • ad hoc to paper existence of argmin for the min-max objective
    Algorithm 1 returns an argmin; for general F this is not guaranteed and is computationally hard.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Second-Order Bounds for [0,1]-Valued Regression via Betting Loss." pith.science (2026). https://pith.science/paper/JX2YA5WC

@misc{pith2026250712584,
  author       = {Pith},
  title        = {Pith review of: Second-Order Bounds for [0,1]-Valued Regression via Betting Loss},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JX2YA5WC}},
  note         = {Machine review of arXiv:2507.12584}
}
abstract

We consider the $[0,1]$-valued regression problem in the i.i.d. setting. In a related problem called cost-sensitive classification, \citet{foster21efficient} have shown that the log loss minimizer achieves an improved generalization bound compared to that of the squared loss minimizer in the sense that the bound scales with the cost of the best classifier, which can be arbitrarily small depending on the problem at hand. Such a result is often called a first-order bound. For $[0,1]$-valued regression, we first show that the log loss minimizer leads to a similar first-order bound. We then ask if there exists a loss function that achieves a variance-dependent bound (also known as a second order bound), which is a strict improvement upon first-order bounds. We answer this question in the affirmative by proposing a novel loss function called the betting loss. Our result is ``variance-adaptive'' in the sense that the bound is attained \textit{without any knowledge about the variance}, which is in contrast to modeling label (or reward) variance or the label distribution itself explicitly as part of the function class such as distributional reinforcement learning.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 34 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Open problem: First-order regret bounds for contextual bandits

    Agarwal, A., Krishnamurthy, A., Langford, J., Luo, H., et al. Open problem: First-order regret bounds for contextual bandits. In Proceedings of the Conference on Learning Theory (COLT), pages 4--7, 2017

    work page 2017

  3. [3]

    On least squares and linear combination of observations

    Aitkin, A. On least squares and linear combination of observations. Proceedings of the Royal Society of Edinburgh, 55: 0 42--48, 1935

    work page 1935

  4. [4]

    Make the minority great again: First-order regret bound for contextual bandits

    Allen-Zhu, Z., Bubeck, S., and Li, Y. Make the minority great again: First-order regret bound for contextual bandits. In Proceedings of the International Conference on Machine Learning (ICML), pages 186--194, 2018

    work page 2018

  5. [5]

    Hannan Consistency in On-Line Learning in Case of Unbounded Losses Under Partial Monitoring

    Allenberg, C., Auer, P., Gy \" o rfi, L., and Ottucs \' a k, G. Hannan Consistency in On-Line Learning in Case of Unbounded Losses Under Partial Monitoring . In Algorithmic Learning Theory (ALT), pages 229--243. 2006

    work page 2006

  6. [6]

    Improved second-order bounds for prediction with expert advice

    Cesa-Bianchi, N., Mansour, Y., and Stoltz, G. Improved second-order bounds for prediction with expert advice. Machine Learning, 66: 0 321--352, 2007

    work page 2007

  7. [7]

    Foster, D. J. and Krishnamurthy, A. Efficient first-order contextual bandits: Prediction, allocation, and triangular discrimination. Advances in Neural Information Processing Systems (NeurIPS), pages 18907--18919, 2021

    work page 2021

  8. [8]

    and Schapire, R

    Freund, Y. and Schapire, R. E. A decision-theoretic generalization of on-line learning and an application to boosting. Journal of computer and system sciences, 55 0 (1): 0 119--139, 1997

    work page 1997

Show all 36 references
  1. [9]

    Regression with input-dependent noise: A gaussian process treatment

    Goldberg, P., Williams, C., and Bishop, C. Regression with input-dependent noise: A gaussian process treatment. Advances in Neural Information Processing Systems (NeurIPS), 1997

    1997

  2. [10]

    and Kale, S

    Hazan, E. and Kale, S. Extracting certainty from uncertainty: Regret bounded by variation in costs. Machine learning, 80: 0 165--188, 2010

    2010

  3. [11]

    and Kale, S

    Hazan, E. and Kale, S. Better algorithms for benign bandits. Journal of Machine Learning Research, 12 0 (4), 2011

    2011

  4. [12]

    Tight first- and second-order regret bounds for adversarial linear bandits

    Ito, S., Hirahara, S., Soma, T., and Yoshida, Y. Tight first- and second-order regret bounds for adversarial linear bandits. In Advances in Neural Information Processing Systems (NeurIPS), volume 33, pages 2028--2038, 2020

    2020

  5. [13]

    How does variance shape the regret in contextual bandits? In Advances in Neural Information Processing Systems (NeurIPS), 2024

    Jia, Z., Qian, J., Rakhlin, A., and Wei, C.-Y. How does variance shape the regret in contextual bandits? In Advances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  6. [14]

    and Kim, J

    Jun, K.-S. and Kim, J. Noise-adaptive confidence sets for linear bandits and application to bayesian optimization. In Proceedings of the International Conference on Machine Learning (ICML), 2024

    2024

  7. [15]

    Most likely heteroscedastic gaussian process regression

    Kersting, K., Plagemann, C., Pfaff, P., and Burgard, W. Most likely heteroscedastic gaussian process regression. In ACM International Conference Proceeding Series, volume 227, 2007

    2007

  8. [16]

    Improved regret analysis for variance-adaptive linear bandits and horizon-free linear mixture mdps

    Kim, Y., Yang, I., and Jun, K.-S. Improved regret analysis for variance-adaptive linear bandits and horizon-free linear mixture mdps . In Advances in Neural Information Processing Systems (NeurIPS), 2022

    2022

  9. [17]

    and Krause, A

    Kirschner, J. and Krause, A. Information directed sampling and bandits with heteroscedastic noise. In Proceedings of the Conference on Learning Theory (COLT), pages 358--384. PMLR, 2018

    2018

  10. [18]

    and Szepesv \' a ri, C

    Lattimore, T. and Szepesv \' a ri, C. Bandit Algorithms . 2018. URL http://downloads.tor-lattimore.com/book.pdf

    2018

  11. [19]

    URL https://cs.bme.hu/ gergo/files/tutorial.pdf

    Neu, G. URL https://cs.bme.hu/ gergo/files/tutorial.pdf

  12. [20]

    and Jun, K.-S

    Orabona, F. and Jun, K.-S. Tight concentrations and confidence sequences from the regret of universal portfolio. IEEE Transactions on Information Theory, 70 0 (1): 0 436--455, 2024. doi:10.1109/TIT.2023.3330187

    2024

  13. [21]

    Second order bounds for contextual bandits with function approximation

    Pacchiano, A. Second order bounds for contextual bandits with function approximation. In Proceedings of the International Conference on Learning Representations (ICLR), 2025

    2025

  14. [22]

    and Van Roy , B

    Russo, D. and Van Roy , B. Eluder dimension and the sample complexity of optimistic exploration . In Advances in Neural Information Processing Systems (NeurIPS), pages 2256--2264, 2013

    2013

  15. [23]

    Incomplete information and internal regret in prediction of individual sequences

    Stoltz, G. Incomplete information and internal regret in prediction of individual sequences. PhD thesis, Universit \'e Paris Sud-Paris XI, 2005

    2005

  16. [24]

    Tsybakov, A. B. Introduction to Nonparametric Estimation. Springer Series in Statistics. Springer, 2004

    2004

  17. [25]

    J., Chen, Y., Simchowitz, M., Du, S., and Jamieson, K

    Wagenmaker, A. J., Chen, Y., Simchowitz, M., Du, S., and Jamieson, K. First-order regret in reinforcement learning with linear function approximation: A robust estimation approach. In Proceedings of the International Conference on Machine Learning (ICML), pages 22384--22429, 2022

    2022

  18. [26]

    Wainwright, M. J. High-dimensional statistics: A non-asymptotic viewpoint, volume 48. Cambridge university press, 2019

    2019

  19. [27]

    More benefits of being distributional: Second-order bounds for reinforcement learning

    Wang, K., Oertell, O., Agarwal, A., Kallus, N., and Sun, W. More benefits of being distributional: Second-order bounds for reinforcement learning. In Proceedings of the International Conference on Machine Learning (ICML), 2024 a

    2024

  20. [28]

    J., Sreedhar, M

    Wang, Z., Dong, Y., Delalleau, O., Zeng, J., Shen, G., Egert, D., Zhang, J. J., Sreedhar, M. N., and Kuchaiev, O. Helpsteer2: Open-source dataset for training top-performing reward models. arXiv preprint arXiv:2406.08673, 2024 b

    2024 arXiv

  21. [29]

    and Ramdas, A

    Waudby-Smith, I. and Ramdas, A. Estimating means of bounded random variables by betting. Journal of the Royal Statistical Society Series B: Statistical Methodology, 2023

    2023

  22. [30]

    Experimental designs for heteroskedastic variance

    Weltz, J., Fiez, T., Volfovsky, A., Laber, E., Mason, B., Jain, L., et al. Experimental designs for heteroskedastic variance. Advances in Neural Information Processing Systems (NeurIPS), 2023

    2023

  23. [31]

    Optimal comparator adaptive online learning with switching cost

    Zhang, Z., Cutkosky, A., and Paschalidis, Y. Optimal comparator adaptive online learning with switching cost. Advances in Neural Information Processing Systems (NeurIPS), pages 23936--23950, 2022

    2022

  24. [32]

    Variance-dependent regret bounds for linear bandits and reinforcement learning: Adaptivity and computational efficiency

    Zhao, H., He, J., Zhou, D., Zhang, T., and Gu, Q. Variance-dependent regret bounds for linear bandits and reinforcement learning: Adaptivity and computational efficiency. In Proceedings of the Conference on Learning Theory (COLT), volume 195 of Proceedings of Machine Learning ...

    2023

  25. [33]

    Optimal online generalized linear regression with stochastic noise and its application to heteroscedastic bandits

    Zhao, H., Zhou, D., He, J., and Gu, Q. Optimal online generalized linear regression with stochastic noise and its application to heteroscedastic bandits. In Proceedings of the International Conference on Machine Learning (ICML), pages 42259--42279, 2023 b

    2023

  26. [34]

    Adaptive experimentation when you can't experiment

    Zhao, Y., Jun, K.-S., Fiez, T., and Jain, L. Adaptive experimentation when you can't experiment. In Advances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  27. [35]

    and Gu, Q

    Zhou, D. and Gu, Q. Computationally efficient horizon-free reinforcement learning for linear mixture mdps. Advances in Neural Information Processing Systems (NeurIPS), pages 36337--36349, 2022

    2022

  28. [36]

    Nearly minimax optimal reinforcement learning for linear mixture markov decision processes

    Zhou, D., Gu, Q., and Szepesvari, C. Nearly minimax optimal reinforcement learning for linear mixture markov decision processes . In Proceedings of the Conference on Learning Theory (COLT), pages 4532--4576. PMLR, 2021

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.