acceptodds
Under review as a conference paper at ICLR 2027

Contextual Swap Regret via Localized Martingale Absorption

Abstract

Online forecasts can be assessed through errors that remain correctable in the issued record. Contextual swap regret measures the reduction in cumulative squared loss from contextual replacements and supports calibration guarantees. Distribution-weighted guarantees do not directly control randomized forecasters' realized records. Over prediction rounds, at fixed dimension and comparator radius, existing methods attain distribution-weighted regret, whereas the prior realized bound is with high probability. For affine replacements, we improve the latter to , matching the distribution-weighted horizon exponent. Sampling fluctuations shrink with replacement distance. Our analysis exploits this relationship to let square loss absorb the sampling cost, preserving a finer prediction grid. Lower bounds establish the optimal horizon exponent at fixed confidence. We also characterize expected minimax regret across dimensions and horizons up to logarithmic factors for every fixed , where bounds replacement parameter norms. Experiments assess the grid choice on synthetic streams and census income prediction data. Within a shared ridge implementation, the finer grid reduces median realized regret to of the coarse grid value at on synthetic data, and improves realized regret and affine calibration error on the census benchmarks.

Primary area:
learning theory
Keywords:
online learning, contextual swap regret, multicalibration, martingale concentration, high-probability regret bounds

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_edQQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper studies the rates at which contextual swap regret can be guaranteed. The setup is as follows: At every round a real valued label y_t is realized, and just before the realization, the algorithm must commit to a forecast p_t. The algorithm seeks to minimize squared error. The weakest regret benchmark, external regret, compares the algorithm's squared loss to that of the best single prediction p* in hindsight. Swap regret compares to a stronger comparator: each distinct prediction value v, it allows swapping out all predictions {p_t : p_t = v} with a different fixed value p*_v. External regret is the special case in which p*_v = p* for all v. Contextual swap regret goes one step further, and allows swapping out predictions in the set {p_t : p_t = v} with a function f_v(x) of the context. Swap regret is the special case of contextual swap regret where each f_v(x) = p*_v is a constant. This paper studies the case in which the functions f_v are affine in the context.

There is an important technical distinction in how we might measure swap regret. We could measure it in expectation over the randomness of the prediction distribution ("pseudo" swap regret), or with respect to the realizations? Prior work left a gap. *(Well, actually there is some more recent prior work that has been on arxiv for a few months that complicates the distinction --- see the related work discussion in weaknesses) It gave a T^{1/3} bound for pseudo-swap regret, but only T^{3/5} for realized. This paper closes the gap and gives upper and lower bounds of T^{1/3} for realized pseudo-swap regret.

Critical strengths

The theorem is correct and makes real progress on the problem of pinning down our understanding of contextual swap reget.

Critical weaknesses

This paper is quite hard to read; it is written in a noticable LLM voice. It could use a heavy human pass, both for language and organization. Other than writing, the biggest issue with the paper is its incomplete discussion of related work, of which there is a lot. I enumerate some of the missing related work below (with LLM assistance, although all connections are verified). Some is just missing historical context, but there is also a recent preprint that has been on arxiv for several months that obtains similar results and should be discussed/compared.

Missing related work

The related-work discussion needs a substantive revision. Below are some uncited papers and how they are related; some are just for completeness of the historical context, but others have direct relevance to the novelty of the results claimed in this paper:

1. “High-Dimensional Prediction for Sequential Decision Making”

Authors: Georgy Noarov, Ramya Ramalingam, Aaron Roth, and Stephan Xie. 2023 preprint; ICML 2025. https://arxiv.org/abs/2310.17651 https://proceedings.mlr.press/v267/noarov25b.html

This paper should be cited in the calibration history in Section 2 and in the discussion of Corollary 5. It gives the first efficient online multicalibration algorithm achieving cumulative O(T^{2/3}) ECE, equivalently normalized O(T^{-1/3}) ECE. Ghuge et al. later gave a different algorithm recovering this result and is cited, but the original paper is uncited.

2. “Simultaneous Swap Regret Minimization via KL-Calibration”

Authors: Haipeng Luo, Spandan Senapati, and Vatsal Sharan. 2025. https://arxiv.org/abs/2502.16387

This is a different paper from the currently cited “Improved Bounds for Swap Multicalibration and Swap Omniprediction,” despite having the same authors. It should be cited as a particularly close technical antecedent to Lemmas 2 and 3.

Its Theorem 3 gives an algorithm-independent, high-probability conversion from pseudo squared calibration error to realized squared calibration error: the latter is at most a constant multiple of the former plus an additive O(K log(K/delta)) term for K prediction values. Consequently, the conversion preserves the cube-root horizon exponent in the noncontextual setting.

This is structurally close to the present conversion, although is not written for contextual regret. The authors should explain what additional difficulty the affine class creates and how their localization/pullback argument addresses it. The relevant novelty is potentially this contextual extension, not the general observation that sampling need not worsen the cube-root exponent.

3. “Fast Rates for Swap-Agnostic Learning of Proper Losses”

Author: Princewill Okoroafor. 2026 preprint. https://arxiv.org/abs/2607.28856

This is a direct comparison for the main upper-bound claim. Theorem 4.5 gives high-probability realized hypothesis-swap regret of order \widetilde O(T^{1/3} (log |H|)^{2/3}) for a finite bounded hypothesis class H and a fixed Lipschitz proper loss, including squared loss. The benchmark permits a different hypothesis at each issued prediction value. The proof uses a second-order guarantee and a negative quadratic activity term to cancel the variance-related cost, making it relevant to the manuscript's technical narrative as well as its rate.

There is also a straightforward statistical implication for the present affine class. Take an epsilon-net of the affine parameter ball and clip the resulting functions to [0,1]. Its logarithmic size is O_R((d+1) log(1/epsilon)). Clipping can only improve a comparator's squared loss, and approximation adds O(T epsilon) to regret. Taking epsilon=T^{-2} therefore recovers the same \widetilde O_R(T^{1/3}d^{2/3}) dependence for binary outcomes.

This net construction is not computationally efficient, but it does preclude an unqualified claim that the statistical rate was previously unknown. The paper also gives an expected Omega(T^{1/3}) lower bound for a fixed three-hypothesis class in Theorem 4.7, using the recalibration lower bound discussed below.

4. “Optimal Recalibration of an Online Predictor”

Authors: Lunjia Hu, Kevin Tian, and Chutong Yang. 2026 preprint. https://arxiv.org/abs/2607.19689

This paper should be cited in Section 5 and Appendix B as a direct predecessor of the scalar lower bound. Its Theorem 13 proves that simultaneously achieving expected normalized calibration error epsilon and expected average squared-loss excess epsilon^2 relative to supplied forecasts requires T=Omega(epsilon^{-3}). Its hard instance cycles through evenly spaced forecast values in [1/4,3/4] and generates independent Bernoulli labels with those means. This is the same basic construction used in the present scalar lower bound, and it can be sampled before the learner's randomization.

The expected Omega(T^{1/3}) contextual swap-regret lower bound follows directly in the present dimension-one, radius-one setting. Take the supplied forecast q_t as the context. The affine class contains the identity comparator f(x)=x and every constant comparator. Thus contextual swap regret S_T dominates both the excess squared loss relative to q_t and the squared calibration error. In particular, E[ECE] <= sqrt(E[S_T]/T), while expected average excess loss is at most E[S_T]/T. An expected regret guarantee o(T^{1/3}) would contradict the recalibration lower bound.

The manuscript should distinguish the earlier expected lower bound and construction from its new probability guarantee, rather than presenting the entire scalar obstruction without attribution.

5. “Optimal Lower Bounds for Online Multicalibration”

Authors: Natalie Collina, Jiuyao Lu, Georgy Noarov, and Aaron Roth. FOCS 2026. https://arxiv.org/abs/2601.05245

This paper establishes optimal cube-root normalized multicalibration lower bounds.

It should be cited both in the discussion of optimal calibration rates and for the provenance of the lower-bound construction. Hu, Tian, and Yang explicitly identify their cyclic Bernoulli construction with the construction in this paper. The present manuscript uses the same basic arrangement of repeatedly visited, separated Bernoulli means and a tradeoff between staying near those means and correcting realized noise.

This is probably the originator of this kind of lower bound instance.

6. “Multicalibration as Boosting for Regression”

Authors: Ira Globus-Harris, Declan Harrison, Michael Kearns, Aaron Roth, and Jessica Sorrell. ICML 2023. https://arxiv.org/abs/2301.13767

This paper should be cited for the connection between multicalibration, regression within prediction level sets, and squared-loss swap-like optimality. It characterizes multicalibration through the absence of sufficiently useful squared-loss corrections and develops a regression-oracle boosting algorithm organized around prediction level sets.

This is a direct conceptual antecedent to the present interpretation of contextual replacements as an audit for correctable errors, as well as to the regret-to-multicalibration argument in Corollary 5. The underlying idea that a residual correlation can be turned into a squared-loss improvement is not new to the present paper. The discussion should explain whether its deterministic reduction is a specialization, a quantitative refinement, or a same-radius variant of earlier relationships.

7. “Efficient Swap Multicalibration of Elicitable Properties”

Authors: Lunjia Hu, Haipeng Luo, Spandan Senapati, and Vatsal Sharan. 2025 preprint. https://arxiv.org/abs/2511.04907

This paper should be included in the discussion of the calibration frontier and Corollary 5. It gives an oracle-efficient reduction from online agnostic learning to swap multicalibration for general elicitable properties and bounded hypothesis classes. For the mean and squared calibration-error convention, its guarantees include cumulative \widetilde O(T^{1/3}) error, yielding normalized squared error \widetilde O(T^{-2/3}) and normalized l1 error \widetilde O(T^{-1/3}).

The comparison is not only about matching the horizon exponent. Its linear-class corollary uses OGD and gives dimension-independent error bounds under its unit-norm assumptions, so the manuscript should explicitly compare test classes, normalization, dimension dependence, and computational assumptions.

Questions for authors

Please describe how you plan to revise the papers exposition to make it easier for a human being to follow, and describe how you will incorporate the missing related work. Do you agree that Okoroafor gives a comparable (albeit perhaps computationally inefficient) upper bound?

1 comment
AuthorsOpenReview ↗

Comparable prior bounds, the new results, and a clearer exposition

Q(c). Does Okoroafor already give a comparable upper bound?

Yes. Okoroafor's Theorem 4.5 gives a comparable statistical upper bound for binary outcomes. Our advance is an efficient construction on the continuous affine class, together with probability and dimension lower bounds. The original learner updates directly on the coefficient ball; the revision explicitly compares it with the finite-class result.

Applying Õ(T^(1/3)(log|H|)^(2/3)) to a clipped ε-net of our ball gives log|H|=O_R((d+1)log(1/ε)) and approximation cost at most 2√2 Tε. Taking ε=T^(-2) recovers Õ_R(T^(1/3)d^(2/3)). Okoroafor also uses a negative quadratic term to cancel second-order cost. The revised Introduction and Related Work credit these connections.

Applying that algorithm directly to the cover maintains (N+1)|H| weights. Our ONS learner works on the continuous ball, with O(Nd²) updates plus projection and a stationary-distribution solve; the cover is used only in the proof. Our bound permits real outcomes in [0,1] on oblivious sequences; Okoroafor permits binary outcomes with adaptation before each fresh prediction draw. Table 1 and Appendix F.1 compare these protocols and costs.

W/Q(b). How are the seven missing works incorporated?

All seven are cited and compared in the revision. We distinguish continuous-class computation, uniform sampling control within prediction buckets, fixed-confidence lower bounds and the dimension–horizon expected minimax characterization.

  1. Noarov–Ramalingam–Roth–Xie. We credit their efficient online multicalibration result before Ghuge: for finite events and discrete bins, expected cumulative Õ(T^(2/3)) error, or normalized Õ(T^(-1/3)), with its computational access stated.

  2. Luo–Senapati–Sharan, KL-Calibration. Their Theorem 3 already gives Cal₂≤6 PCal₂+96K log(4K/δ) with probability at least 1−δ, preserving the scalar cube-root exponent. Our extension controls all permitted affine replacements in every bucket simultaneously: localize deviations by squared replacement distance, then absorb them by feasible mixing with the constant prediction. Lemmas 2–3 explain this continuous-class step.

  3. Okoroafor. The cover recovers the same binary-outcome statistical rate, while direct enumeration maintains (N+1)|H| weights. Our continuous ONS updates avoid this enumeration. Table 1 and F.1 compare the rate, outcome protocol and computation as above.

  4. Hu–Tian–Yang. Taking their supplied forecast as context makes identity and constant corrections available, yielding an expected scalar Ω(T^(1/3)) swap-regret obstruction; Okoroafor's Theorem 4.7 also gives an expectation bound. Our Theorem 6 proves P(S≥cδT^(1/3))≥1−δ, with absolute c under the stated δ-dependent horizon condition. The revision separates the prior expectation result from this probability guarantee.

  5. Collina–Lu–Noarov–Roth. We credit the cyclic Bernoulli construction used by Hu–Tian–Yang. Our dimensional construction must also satisfy one joint coefficient-norm constraint, including at radius one. It yields expected Ω(T^(1/3)d^(2/3)) under the theorem's horizon condition. With the upper bounds, the expected minimax order is min{T^(1/3)d^(2/3),T^(3/5)} up to logarithms, for fixed R≥1.

  6. Globus-Harris et al. Their regression–calibration connection uses level-set regression, residual correlation and squared-loss improvement. Our bounded coefficient ball makes arbitrary shifts or rescaling infeasible. Feasible mixing gives the quantitative same-radius form MC₁≤C_mc√(S/T), MC₂≤C_mc²S/T (A.5).

  7. Hu–Luo–Senapati–Sharan. Their OGD reduction for unit-norm linear tests gives dimension-independent cumulative squared swap-calibration error Q=Õ(T^(1/3)), hence Q/T=Õ(T^(-2/3)) and normalized L1 error Õ(T^(-1/3)). Q measures residual/test correlations within buckets; our S measures the largest squared-loss reduction from permitted affine replacements. Their rates are not interchangeable. Our corollary derives MC₂=Õ(T^(-2/3)d^(2/3)) and MC₁=Õ(T^(-1/3)d^(1/3)) from S after radius scaling. Related Work compares metrics, test classes and oracle access; our main result directly controls replacement gain and characterizes its expected minimax order.

W/Q(a). How is the argument made easier to follow?

The revision defines realized and probability-weighted records before the theorem, explaining why optimizing corrections prevents identifying pseudo-regret with expected realized regret. It then connects the deviation term, squared replacement distance, feasible midpoint identity and finer grid. The original conceptual figure is integrated into the Introduction; the new proof figure distinguishes the exact identity from the high-probability bound. Table 1 and the lower-bound discussion identify prior rates and constructions alongside our computation, probability and dimension results.

Reviewer_yysiOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper studies the contextual swap regret problem for an online linear regression problem with squared loss. The prior state-of-the-art was from Luo et al. 2025, who obtained a bound of order T1/3d2/3T^{1/3}d^{2/3} for the pseudo-version of the swap regret and a weaker bound T3/5d2/5T^{3/5}d^{2/5} for the actual swap regret. The first contribution of this paper is to show that the same algorithm of Luo et al. 2025 (with the parameters set right) can in fact already achieve T1/3d2/3T^{1/3}d^{2/3} for the actual swap regret, as long as one analyzes the martingale concentration in a more careful and localized way and mixes the comparator in the regret with the current prediction of interest. This improved actual swap regret bound also implies similar improvements for swap multicalibration based on known reductions from e.g. Luo et al. 2025.

The paper then points out the upper bound can be improved to min⁡T1/3d2/3,T3/5\min\\{T^{1/3}d^{2/3}, T^{3/5}\\}, where the latter dimension-independent bound is achieved by using OGD instead of ONS in the Blum-Mansour reduction. More importantly, the paper shows a matching lower bound, thus fully characterizing the minimax swap regret of this problem. Finally, the paper concludes with some experimental results, confirming that choosing the parameters as the theory of this paper predicts does lead to better performance compared to Luo et al. 2025.

Critical strengths
  1. The problem studied is fundamental and receives increasing interest in the community due to its connection to omniprediction, (multi)calibration, and more generally trustworthy machine learning.

  2. The paper closes the existing gap between pseudo swap regret and actual swap regret.

  3. More importantly, the paper fully characterizes the minimax swap regret by providing a matching lower bound.

Critical weaknesses

The main reason I do not recommend acceptance is that the paper is poorly written and, in my opinion, clearly written by AI for the majority part of it. I don't think I need to list all the writing issues and/or "evidence" of AI-generated contents, but I think one can already draw their own conclusion from Figures 1 and 2: they are completely not understandable without further explanations and also in my mind fairly useless in terms of illustrating the key ideas of the proof.

Questions for authors

I don't really have questions whose answers could materially affect my assessment.

1 comment
AuthorsOpenReview ↗

How the original figures explain the proof, and what the revision clarifies

W. How do Figures 1 and 2 explain the proof?

Figure 1 explains why localized analysis permits a finer prediction grid; Figure 2 shows the three proof steps: localized deviation control, feasible comparator mixing, and absorption by squared loss. These are the localization and mixing mechanisms you identify in your Summary. We disagree that the figures contribute almost nothing to explaining the proof, and clarify their correspondence with the original text below.

Figure 1: why does changing the deviation analysis allow a finer grid? The original figure compares two treatments of sampling cost. A uniform worst-case analysis charges a cost in every prediction bucket, so finer grids accumulate more cost. Localization instead makes that cost depend on correction size, allowing squared loss to absorb it and retain a finer grid. The original Introduction (lines 46–55) and caption (lines 73–77) explain this contrast and why the proof needs both localization and comparator mixing.

Figure 2: how is the deviation term absorbed? Panel (a) corresponds to Lemma 2: squared distance between a correction and the current prediction controls sampling deviation. Panel (b) depicts the feasible midpoint; 0.50, 0.76 and 0.63 are the current prediction, comparator value and midpoint. The class is convex and contains constant predictions, so f_half=(p+f)/2 remains feasible. If Δ(f) is the squared-loss improvement from replacing p by f(x), then

2Δ(f1/2)=Δ(f)+12(p−f(x))2.2\Delta(f_{1/2})=\Delta(f)+\tfrac12\bigl(p-f(x)\bigr)^2.

The extra half squared-distance term absorbs Lemma 2's localized deviation term at coefficient 1/2. Maximizing within each bucket and summing gives Lemma 3's conversion from pseudo to realized regret. Panel (c) depicts this absorption and the resulting balance of grid error and learning cost. The original text (lines 249–269) explains these steps, with Lemma 3 leading to Theorem 4.

The revision makes these correspondences explicit. We integrate original Figure 1's conceptual explanation into the Introduction and rebuild the proof figure around the identity and its substitution into the localized bound. The caption distinguishes the exact identity from the high-probability inequality, and the text states each lemma's purpose before using it. Readers can follow the formulas without translating labels such as “offset” and “credit” into mathematical relations themselves.

AI-assisted writing was disclosed on original p. 10, including assistance with drafting, editing and presentation, with author review and responsibility for the content. We welcome concrete readability concerns and ask that assessments identify the text, figures or derivations that obstruct understanding so that we can address them directly.

Your review recognizes the closing of the pseudo/actual swap-regret gap and the expected minimax characterization from matching lower bounds. We ask you to reconsider the overall assessment in light of the original explanations and the revisions above. If a specific step still affects your judgment, please identify it so that we can clarify it directly.

Reviewer_MvfEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies contextual swap regret, which measures how much error in online predictions could still be reduced through corrections based on sample features. The motivation is that existing guarantees are stronger for performance weighted by prediction probabilities than for the predictions actually produced by randomized models. The authors improve the analysis using the fact that smaller corrections lead to smaller sampling fluctuations. This allows an existing learning framework to use more closely spaced prediction values. Its main contributions are improving the regret bound from roughly T^(3/5) to T^(1/3) and proving that this growth rate is the best possible for a fixed number of features under the stated assumptions. The paper provides mathematical proofs and experiments on synthetic data and the Adult and ACSIncome datasets. Within the same prediction framework, the results show lower regret and better agreement between predicted probabilities and observed outcomes.

Critical strengths

The paper is interesting and got some strong points:

  1. The paper gives a stronger guarantee for the predictions actually issued by a randomized model. Its improved analysis of sampling fluctuations shows that the cumulative error still correctable afterwards can grow more slowly as the number of predictions increases.

  2. The paper proves that the improved growth exponent is the best possible for a fixed number of features. No algorithm can guarantee a smaller exponent across all allowed data sequences. The analysis further explains how this limit changes as the number of features increases, clarifying the scope for further improvement.

Critical weaknesses

In my opinion, a couple of issues that stand out as potential weaknesses:

  1. The practical benefit over simpler models is unclear. The main experiments show that the method works better when it can choose from more probability values. But a simpler model performs similarly or better in some tests with more input features. The paper therefore needs to explain when the extra complexity is useful.

  2. The theoretical guarantee depends strongly on the chosen features. Regret measures how much prediction error can be reduced by linear corrections using these features. Low regret can therefore leave systematic errors from nonlinear relationships undetected. Covering these errors may require additional features, which can weaken the error bound and increase computational costs.

Questions for authors

A couple of questions that I would like the authors to respond to:

  1. In some experiments with more input features, a single ridge model achieves similar or better results. What explains this pattern? Based on the existing results, under what conditions does the proposed method’s added complexity provide consistent benefits?

  2. Which prediction errors caused by nonlinear relationships or feature interactions can the chosen features detect? For which groups do the reported low regret and calibration error provide evidence of reliable predictions?

1 comment
AuthorsOpenReview ↗

When contextual corrections help, and what their guarantees cover

W1/Q1. When is the extra complexity useful, and why can single ridge do better?

The extra bucket structure lets different prediction levels use different error corrections; consistent gains are clearest on long, low-dimensional streams and the longer Adult/ACS streams. Single ridge uses one shared model. In original Table 7, at d=2 and T=10,000,30,000,100,000, the bucketed method wins on all 20 seeds at each horizon. At T=30,000, median affine swap regret falls from 261.3 to 124.9 on Adult and from 283.7 to 171.1 on ACS, reductions of about 52% and 40%. Affine calibration error also falls from .0468 to .0286 and .0475 to .0266 (Table 8).

Single ridge's comparable or better high-dimensional performance can be understood through the changing benefit of separate fits. Following your suggestion, we decomposed prediction loss into unrounded prediction loss and rounding variance. As dimension grows, the bucketed method's advantage in unrounded loss disappears and reverses, while its rounding-variance advantage approaches zero. This agrees with the original comparison: when a shared fit already performs similarly or better, the extra flexibility does not improve performance. Single ridge also has lower affine calibration error on shorter Adult/ACS streams. The choice therefore depends on stream length and the measured benefit of separate fits; more features alone do not imply a larger benefit.

Where gains are clear, the extra computation buys substantial error reduction. At T=30,000, bucketed/single-ridge learner times are 6.98/2.85 seconds on Adult and 9.50/2.90 on ACS. A few extra seconds accompany the 52%/40% regret reductions and improved calibration. These comparisons concern the ridge implementations in Tables 7–8.

W2/Q2(a). Which nonlinear errors and feature interactions can be detected?

The theory covers affine corrections representable by the chosen features and permitted coefficient norms within each prediction bucket. Different predictions may use different corrections; constants include ordinary calibration. Low regret means these permitted corrections offer little cumulative squared-loss improvement.

A specific nonlinear relationship or interaction can be encoded as an additional feature. For example, adding age × sex lets the audit test that interaction, subject to feature normalization and coefficient norms. The current Adult/ACS experiments include no explicit interaction features; this describes how to expand coverage. With D features, regret is Õ(T^(1/3)D^(2/3)), normalized L1 calibration error is Õ(T^(-1/3)D^(1/3)), and ONS updates cost O(ND²), plus projection and a stationary-distribution solve. Broader coverage thus has explicit statistical and computational costs.

Q2(b). Which groups have empirical evidence of improved calibration?

The directly evaluated groups are female, male, White alone, Black alone, age≤35 and age>50 on Adult and ACS. The original experiments compared coarse and fine prediction grids; following your suggestion, we reanalyzed the saved prediction traces by group. At T=30,000, the finer grid has lower L1 group calibration error on all 20 seeds in every one of the 12 dataset/group combinations, using the same ten evaluation bins. This compares fine and coarse versions of the bucketed method.

These group results come from direct measurement; theoretical coverage depends on which tests the actual features represent. None of the six raw indicators is exactly affine after the current normalization. The revision therefore reports both the guarantee for permitted affine corrections and the measured calibration improvements for these six groups (Appendix F, Table 10).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.