acceptodds
Under review as a conference paper at ICLR 2027

Exploring Multiple High-Scoring Subspaces in Generative Flow Networks

Abstract

Generative Flow Networks (GFlowNets) construct combinatorial objects through sequences of elementary actions, but exploration in large spaces can spend much of a finite sampling budget on low-reward regions. We propose CBFlowNet, an adaptive subspace exploration framework that periodically selects subsets of action types to guide trajectory collection. Its controller learns a set-level reward predictor from the mean terminal rewards collected under selected subsets. Policy-conditioned arm statistics, obtained from unrestricted evaluation, serve as auxiliary features rather than sufficient descriptions of subspace value. A Deep Sets scorer combines these features with action identities to model interactions, while bandit-inspired exploration encourages alternative selections as the GFlowNet evolves. Our analysis establishes trajectory-support and zero-loss target-consistency properties and decomposes the errors affecting current-policy set selection. Experiments on molecular, RNA, and bit-sequence design demonstrate improvements in high-reward candidate discovery. Implementations are available at https://anonymous.4open.science/r/ICLR-41377.

Primary area:
reinforcement learning
Keywords:
Reinforcement learning, Generative flow networks

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 2 reviews, above 27% of the others.
Reviewer_KpumOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces CBFlowNet, a method for improving exploration in GFlowNets under limited sampling budgets. It uses a learned controller to periodically select subsets of action types, based on set-level reward predictions and a bandit-style exploration strategy. The selected subsets guide trajectory collection toward different subspaces during training. The paper also provides theoretical analyses of trajectory-support preservation, zero-loss target consistency, and subspace-selection errors. Experiments on molecular design, RNA design, and bit-sequence generation show improvements in finding high-reward candidates.

Critical strengths
  1. Strong empirical results in molecular mode discovery. CBFlowNet discovers approximately 41% more high-reward modes than QGFN under the matched total reward-evaluation budget reported in Table 5. This is a useful comparison, as it suggests that the improvement is not simply due to additional reward evaluations.

  2. A reasonable approach to learning action-subset utilities. Rather than relying only on individual action scores, CBFlowNet uses set-level reward feedback to learn which action combinations are worth exploring. The ablations in Figure 6 show that Deep Sets-based selection consistently outperforms Top-K and random selection across the tested arm-scoring variants, suggesting that learned subset selection is more effective than these simpler alternatives.

Critical weaknesses
  1. Unclear generalization from complete to partial sets. The Deep Sets scorer is trained on complete sets but used to score partial sets during greedy subset construction (Appendix F). Good predictions on complete sets do not necessarily imply reliable rankings of partial sets, which are what the selection procedure actually depends on. The authors acknowledge this issue, but there is no direct evaluation of these intermediate rankings. The theoretical analysis accounts for the effects of partial-set queries in the selection residual $e_t$, without establishing that this residual is small. The main concern is therefore whether the scorer can reliably guide subset construction in practice.

  2. The ablations do not fully isolate the benefit of interaction-aware and adaptive selection. Deep Sets is trained with set-level reward labels, while Top-K relies on individual arm scores. It is therefore unclear whether modeling non-additive interactions provides a benefit beyond learning from set-level supervision. Also, Appendix E.1 shows that a fixed four-arm subspace already contains all target modes in the bit-sequence task, making it difficult to assess the benefit of switching subspaces in this benchmark. The matched-budget molecular results support the overall method’s effectiveness, but do not by themselves resolve these mechanism-level attribution questions.

Questions for authors
  1. Given the complete-to-partial-set mismatch discussed in Appendix F, is there empirical evidence that scores on partial sets reliably guide greedy construction toward high-quality complete sets? For example, at frozen policy checkpoints, how does the sampling yield of the greedily constructed set compare with that of the best feasible complete set under the same cardinality constraints? Could you report this comparison both with and without the dropout exploration bonus?

  2. Would Deep Sets still outperform an additive predictor trained with the same set-level supervision protocol (Sk,vk,Yk)(S_k, v_k, Y_k) and using the same action identities and auxiliary features, with exploration settings and reward-evaluation budgets matched? The current Top-K comparison differs in both the use of set-level supervision and the ability to model action interactions, leaving unclear whether non-additive interaction modeling provides a benefit beyond learning from set-level feedback.

  3. How is the fixed subset chosen in the Hard Pruning baseline? Are the subset size, gating and exploration settings, and total reward-evaluation budget matched to CBFlowNet? In particular, does Hard Pruning freeze a subset selected by the same learned procedure after a comparable warm-up, or does it use a different selection rule?

10 comments
AuthorsOpenReview ↗

Response to Reviewer Kpum(part 1/2)

Dear Reviewer Kpum Thank you for your valuable comments, which are crucial to improve of our paper. We first summarize the proposed weaknesses and questions(we noticed that some weaknesses and questions are alike, but some weaknesses contain several parts to clarify).

W1. Unclear generalization from complete to partial sets. Is there empirical evidence that scores on partial sets reliably guide greedy construction toward high-quality complete sets?

Thank you for highlighting this issue. We directly evaluated the rankings of candidate additions at partial sets encountered during greedy construction. With the GFlowNet, Deep Sets scorer, and arm features frozen(from iteration 50000), we examined prefixes of sizes 5, 10, 20, and 29. For each prefix, we evaluated ten candidate additions, including the greedy choice. Each resulting partial set was completed to 30 blocks with the same suffixes, and each completion was evaluated using 400 samples under the same sampling settings. We report the within-prefix Spearman correlation between prediction means:

ρ=1−6∑i=110(r^i−ri)210(102−1).\rho=1-\frac{6\sum_{i=1}^{10}(\hat r_i-r_i)^2}{10(10^2-1)}.

Higher values indicate better agreement between the two rankings.

Prefix Size5102029
ρ\rho0.580.680.770.81

The positive correlations across all evaluated prefix sizes indicate that higher predicted scores tend to correspond to higher completion rewards, supporting their usefulness for guiding greedy construction. Ranking agreement increases from 0.58 at prefix size 5 to 0.81 at size 29, suggesting more accurate rankings as the partial sets approach the target size.

Q1. How does the sampling yield of the greedily constructed set compare with that of the best feasible complete set under the same cardinality constraints with and without exploration bonus.

According to your suggestion, we compare the constructed complete set with the best found reference set in molecule generation task. At a frozen policy checkpoint, we perform multi-start swap search over complete 30-block sets, initialized from random, Top-30, and greedy selections. We evaluate 2,000 candidate sets using GFN sampled candidates for each set, which provides a high-budget best-found reference. We compare the best set with greedy selections with and without the dropout bonus, the results is shown below. We report the mean transformed training reward, mean raw proxy score, and high-score yield, defined as the fraction of evaluated molecules whose raw proxy score exceeds 7.5.

Subspace selectionMean rewardMean proxy scoreHigh-score yield
Random0.10855.52360.00%
With exploration bonus0.52287.237648.75%
Without exploration bonus0.53577.276450.25%
Best-found subspace0.54167.291849.75%

Greedy selection without the exploration bonus achieves a mean reward close to the best-found reference (0.5357 vs. 0.5416), with comparable proxy scores and high-score yield (50.25% vs. 49.75%). Both greedy variants substantially outperform random selection.

W2 and Q2. Would Deep Sets still outperform an additive predictor trained with the same set-level supervision protocol.

Thank you for highlighting this confounding factor. To address this concern, we conducted an additional experiment on molecular generation with an additive predictor trained using the same set-level supervision, action identities, and auxiliary features as Deep Sets, with matched exploration settings and reward-evaluation budgets. All other experimental settings were kept unchanged. This baseline learns from set-level feedback while restricting its predictions to additive contributions from individual actions.

MethodModes with R > 7.5Modes with R > 8Top-1000 reward
Deep Sets14,8333,0128.432
Additive Predictor13,1192,5568.408
Top-K12,8192,4238.399

The additive predictor slightly improves over Top-K. Deep Sets retains an advantage over the additive predictor under the matched protocol: it discovers 13.1% more modes with R > 7.5 and 17.8% more modes with R > 8, with a smaller improvement in Top-1000 reward.

These results provide preliminary evidence that non-additive interaction modeling offers benefits beyond set-level supervision alone, particularly for discovering high-reward modes in this task.

AuthorsOpenReview ↗

Response to Reviewer Kpum(part 2/2)

W2 part 2. A fixed four-arm subspace already contains all target modes in the bit-sequence task, making it difficult to assess the benefit of switching subspaces in this benchmark.

Thank you for pointing this out. We would like to clarify two points.

First, this complete coverage results from the deliberate construction of the toy environment(bitseq), which serves as an illustrative demonstration. In molecular generation, there is no corresponding guarantee built into the task that a small, fixed action subspace contains all high-reward modes. The coverage property of this toy example therefore does not necessarily generalize to the molecular setting.

Second, complete coverage does not imply efficient discovery. Within a given subspace, some modes may be considerably easier for the GFlowNet to discover than others. After the more accessible modes have been found, discovering additional high-reward modes may become increasingly difficult. A mode that is difficult to discover in one subspace may be more accessible in another, where different available actions provide alternative generation trajectories. Switching subspaces can therefore offer complementary exploration opportunities and potentially improve mode discovery under a fixed reward-evaluation budget, even when a single subspace contains all target modes.

Q3. How is the fixed subset chosen in the Hard Pruning baseline? Are the subset size, gating and exploration settings, and total reward-evaluation budget matched to CBFlowNet? In particular, does Hard Pruning freeze a subset selected by the same learned procedure after a comparable warm-up, or does it use a different selection rule?

Thank you for raising this point. In the original Hard Pruning baseline, we uniformly sample 30 actions from the full action set and keep this subset fixed. Thus, the subset is selected randomly, rather than by CBFlowNet’s learned selection procedure.

To directly assess the benefit of continued adaptive subspace selection, we introduce a controlled Warm-up Hard Pruning baseline. We first train CBFlowNet for 10,000 steps, during which the Deep Sets scorer is also trained. We then fork the same trained checkpoint into two branches:

CBFlowNet (continued adaptive selection): continues training with adaptive subspace selection.

Warm-up Hard Pruning: freezes the (K=30) subset selected by the learned Deep Sets scorer at the branching point and continues training within this fixed subset.

The two branches therefore share exactly the same warm-up training history and learned subset-selection procedure. We match the subset size (K=30), hard-gating strategy, exploration parameter (ϵ\epsilon), and total reward-evaluation budget across methods, with all other training settings held constant. The 10,000-step warm-up, including the reward evaluations used to train the scorer, is included in each branch’s total budget. Both branches consequently receive the same remaining reward-evaluation budget after the split.

MethodModes with R > 7.5Modes with R > 8Top-1000 reward
Deep Sets14,8333,0128.432
Warm-up Hard-pruning50167898.379
Hard Pruning24673158.262

Warm-up Hard-pruning outperforms random Hard Pruning, confirming the benefit of reward-informed subset selection. Nevertheless, Deep Sets discovers almost 3 times and 4 times as many modes at R>7.5 and R>8, respectively, and achieves a higher top-1000 reward. These results support the benefit of continued adaptive subspace selection over one-time pruning(W2 part2). They also complement our response to W2: subspace coverage alone does not guarantee efficient mode discovery, and switching subspaces can provide complementary exploration opportunities.

Reviewer_KpumOpenReview ↗

Response to the Revision

Thank you for the clarifications, which address my original concerns. I have two further questions about the experimental settings and results.

  1. Table 7 and Appendix H specify β=8\beta=8 for the RNA experiments. However, the CBFlowNet mode-discovery curves for RNA1, RNA2, and RNA3 in Figure 4 appear to match the curves labeled β=20\beta=20 in Figures 9, 10, and 11 at every plotted point. The RNA1 and RNA2 average-reward curves also appear to match their corresponding β=20\beta=20 curves.

    Since β\beta changes the training target Rtar(x)=R(x)βR_{\mathrm{tar}}(x)=R(x)^\beta, could you clarify:

    • Which β\beta was used for each method in Figure 4, and whether any table entries, figure labels, or descriptions need correction?
    • Which configurations, seeds, and runs produced the CBFlowNet curves in Figure 4 and Figures 9 to 11?
    • Whether all methods used the same β\beta or were tuned separately under comparable protocols? If tuned separately, please describe the protocol.
  2. Algorithm 1 places full-vocabulary initialization before the TT training steps. If Table 4's 100,000 steps refer to this subsequent loop, the batch size of 4 would already account for Table 5's 400k training evaluations. Does the 500k total include both training and unrestricted-evaluation calls during initialization, and what stopping rule enforces this budget?

    Please also clarify how Appendix J's "100 evaluations per 400 sampling steps" corresponds to the 100k bandit evaluations in Table 5. Please specify what "sampling steps" means, the evaluation batch size bevb_{\mathrm{ev}}, and the evaluation frequency, along with which samples contribute to the reported mode counts and Top-1000 reward.

    For QGFN, the strongest baseline in Table 5, please report the variant, pp, nn-step setting, any relevant schedule, reward exponent, and tuning protocol.

The additional experiments have strengthened the paper. Clarifying these points would help me decide whether to recommend acceptance.

Reviewer_bjySOpenReview ↗
Rating
1(−1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes CBFlowNet, an adaptive exploration strategy for GFN, inspried from combinatorial multi-armed bandits. It periodically selects a super arm ( a subset of the complete action types) , and construct gated policy based on historically data. By mixing the gated policy with unform policy for data-collection, the author seek to better exploration for data-collection of GFlowNet learning. And the training objective still follows the classic Trajectory Balance (TB) objective. Experiments on molecular design, three RNA tasks, and bit-sequence generation report improved high-reward discovery.

Critical strengths
  1. The method addresses a practical exploration problem from a new perspective, the combinatorial multi-armed bandit.
  2. The molecular results provide encouraging empirical evidence about high reward discovery.
Critical weaknesses
  1. The reward label YkY_k, which is the averge of the cumulative reward of a epoch interval, keeps changing as GFN training progresses. Then, the Deep Sets predictor is trained based on reward, to inform the construction of new subset. Consequently, successful prediction quility of current subset quality are not assured and can be highly unreliable. So is the following selection procedure for new subset.

  2. Section 4, which describes the main contribution of this paper, is not well organized, resulting in poor readability. For example, Eq. (5) is introduced without explaining why it is needed or how it is used. The definition of a subset and its construction procedure are separated between section 4.1.1 and section 4.1.6. At line 201, the authors state that “two sampling phases supply different controller inputs,” but I could not find clear definitions of the controller or the sampling procedures around this statement. Similarly, the training label for the Deep Sets scorer is introduced in section 4.1.3, but its connection to the scorer is not explained until section 4.1.6, where the scorer itself is introduced.

  3. The presentation of the proposed algorithm in Algorithm 1 is very unclear. Considerable detail is hidden behind operations such as GFLOWNET.TRAIN and DeepSetsGreedySelect, and I could not find any explanations of what these operations involve throughout the paper.

  4. The theoretical analysis in Appendix B provides limited justification for the effectiveness of the proposed algorithm. Although the authors acknowledge its limitations, these substantially restrict the explanatory value of the theoretical contribution.

  5. The main experiments use hard gating, whereas the main text (section 4.1.2) describe soft gating, which as the one of main component should be fouced.

  6. The goal of GFN is to achieve sampling complying with R(x)/∑xR(x)R(x)/\sum_x R(x). It is a problem of distribution matching, rather than mode discovery of RR or max⁡xR(x)\max_x R(x). Direct experiments and metrics measuring the discrepancy between the learned terminal distribution and the target distribution are missing.

Questions for authors
  1. In Algorithm 1, could you provide complete pseudocode for the subset-selection procedure and make the connections between the equations of section 4.1.2-4.1.6 explicit?

  2. How to compute the heuristic Monte Carlo dropout bonus σ^ϕ\hat\sigma_\phi?.

  3. What is purpose of introducing (4)? It seems to be irrelevant to the proposed Algorithm.

  4. Could you provide direct evidence that the Deep Sets scorer produces useful rankings for the partial subsets queried during greedy construction?

8 comments
AuthorsOpenReview ↗

Response to Reviewer bjyS(Part 1/2)

Dear Reviewer bjyS, Thank you for your valuable comments, which are crucial to improve of our paper. We first summarize the proposed weaknesses and questions. We would like to clarify your concerns point by point in the following.

W1.The reward label, which is the averge of the cumulative reward of a epoch interval, keeps changing as GFN training progresses. Whether the subset chosen by Deep Sets can reflect the true quality?

We agree that the interval-level reward labels are non-stationary because the GFlowNet policy evolves during training. However, our formulation explicitly accounts for this dependence: subspace quality is evaluated with respect to the current policy, and our analysis includes policy drift and the mismatch between interval-level supervision and current-policy utility. The Deep Sets scorer is updated throughout training using newly collected subspace-level feedback.

To directly assess whether Deep Sets predictions reflect subspace quality, we conducted an additional experiment on the molecular generation task with 16 different subspaces. At each evaluated training checkpoint, we generated 4,000 samples under each subspace using the GFlowNet at that checkpoint. We used these samples to estimate subspace quality and construct an empirical reference ranking, which we compared with the ranking predicted by Deep Sets using Spearman’s rank correlation:

ρ=1−6∑i=116(r^i−ri)216(162−1).\rho=1-\frac{6\sum_{i=1}^{16}(\hat r_i-r_i)^2}{16(16^2-1)}.

Higher values indicate better agreement between the two rankings.

Training iterationsSpearman’s ρ\rho
1,0000.25
5,0000.65
10,0000.81
50,0000.85

The correlation increased from 0.25 at 1,000 training iterations to 0.65 at 5,000 iterations, reaching 0.81 at 10,000 iterations. These results demonstrate increasingly strong agreement between the predicted and empirical rankings, providing evidence that Deep Sets can learn to identify promising subspaces despite the non-stationary reward labels.

W2. Presentation organization.

Thanks for pointing this out. We have reorganized Section 4 to introduce the controller and both sampling procedures upfront. Subset definition and construction, together with the Deep Sets scorer and its training label, are now presented in Section 4.1.1.

W3. Considerable detail is hidden behind operations such as GFLOWNET.TRAIN and DeepSetsGreedySelect.

Thanks for pointing this out. We have revised our manuscript. Algorithm 1 now explicitly describes sampling, GFlowNet updates, evaluation, and scorer training. It also includes the greedy selection procedure: start with the highest-scoring arm, then repeatedly add the arm maximizing the candidate set’s prediction mean plus dropout standard deviation until KK arms are selected. Appendix F provides further details.

W4. Theorectical limitation.

We have already acknowledge this limitation in our original manuscript, Appendix B does not establish an end-to-end performance guarantee. However, it provides concrete insights, Proposition B.1 establishes support preservation and zero-loss TB target consistency. Lemma B.2 and Theorem B.3 characterize how statistical uncertainty, policy drift, feedback bias, prediction error, and search error affect subspace selection. In particular, the analysis explains why accurate arm statistics alone do not guarantee good set selection. Although the uncontrolled residuals limit the performance guarantee, these results clarify the method’s properties and potential failure sources. Practical effectiveness is supported separately by the selection ablations and the matched total reward-evaluation comparison in Appendix G.1.

W5. The main experiments use hard gating, whereas the main text describe soft gating, which as the one of main component should be focused.

Thanks for the suggestion. We have revised our manuscript. Section 4.1.2 now defines both variants, presenting hard gating as the default used in the main experiments and soft gating as an alternative, with their shared exploration mechanism.

Q1. pseudocode for the subset-selection procedure and make the connections explicit?

Thank you for your suggestion. We have revised our manuscript accordingly. We have expanded Algorithm 1 to explicitly describe greedy subset selection, including initialization and iterative arm addition based on the prediction mean plus the dropout bonus. Section 4 now clarifies how unrestricted evaluation provides arm features, interval rewards train the Deep Sets scorer, and the scorer selects subsets that guide subsequent trajectory collection.

Q2. How to compute the heuristic Monte Carlo dropout bonus?

Thanks for your question. It is given in Appendix F in our original manuscript. For each candidate set, we run eight forward passes with different dropout masks while fixing the model parameters and input features. Their standard deviation is the exploration bonus, added to their mean prediction.

AuthorsOpenReview ↗

Response to Reviewer bjyS(Part 2/2)

W6. GFN is a problem of distribution matching, rather than mode discovery. Direct experiments and metrics measuring the discrepancy between the learned terminal distribution and the target distribution are missing.

We agree that distribution matching is a defining objective of GFlowNets and that mode-discovery metrics alone do not establish distributional accuracy. Our work focuses on improving finite-budget discovery of diverse high-reward candidates, a practical objective also studied in LSGFN [1] and QGFN [2]. In CBFlowNet, subspace guidance modifies trajectory collection, while the underlying GFlowNet retains its reward-proportional training objective.

Following your suggestion, we added an enumerable 3D HyperGrid experiment with horizon H=16H=16. Enumerating the terminal states allows us to compute the exact reward-proportional target PR(x)=R(x)/ZP_R(x)=R(x)/Z, where Z=∑xR(x)Z=\sum_x R(x). We evaluate the generated terminal distribution PgenP_{\mathrm{gen}} using

Global TV=12∑x∣Pgen(x)−PR(x)∣.\mathrm{Global\ TV} = \frac{1}{2}\sum_x |P_{\mathrm{gen}}(x)-P_R(x)|.

We additionally report the signed probability-mass error over the high-reward mode region AA:

ΔHigh-R=∑x∈APgen(x)−∑x∈APR(x).\Delta_{\mathrm{High\text{-}R}} = \sum_{x\in A} P_{\mathrm{gen}}(x) - \sum_{x\in A} P_R(x).

Positive values indicate excess probability allocated to this region, negative values indicate insufficient allocation, and values closer to zero indicate more accurate aggregate probability allocation.

For CBFlowNet, we distinguish two evaluation settings using the same trained checkpoint: guided sampling applies subspace guidance, whereas ungated sampling disables this guidance and samples over the full action space. The ungated setting requires no retraining.

MethodGlobal TVΔ\Delta High-R Mass
TB0.009-0.007
QGFN0.045+0.019
CBFlowNet (guided sampling)0.021+0.013
CBFlowNet (ungated sampling)0.007+0.003

Ungated CBFlowNet achieves a Global TV distance of 0.007, comparable to TB's 0.009, with a high-reward mode-mass error of only +0.003. These results show that the learned full-space sampler closely approximates the reward-proportional target in this experiment.

Applying subspace guidance to the same checkpoint increases Global TV to 0.021 and allocates an additional probability to the high-reward region relative to the target. Guided CBFlowNet nevertheless has a lower Global TV distance than QGFN in this comparison.

[1] Kim, Minsu, et al. "Local search gflownets." International Conference on Learning Representations. 2024. [2] Lau, Elaine, et al. "Qgfn: Controllable greediness with action values." Advances in Neural Information Processing Systems 37 (2024): 81645–81676.

Q3. What is purpose of introducing (4)? It seems to be irrelevant to the proposed Algorithm.

Thanks for the question. It is relevant to the proposed algorithm. Original Eq. (4), now Eq. (3), defines the interval-level reward used to train Deep Sets through mean-squared error. The trained scorer then guides subset selection. This connection is explicit in Section 4.1.1 and Algorithm 1(line 15).

Q4. Could you provide direct evidence that the Deep Sets scorer produces useful rankings for the partial subsets queried during greedy construction?

Thank you for highlighting this issue. We directly evaluated the rankings of candidate additions at partial sets encountered during greedy construction. With the GFlowNet, Deep Sets scorer, and arm features frozen(from iteration 50000), we examined prefixes of sizes 5, 10, 20, and 29. For each prefix, we evaluated ten candidate additions, including the greedy choice. Each resulting partial set was randomly completed to 30 blocks with the same suffixes, and each completion was evaluated using 400 samples under the same sampling settings. We report the within-prefix Spearman correlation between prediction means:

ρ=1−6∑i=110(r^i−ri)210(102−1).\rho=1-\frac{6\sum_{i=1}^{10}(\hat r_i-r_i)^2}{10(10^2-1)}.

Higher values indicate better agreement between the two rankings.

Prefix Size5102029
ρ\rho0.580.680.770.81

The positive correlations across all evaluated prefix sizes indicate that higher predicted scores tend to correspond to higher completion rewards, supporting their usefulness for guiding greedy construction. Ranking agreement increases from 0.58 at prefix size 5 to 0.81 at size 29, suggesting more accurate rankings as the partial sets approach the target size.

Please note that all revisions only improve the clarity of the presentation and do not change the proposed method.

AuthorsOpenReview ↗

Kind Follow-Up on Further Feedback

Dear Reviewer bjyS,

Thank you again for your careful review and constructive suggestions. We would appreciate your feedback on the responses and revisions we posted on November 12.

To briefly highlight the changes most relevant to your concerns, we have:

  • Reorganized Section 4 and expanded Algorithm 1 to make the sampling procedures, scorer training, and greedy subset construction explicit.
  • Added empirical evaluations of both complete-set rankings and candidate-addition rankings during partial-set construction.
  • Added an enumerable HyperGrid experiment to directly evaluate distribution matching, distinguishing the ungated learned policy from guided sampling.
  • Clarified that hard gating is the default in the main experiments and made the limitations of the theoretical analysis more explicit.

With the Paper Discussion deadline (November 18) approaching, when you have a chance, could you let us know whether these revisions address your main concerns, or whether any specific issues remain unresolved? We would be happy to provide further clarification to support your assessment.

Thank you for your time and consideration.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Concern about Reviewer bjyS’s unexplained post-rebuttal recommendation

Dear Area Chair,

We are writing to flag a concern about Reviewer bjyS’s response to our rebuttal. The reviewer stated that they had read our responses and would maintain their recommendation to reject. Their follow-up did not identify which concerns remained unresolved, explain why our answers were insufficient, or raise any additional concern.

Our objection is not that the reviewer declined to change their recommendation. It is that their comment states an outcome without explaining it. Neither we nor the AC can tell from that comment which issue the reviewer still considers grounds for rejection.

Could you please request a specific clarification from the reviewer: Which concern remains decisive, and why does our response not resolve it? If the reviewer provides no clarification, could you please assess the original criticisms against our rebuttal directly, rather than treat the unchanged recommendation as evidence that the rebuttal was inadequate.

Best regards, The Authors of Paper

Dear Area Chair,

One additional point about the timeline: we posted our rebuttal on November 12, but Reviewer bjyS replied only about 20 minutes before the discussion closed on November 19. That left us little practical opportunity to ask a follow-up question and receive an answer before the deadline. And their entire reply was a single sentence stating that they had reviewed our responses and would maintain their recommendation to reject. They did not identify a single concern that remained unresolved or explain where our rebuttal fell short. With only 20 minutes left, we had little opportunity to seek clarification. We hope that you can this timing into account.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.