acceptodds
Under review as a conference paper at ICLR 2027

Contextual Combinatorial Bandits with Post-Serving Contexts

Abstract

We study contextual combinatorial bandits with post-serving contexts (CMAB-PC), where the learner selects of base arms using their pre-serving contexts, while additional contexts that also determine the rewards, such as follow-up engagement in slate recommendation, are revealed only after the selected arms are served. We first propose OFUC-PC, which is jointly optimistic over the reward parameter and the mean post-serving contexts, and analyze it with a new noisy batched truncated elliptical potential lemma (NBT-EPL). OFUC-PC achieves a leading regret of , where is the learnability rate of the post-serving mappings and and are the pre-serving and augmented context dimensions, and a minimax lower bound matches this rate for additive rewards and . We then propose CUCB-PC, an oracle-efficient alternative with the same leading reward-learning term, and variance-aware versions of both algorithms that improve this term from to under variance-modulated smoothness. Experiments on synthetic and MovieLens environments show that our algorithms incur lower regret than the baselines and that the oracle-efficient algorithms are – faster than the joint ones.

Primary area:
learning theory
Keywords:
contextual bandits, combinatorial bandits, post-serving contexts, regret analysis, online learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_CbJdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies contextual combinatorial bandits with post-serving contexts, where multiple arms must be selected before observing additional features that influence their rewards. The authors propose jointly optimistic OFUC$^2$-PC and oracle-efficient C$^2$UCB-PC, using a new noisy batched truncated elliptical potential lemma to analyze uncertainty from reward estimation and post-serving context prediction. The analysis establishes matching regret upper and lower rates, up to logarithmic factors, for additive rewards with context learnability rate α=1/2\alpha=1/2 under the stated parameter conditions, and variance-aware variants sharpen the reward-learning term under variance-modulated smoothness. Experiments on synthetic and MovieLens-based semi-synthetic environments show lower regret than the tested baselines, with oracle-efficient variants running approximately 18–36 times faster than their joint-optimization counterparts.

Critical strengths
  1. Well-motivated problem formulation: Addresses combinatorial decisions where reward-relevant contexts become available only after serving, explicitly accounting for uncertainty that conventional contextual bandits overlook.

  2. Substantive theoretical contribution: Develops a noisy batched truncated elliptical potential lemma that handles noisy contexts and batched feedback without a non-degenerate noise covariance assumption, supporting near-minimax regret guarantees for additive rewards with α=1/2\alpha=1/2 under the stated conditions.

  3. Effective algorithmic alternatives: Provides oracle-efficient and variance-aware variants with strong regret guarantees; synthetic and semi-synthetic experiments demonstrate lower regret than the tested baselines and approximately 18–36× faster execution for oracle-efficient methods than their joint-optimization counterparts.

Critical weaknesses
  1. Restrictive modeling assumptions: The guarantees rely on a shared linear reward model, uniformly controlled post-serving prediction errors, and conditionally independent arm noises. Robustness to model misspecification, correlated feedback, and interactions between selected arms remains unclear.

  2. Limited experimental realism and scale: Evaluation uses synthetic and MovieLens-based semi-synthetic feedback, primarily with only 10 arms and a single reward structure. Runtime experiments extend to just 14 arms, providing limited evidence for large-scale applications or genuine post-serving behavior.

  3. Gap between optimization guarantees and implementation: The joint algorithms assume exact optimistic optimization, while experiments use approximate multi-start SLSQP without quantifying optimization error. The oracle-efficient variants are tested with simple top-KK selection, leaving performance under harder combinatorial constraints unverified.

Questions for authors

See weakness.

2 comments
AuthorsOpenReview ↗

Response to Reviewer CbJd

1. Restrictive Modeling Assumptions

We thank the reviewer for highlighting these important modeling considerations.

Our assumptions follow established contextual and combinatorial bandit frameworks and serve specific purposes in our analysis. The shared linear reward model is a standard formulation in linear contextual bandits, enabling pooled parameter estimation and UCB confidence construction. The learnability conditions, following Wang et al. (2023) and Oh (2026), characterize prediction accuracy without restricting post-serving mappings to a particular parametric form. Conditional independence and zero-mean noise assumptions are commonly adopted in semi-bandit analysis to establish concentration guarantees, particularly for controlling cumulative uncertainty through NBT-EPL.

Our framework accommodates a broad range of settings beyond simple linear and additive models. Specifically, post-serving mappings can be nonlinear as long as the stated learnability conditions hold, while combinatorial reward functions can be nonlinear and non-additive under standard monotonicity and bounded-smoothness conditions. Furthermore, NBT-EPL does not require non-degenerate post-serving noise covariance. The shared linear reward model is used for our current parameter estimation and regret analysis, but parameter sharing is not required by NBT-EPL itself. An extension to arm-specific linear reward parameters is possible in principle.

The primary focus of our work is to address the challenges arising from noisy post-serving contexts and simultaneous combinatorial selection. Our main contributions include the NBT-EPL, which enables new regret guarantees, a matching minimax lower bound under the specified conditions, and the oracle-efficient C$^2$UCB-PC algorithm.

We agree that extending our framework to more general settings, such as misspecified reward models, correlated feedback, and unknown interactions among selected arms, is an interesting direction. These settings introduce additional challenges in parameter estimation and uncertainty analysis that go beyond the scope of the present work. We believe that exploring these more general settings would be a valuable direction for future research.

2. Limited Experimental Realism and Scale

We thank the reviewer for highlighting these concerns. Our additional experiments show that the proposed algorithms maintain their regret and computational advantages with substantially larger candidate-arm pools.

Larger synthetic experiments.

We evaluate all seven methods with m∈10,30,50,100m\in\\{10,30,50,100\\}, K=5K=5, and T=20,000T=20{,}000, under linear and sine mappings with five seeds.

Across all configurations, our four proposed methods achieve lower mean final pseudo-regret than any of the tested baselines.

Large-scale results at m=100m=100, K=5K=5. Final cumulative pseudo-regret (mean ±\pm SE).

MethodLinearSine
X-only LinUCB956.98±29.88956.98\pm29.88779.76±26.39779.76\pm26.39
OFUC$^2$-PC244.67±12.46244.67\pm12.46262.41±11.73262.41\pm11.73
C$^2$UCB-PC356.14±10.20356.14\pm10.20346.63±11.37346.63\pm11.37
VA-OFUC$^2$-PC208.99±7.60208.99\pm7.60211.64±9.50211.64\pm9.50
VA-C$^2$UCB-PC311.48±11.02311.48\pm11.02320.37±10.64320.37\pm10.64

At m=100m=100, the oracle-efficient variants achieve 29.929.9--38.8×38.8\times runtime speedups over their joint-optimization counterparts.

Expanded MovieLens-1M experiment.

We increase the candidate pool from 10 to all 61 movies satisfying the original eligibility threshold of at least 500 construction-user ratings, without relaxing that criterion. With m=61m=61, K=3K=3, T=20,000T=20{,}000, and five seeds, all four proposed methods achieve lower mean final pseudo-regret than any of the tested baselines. The oracle-efficient variants achieve approximately 1010--13×13\times runtime speedups. We acknowledge that our experiments remain semi-synthetic and use a single reward structure. Evaluations with real post-serving feedback and other reward functions remain future work.

We have included the additional results in Appendix I (Figures 7–9) and Appendix J (Figures 11–12) of the revised manuscript for the reviewer's reference.

3. Optimization Guarantees and Implementation

We thank the reviewer for raising these concerns. Our additional experiments show that the SLSQP implementation is empirically stable. Moreover, our theoretical guarantees extend beyond Top-K selection under the stated optimization assumptions, and new experiments provide supporting evidence under partition constraints. We explain these results below.

Theoretical generalization beyond Top-K selection: Our algorithms and regret analysis are not restricted to Top-K selection. They apply to general feasible action families S\mathcal S in which at most KK arms are selected per round. To illustrate this generality, we consider partition constraints.

Let the mm arms be partitioned into KK disjoint groups G1,…,GKG_1,\ldots,G_K. Consider the partition-constrained family

Spart=S⊆[m]:∣S∩Gj∣=1,j∈[K],\mathcal S_{\mathrm{part}} =\\{S\subseteq[m]:|S\cap G_j|=1,\\ j\in[K]\\},

where every feasible action contains exactly KK arms.

For OFUC$^2$-PC, under the joint confidence event, the true reward parameter and mean post-serving contexts belong to their respective confidence sets. Hence,

Ft(St⋆;μt)≤Ut(St⋆)≤Ut(St)=Ft(St;μ~t),F_t(S_t^\star;\mu_t) \le U_t(S_t^\star) \le U_t(S_t) =F_t(S_t;\widetilde\mu_t),

where StS_t maximizes the joint optimistic objective over Spart\mathcal S_{\mathrm{part}}. This argument requires only optimization over the feasible family and does not rely on the Top-K structure.

The regret analysis remains unchanged. The post-serving context-learning bound depends on the total number of selected-arm observations, which satisfies

∑i=1mni=∑t=1T∣St∣≤KT.\sum_{i=1}^{m}n_i =\sum_{t=1}^{T}|S_t| \le KT.

Moreover, NBT-EPL (Lemma 1) applies to any selection policy choosing at most KK arms per round, independently of the feasible family's structure. Therefore, both the context-learning and reward-learning bounds remain valid, recovering the regret guarantee in Theorem 1.

For C$^2$UCB-PC, the monotonicity assumption and an exact combinatorial optimization oracle over Spart\mathcal S_{\mathrm{part}} preserve the optimism argument in Appendix F. Thus, Theorem 3 continues to hold. The variance-aware guarantees in Theorem 4 extend similarly under their additional assumptions.

Consequently, our existing regret upper bounds apply to partition constraints without changing their rates. We further evaluate this theoretical generality through the partition-constrained experiments below.

SLSQP restart sensitivity.

We compare restart caps of 6, 12, and 24 under m=10m=10, K=5K=5, T=20,000T=20{,}000, and five seeds.

Restart sensitivity: final cumulative pseudo-regret (mean ±\pm SE).

MethodMappingCap 6Cap 12Cap 24
OFUC$^2$-PCLinear71.83±4.0971.83\pm4.0972.55±4.1172.55\pm4.1172.43±4.0372.43\pm4.03
Sine77.66±5.1077.66\pm5.1077.57±5.2377.57\pm5.2377.77±5.3677.77\pm5.36
VA-OFUC$^2$-PCLinear53.71±2.4253.71\pm2.4254.36±2.5254.36\pm2.5254.29±2.5354.29\pm2.53
Sine58.12±3.5658.12\pm3.5657.42±3.4357.42\pm3.4357.39±3.5157.39\pm3.51

The final pseudo-regret shows no consistent improvement with additional SLSQP restarts. Increasing the cap from 6 to 24 raised runtime by 19.619.6--31.331.3\\%, while the paired 95% confidence intervals for cumulative regret differences included zero.

These results support empirical stability, although they do not certify global optimality.

Beyond Top-KK selection.

We additionally evaluate all seven algorithms under partition constraints with m=50m=50 arms divided into five groups of ten, selecting exactly one arm per group.

The feasible family is

\mathcal S_{\mathrm{part}} = \left\\{ S\subseteq[m]: |S\cap G_j|=1,\\; j=1,\ldots,5 \right\\}.

Partition-constrained results (m=50m=50, K=5K=5). Final cumulative pseudo-regret (mean ±\pm SE).

MethodLinearSine
X-only LinUCB772.34±32.88772.34\pm32.88609.62±26.90609.62\pm26.90
OFUC$^2$-PC166.54±10.42166.54\pm10.42171.59±9.08171.59\pm9.08
C$^2$UCB-PC248.73±14.37248.73\pm14.37248.02±11.19248.02\pm11.19
VA-OFUC$^2$-PC134.21±7.92134.21\pm7.92141.15±10.95141.15\pm10.95
VA-C$^2$UCB-PC211.38±12.91211.38\pm12.91216.68±13.01216.68\pm13.01

Across both environments, all four proposed methods achieve lower mean cumulative pseudo-regret than the tested baselines. All 24 paired 95$\%$ confidence intervals exclude zero, supporting the observed improvements. The oracle-efficient methods are 25.7–32.2 times faster than their joint counterparts. These results show that our methods also work beyond Top-KK selection. Testing more general combinatorial constraints remains future work.

We have included the additional results in Appendix I (Figures 5–6) and Appendix K (Figures 13–15) of the revised manuscript for the reviewer's reference.

We would be happy to provide further clarification or address any additional questions you may have.

Reviewer_72cHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies contextual combinatorial bandits with post-serving contexts, extending the framework of Wang et al. (2023) to combinatorial action selection. The authors propose OFUC2-PC based on joint optimism and C2UCB-PC as a computationally efficient alternative. They establish regret bounds using a new noisy batched truncated elliptical potential lemma (NBT-EPL), together with matching lower bounds for additive rewards. Variance-aware extensions and numerical experiments are also presented.

Critical strengths

The paper considers a meaningful extension of contextual bandits with post-serving contexts to combinatorial actions, with potential applications in recommendation systems. The theoretical analysis is relatively comprehensive, including regret upper and lower bounds.

Critical weaknesses
  1. The proposed framework appears to largely combine the post-serving contextual bandit model of Wang et al. (2023) with existing contextual combinatorial bandit techniques. OFUC2-PC follows the joint optimism approach of Wang et al. (2023), while C2UCB-PC extends the standard combinatorial UCB principle. The variance-aware extensions also build on established techniques. The main technical novelty seems to be the NBT-EPL. It would be helpful to better clarify which difficulties cannot be addressed by directly combining the generalized elliptical potential lemma of Wang et al. (2023) and the truncated elliptical potential analysis of Takemura et al. (2021).

  2. Assumptions 1 and 2 require estimators of the unknown post-serving mappings with high-probability prediction guarantees. These assumptions are strong on post-serving context learning, where it basically implies the optimistic learning of post context feature mapping.

  3. The experiments compare against a plug-in LinUCB baseline adapted from Wang et al. (2023), but not against a combinatorial adaptation of their stronger poLinUCB algorithm. Such a comparison would help clarify the algorithmic improvements beyond existing post-serving contextual bandit techniques.

Questions for authors

Please see the weakness above.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 72cH

1. NBT-EPL Novelty

We thank the reviewer for the thoughtful comments on the novelty of NBT-EPL and the opportunity to clarify its technical contributions.

The main theoretical difficulty is not merely the coexistence of noisy contexts and combinatorial actions, but controlling their interaction in the cumulative confidence widths.

There are two distinct obstacles:

First, noisy realized versus mean contexts. The reward design matrix is updated using noisy realized augmented contexts ut,iu_{t,i}, whereas regret depends on confidence widths evaluated at their conditional means uˉt,i\bar u_{t,i}. Thus, standard potential arguments cannot be applied directly.

Second, simultaneous batched selection. All KK arms in a round are selected using the same pre-round matrix At−1A_{t-1}. Sequentially updating the matrix within the round would not represent the uncertainties used by the algorithm when making its simultaneous selection. Wang et al. (2023) address noisy contexts for single-arm selection under a non-degenerate noise covariance condition. Takemura et al. (2021) address batched selection through truncation, but do not account for the discrepancy between realized and conditional-mean contexts. Neither result can therefore be substituted directly into the other to obtain the required bound.

Our proof introduces a predictable truncation of realized confidence widths and aggregates them within each round. Using the conditional mean-zero property of post-serving noise and conditional independence across selected arms, we establish the key inequality

\left[ W_t\mid\mathcal{F}_t^{\mathrm{pre}} \right] \ge (1-e^{-1})X_t,$$ where $X_t$ is the sum, over all selected arms $i \in S_t$, of the minimum of $1/K$ and the squared confidence width of the conditional-mean augmented context for arm $i$, computed using the inverse of the pre-round design matrix $A_{t-1}$. We then combine this inequality with batched log-determinant increments and a nonnegative supermartingale argument, yielding $$\Gamma_T \le \frac{2}{1-e^{-1}} \left( \log\frac{\det A_T}{\det(\lambda I)} + \log\frac{1}{\delta_E} \right).$$ Importantly, unlike Wang et al. (2023), our NBT-EPL does not require non-degenerate post-serving noise covariance, and its bound contains no inverse dependence on the smallest eigenvalue of the noise covariance matrix. It therefore remains valid even when the post-serving noise is degenerate. This is the technical distinction, and we emphasize it in Section 3.3 and Appendix D. More broadly, NBT-EPL provides a general tool for analyzing batched arm selection with noisy post-serving contexts. It may support future work on new combinatorial selection strategies and more complex action constraints, even when the post-serving noise covariance is degenerate. ### 2. Learnability Assumptions We thank the reviewer for raising this concern. Assumptions 1 and 2 follow the learnability frameworks of Wang et al. (2023) and Oh (2026), respectively. We intentionally adopt these general conditions to accommodate different post-serving mapping classes and learning methods, rather than restricting our analysis to a particular estimator. For standard linear mappings, regularized least squares satisfies Assumption 1 with $\alpha=1/2$ under appropriate conditions. As discussed by Wang et al. (2023), more complex mappings can be studied using nonparametric methods, such as nearest-neighbor regression, manifold regression, and estimation of Hölder-smooth functions. These methods may exhibit slower convergence rates, which can be accommodated by our framework whenever the corresponding high-probability prediction guarantees hold. In particular, Assumptions 1 and 2 allow arbitrarily small positive exponents $\alpha$ and $a$, respectively, reflecting increasingly difficult learning problems. These slower learning rates are explicitly reflected in our regret bounds. The context-learning regret scales as $\widetilde O(T^{1-\alpha})$ under Assumption 1 and $\widetilde O(T^{1-a})$ under Assumption 2. Thus, any fixed positive exponent retains sublinear dependence on $T$, while the bound approaches linear regret as the exponent approaches zero. We acknowledge that arbitrary mappings may not satisfy these conditions. Our guarantees apply to learnable post-serving mappings, rather than all possible mappings.

3. Missing Baseline

We thank the reviewer for highlighting the connection to poLinUCB. We clarify that poLinUCB selects a single arm through joint optimism over the reward parameter and post-serving context, while OFUC$^2$-PC (Algorithm 1) extends the same mechanism to combinatorial super-arm selection. Therefore, we do not include poLinUCB as a separate experimental baseline, as its combinatorial adaptation is already represented by OFUC$^2$-PC.

However, the main theoretical challenge is to control cumulative confidence widths under both noisy observations and simultaneous arm selection. Standard elliptical potential arguments cannot directly provide the required bound. Simply combining Wang et al. (2023) and Takemura et al. (2021) is also insufficient, as neither analysis handles both challenges together.

To overcome these difficulties, we develop NBT-EPL, which avoids the non-degenerate post-serving noise covariance assumption and the σε−4\sigma_\varepsilon^{-4} dependence of Wang et al. (2023). The lemma applies to general batched arm-selection policies under its stated conditions and enables combinatorial regret guarantees matching our minimax lower bound for the specified additive subclass and parameter regime.

Furthermore, we propose C$^2$UCB-PC (Algorithm 2) to overcome the computational cost of joint optimization. It requires only one combinatorial oracle call per round while retaining provable regret guarantees, with substantial runtime improvements demonstrated experimentally. Moreover, under Assumption 5, we establish corresponding variance-aware regret guarantees.

Thus, our contributions extend beyond the joint-optimism principle to new theoretical tools, minimax regret guarantees under the stated conditions, and an oracle-efficient algorithm.

We would be happy to provide further clarification or address any additional questions you may have.

Reviewer_u5wmOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies contextual combinatorial semi bandits: each arm provides a set of reward relevant features when pulled. The main technical contribution of the paper is a noisy batched truncated elliptical potential lemma, which enable proving regret bounds fior joint optimism ; and a version that requires only one combinatorial optimization oracle per iteration. Under the assumptions in the paper; the authors prove a regret upper bound and a matching regret lower bound (up to logs); this holds for learnability rate 1/2 and additional conditions in Thm 2. Simulations results are provided to show the regret/runtime tradeoffs.

Critical strengths
  • the paper resolves an important difficulty in learning from information observed after the decision. It addresses a practical scenario where standard regret analysis fails: the learner selects a set of items before observing their post serving features; the algorithm learns from information revealed after an item is selected, such as watch time of a movie, where in the following iterations those features aren't observed before making the decision. I find this an important result.
  • The authors further give an efficient algorithm with the same regret bound guarantee. The first algorithm searches over possible reward parameters, missing contexts, and sets of items, which can be computationally infeasible. The second algorithm instead assigns each item an optimistim score and makes one call to an oracle that selects the best feasible set given those scores.
  • the paper also provides a matching lower bound under some assumptions.
Critical weaknesses
  • The paper does not establish that the proposed variance weighting is the cause of improved regret guarantee. Section 5 presents additional algorithms that give different weights based on estimated uncertainty. However, this section imposes stronger assumptions on the reward function and requires expected rewards to be bounded away from 0 and 1. Under those assumptions, the same improvement can follow for the simpler unweighted algorithm. So the analysis does not distinguish the benefit of variance weighting from the benefit of stronger assumptions. Additionally the claim that the new guarantee is never worse needs qualification as the bound depends on ν\nu (distance of bound from 0/1).
  • The experiments don't establish improvement over the closest existing post serving algorithm. The experiments compare against a method that ignores the post serving features and ones substitutes predicted features into reward estimation. They omit the one from poLinUCB (an adapted version to select multiple arms), which uses post serving observations.
Questions for authors

Please see the weaknesses.

1 comment
AuthorsOpenReview ↗

Response to Reviewer u5wm

1. Variance-Aware Improvement

We thank the reviewer for highlighting this important issue. We acknowledge that while our variance-aware regret guarantees remain valid under the stated assumptions, the current analysis does not establish that variance weighting alone is responsible for the improved regret bound. In addition, to avoid overstating the comparison between the regret bounds, we have replaced “the new term is never worse” with “so the original K,TK,T scaling is recovered for fixed ν\nu” in Section 5 (Discussion). This revision is highlighted in blue in the revised manuscript.

Our variance-aware results still provide a useful theoretical extension. Building on the variance-modulated smoothness framework of Liu et al. (2022) and the optimistic-variance approach of Liu et al. (2023), we show that variance-aware regret bounds can be established in our setting under the stated assumptions. For future work, we aim to better understand whether variance weighting itself can improve regret bounds and whether tighter guarantees can be obtained under weaker assumptions.

Moreover, we respectfully emphasize that our main contributions are independent of variance weighting. In particular, our NBT-EPL, regret guarantees for OFUC$^2$-PC and C$^2$UCB-PC, minimax lower bound under the stated conditions, and oracle-efficient algorithm do not rely on variance weighting or Assumption 5. Therefore, variance-aware methods serve as an additional extension rather than a necessary component of our main framework.

2. Missing Baseline

We thank the reviewer for highlighting the connection to poLinUCB. Wang et al. (2023) proposed poLinUCB, which selects a single arm by jointly optimizing over the reward parameter and post-serving context within their confidence sets. Our OFUC$^2$-PC (Algorithm 1) directly extends this mechanism to combinatorial super-arm selection by jointly optimizing over a feasible set of arms, a shared reward parameter, and the post-serving contexts of all selected arms. Therefore, we do not include poLinUCB as a separate experimental baseline, as its combinatorial adaptation is already represented by OFUC²-PC.

However, the main theoretical challenge is to control cumulative confidence widths under both noisy observations and simultaneous arm selection. Standard elliptical potential arguments cannot directly provide the required bound. Simply combining Wang et al. (2023) and Takemura et al. (2021) is also insufficient, as neither analysis handles both challenges together.

To overcome these difficulties, we develop NBT-EPL, which avoids the non-degenerate post-serving noise covariance assumption and the σε−4\sigma_\varepsilon^{-4} dependence of Wang et al. (2023). The lemma applies to general batched arm-selection policies under its stated conditions and enables combinatorial regret guarantees matching our minimax lower bound for the specified additive subclass and parameter regime.

Furthermore, we propose C$^2$UCB-PC (Algorithm 2) to overcome the computational cost of joint optimization. It requires only one combinatorial oracle call per round while retaining provable regret guarantees, with substantial runtime improvements demonstrated experimentally.

Thus, our contributions extend beyond the joint-optimism principle to new theoretical tools, minimax regret guarantees under the stated conditions, and an oracle-efficient algorithm.

We would be happy to provide further clarification or address any additional questions you may have.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.