acceptodds
Under review as a conference paper at ICLR 2027

Decoupling Exploration from Optimization in RLVR

Abstract

Modern language models undergo reinforcement learning with verifiable rewards (RLVR) on top of already-trained checkpoints. A key promise of RLVR is the discovery of new reasoning strategies. In principle, a model can discover new ideas absent from its prior training data. In practice, however, augmenting RLVR with strong novelty incentives has seen limited success and can degrade model quality. Because verifiable rewards supervise only a narrow slice of the model's knowledge and behavior, degradations are difficult to recover from. We introduce Exploration-Distillation (ExpDis), a framework that decouples exploration from optimization. ExpDis trains an policy with a novelty bonus in the reward, filters its trajectories for correctness and quality, and distills them into a separate policy. The student policy is then trained without a novelty bonus. We repeat the above procedure for several rounds, alternating between exploration and optimization. Further, parallel explorers can evolve independently to produce diverse distillation targets for the student. Across seven mathematical reasoning benchmarks and two model families, ExpDis outperforms DAPO at the same wall-clock budget. Moreover, we observe improved pass@ scaling, indicating that ExpDis produces models that generate more diverse correct solutions.

Primary area:
foundation or frontier models, including LLMs
Keywords:
reinforcement learning with verifiable rewards, exploration, novelty bonus, distillation, LLM reasoning, diversity

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_jmNqOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Exploration-Distillation (ExpDis), a framework for decoupling exploration from optimization in RLVR. ExpDis trains a separate explorer policy with a novelty-augmented reward, filters its trajectories for correctness and quality, distills the retained trajectories into a student, and finally trains the student with standard correctness-only RLVR.

ExpDis consistently outperforms DAPO across three models in both pass@1 and pass@64, with a single-round version already outperforming DAPO trained for substantially more RL steps. Further gains are obtained by distributing a fixed RL budget across multiple parallel explorers and repeated exploration-distillation rounds.

I find the core method elegant, particularly because it separates aggressive exploration from the deployed optimization policy while avoiding direct off-policy RL. However, the paper does not directly establish that the trajectories retained from the explorer are genuinely unlikely under the student before distillation, and some interpretations of the MMLU-Pro degradation could be better supported. Overall, I find the method simple, well motivated, and empirically strong. My overall assessment is weak accept.

Critical strengths

The main strength of the paper is the elegance of the method. I particularly like the way ExpDis separates exploration from optimization: the explorer can be pushed aggressively toward novelty, even at the cost of degrading its own policy, while only filtered successful trajectories are transferred to the student through SFT. The student then returns to its own distribution and consolidates these discoveries with standard on-policy RL. I find this a particularly clever design, as it gets the benefit of aggressive exploration without either corrupting the deployed policy or introducing a direct off-policy RL objective.

The empirical results are also strong and consistent across three models, with improvements in both pass@1 and pass@64. The paper is clearly written and unusually thorough in its ablations, covering model choice, training data, design components, novelty mechanisms, and hyperparameters.

Critical weaknesses

1. The analysis does not directly establish that the retained explorer trajectories expand the student’s support.

The paper evaluates several notions of diversity, including lexical diversity, answer diversity, entropy, and an embedding-based semantic strategy diversity metric. However, these measurements are primarily properties of the resulting generations and do not directly show that the filtered correct trajectories transferred from the explorer were unlikely under the student before distillation.

This distinction seems important because the reported diversity metrics are not uniformly aligned. For example, ExpDis improves the semantic diversity metric in Table 6, while some lexical and answer-diversity measures are comparable to or even lower than DAPO depending on the configuration. This makes it difficult to infer from aggregate diversity statistics alone that the explorer is consistently discovering reasoning trajectories outside the student's existing support.

A more direct analysis would be to evaluate the retained explorer trajectories under the pre-SFT student policy πt−1\pi_{t-1}, for example through sequence likelihood, token-level NLL, or perplexity, and compare them against trajectories generated directly by the student. This would provide substantially stronger evidence for the paper's support-expansion interpretation.

I would also be interested in a simpler exploration baseline based on increased stochasticity. Table 3 indicates that training rollouts already use T=1.0T=1.0, so a useful comparison would be increasing the temperature to T>1T>1 or an entropy-regularized RL baseline. Figure 9 also appears to include a λ=0\lambda=0 setting; if this corresponds to running the same exploration-distillation pipeline without the novelty reward, it would be helpful to clarify this explicitly and discuss what it reveals about the contribution of novelty itself.

Also, the compute comparisons count RL updates but exclude the SFT stage. Since SFT plays a crucial role in ExpDis, reporting end-to-end training cost including SFT would make the efficiency comparison more complete.

2. The interpretation of the MMLU-Pro degradation could be better supported.

The paper interprets the reduction in MMLU-Pro performance as evidence of knowledge loss under aggressive novelty optimization. I think this interpretation is plausible but stronger than the current evidence supports. The observed degradation is relatively small, and without multiple seeds or confidence intervals it is difficult to distinguish systematic forgetting from optimization variance or broader behavioral changes induced by the novelty objective.

A category-level breakdown would also be informative, particularly given the observation that mathematical categories may improve while aggregate MMLU-Pro performance decreases. More generally, since the paper speculates that degradation may accumulate with stronger or longer novelty optimization, reporting MMLU-Pro performance across different values of λ\lambda or different exploration durations would provide direct evidence for this hypothesis.

Questions for authors
  • Section 3 states that applying the novelty bonus to all rollouts rather than only verifier-correct ones leads to training instability. Could the authors characterize this failure mode more concretely? For example, does the explorer collapse toward high-novelty but incorrect behaviors, causing too few trajectories to survive rejection sampling, or does the instability arise through a different mechanism?
  • The main empirical evaluation is limited to mathematical reasoning. How well do the authors expect ExpDis to generalize to domains such as code generation, tool use, or long-horizon agentic tasks?
2 comments
AuthorsOpenReview ↗

Response to Reviewer jmNq (Part 1/2)

(Part 1/2)

We thank the reviewer for the detailed and constructive feedback. We address your points below, including additional experiments and paper edits found in our revised draft. Figure, table, and appendix numbers below refer to the revised PDF.

W1a. The analysis does not directly establish that the retained explorer trajectories expand the student's support.

Thanks for the suggestion. We ran the analysis you propose (see Figure 18). We score rollouts under the pre-SFT student, which is the base model in round 1, and compare the NLL of the traces the filter retains with the base model's own samples. The student's own samples average 0.242 nats per token, with a 95th percentile of 0.342.

Traces scored under the pre-SFT studentMean NLL per tokenShare beyond the student's 95th percentile
Student's own samples0.2425.0%
Explorer, λ=0, all rollouts0.28518.3%
Explorer, λ=0, filtered0.27813.0%
Explorer, λ=0.5, all rollouts0.31834.6%
Explorer, λ=0.5, filtered0.30125.0%

The last row covers the explorer rollouts that pass the filter (correct, complete and within the length window), which are the traces the student is ultimately trained on. One in four of them is less likely than 95% of the student's own samples, five times the reference rate of 5%. In other words, we observe that our explorers indeed generate unlikely traces under the base model. We have added this experiment to our updated draft.

W1b. Simpler exploration baselines: higher temperature, entropy regularization, and what λ=0 means.

The λ=0 point in Figure 11 is the full ExpDis pipeline (explorer, filter, SFT, student RL) with the novelty term removed. We added the above baselines as well (Figure 17):

MethodAIME24AIME25AIME26MATH500MinervaMean
DAPO (paper)50.0536.9337.6674.2429.2745.63
DAPO, T=1.249.5936.5837.3173.5629.0045.21
DAPO + entropy bonus49.9336.8337.5774.0629.2045.52
ExpDis, λ=050.6137.3839.4374.5629.5646.31
ExpDis, λ=0, T=1.250.3937.2038.7374.4329.4546.04
ExpDis, entropy bonus50.6037.3739.4274.5529.5546.30
ExpDis, λ=0.5 (paper)52.7639.1746.2575.8030.6948.93

Raising the rollout temperature to 1.2 or adding an entropy bonus leaves DAPO at or below its original accuracy. With λ=0, ExpDis performs close to DAPO, so the novelty bonus accounts for most of its gain. Replacing it in the explorer with a higher temperature or an entropy bonus does no better than λ=0, indicating that high temperature and the entropy bonus do not confer exploration benefits on accuracy.

W1c. Compute comparisons exclude SFT.

Thanks for pointing out that our budget matching was confusing in the submitted draft. We want to clarify that runs are wall-clock matched, including the SFT cost and not only the RL cost. Because our filtering scheme selects shorter generations, our models tend to generate marginally shorter traces, which makes up for the small added SFT cost, meaning ExpDis can perform the same number of RL iterations and also the SFT stages in the same time that DAPO performs just the RL iterations. Thanks for this feedback, we have updated our draft to better explain the full training cost and how to compare the runs.

W2. The interpretation of the MMLU-Pro degradation could be better supported (seeds, categories, λ, duration).

Seeds. Change in MMLU-Pro accuracy relative to Base, averaged over three seeds (we will report ten seeds in the camera-ready version and are running them now):

MethodQwen3-1.7B changeQwen3-4B changeMinistral-3-3B change
DAPO+0.11+0.10+0.19
DAPO (4× steps)+0.49+0.40+0.33
DAPO (+novelty)−1.91−1.18−0.79
ExpDis explorer−1.52−0.75−0.54
ExpDis (single-round)+0.81+0.67+0.72
ExpDis+1.59+1.18+1.24
AuthorsOpenReview ↗

Response to Reviewer jmNq (Part 2/2)

(Part 2/2)

Categories (Figures 19 and 20). DAPO (+novelty) is below Base in 13 of 14 MMLU-Pro categories, most in philosophy (−4.4) and history (−3.7), and also in the math category (−1.3). ExpDis is above Base in 12 of 14. MMLU-Redux, GPQA-Diamond and IFEval show the same pattern on all three models. On ZebraLogic, DAPO (+novelty) is equal with Base and ExpDis gains the most:

ModelMethodMMLU-ProMMLU-ReduxGPQA-DiamondIFEvalZebraLogic
Qwen3-1.7BBase (accuracy)56.5073.0834.3969.7162.45
DAPO+0.11+0.06+0.33−0.23+0.81
DAPO (4× steps)+0.49+0.27+0.76−0.40+1.50
DAPO (+novelty)−1.91−0.89−0.95−1.77+0.01
ExpDis+1.59+0.88+1.96+0.32+3.41
Qwen3-4BBase (accuracy)70.4083.4455.0882.0082.08
DAPO+0.10+0.05+0.25−0.24+0.57
DAPO (4× steps)+0.40+0.22+0.55−0.43+1.02
DAPO (+novelty)−1.18−0.40−0.57−0.74+0.05
ExpDis+1.18+0.65+1.36+0.23+2.22
Ministral-3-3BBase (accuracy)25.5055.3839.0248.7612.13
DAPO+0.19+0.10+0.31−0.21+0.64
DAPO (4× steps)+0.33+0.18+0.55−0.43+1.13
DAPO (+novelty)−0.79−0.44−0.33−0.87+0.20
ExpDis+1.24+0.68+1.56+0.25+2.75

Novelty weight (Figures 21 to 23). Change relative to Base when the model is trained with novelty weight λ:

λMMLU-ProMMLU-ReduxGPQA-DiamondIFEvalZebraLogic
0+0.11+0.06+0.33−0.23+0.81
0.5−1.91−0.89−0.95−1.77+0.01
4−6.42−5.55−1.86−8.53−6.05
8−9.54−8.78−2.49−13.21−10.24
16−13.81−13.19−3.35−19.61−15.97
32−19.63−19.21−4.53−28.33−23.79

Math accuracy for the single model and for the ExpDis student:

MethodAIME24AIME25AIME26MATH500MinervaMean
Single model, λ=446.6934.4435.1369.2627.3142.57
Single model, λ=844.0132.4633.1165.2825.7440.12
Single model, λ=1639.3129.0029.5858.3122.9935.84
Single model, λ=3232.2623.8024.2747.8518.8729.41
ExpDis student, λ=451.0937.7940.9774.8429.8246.90
ExpDis student, λ=850.5337.3239.1774.5129.5246.21
ExpDis student, λ=1650.1437.0037.9574.2929.3245.74
ExpDis student, λ=3249.7436.6937.4373.7829.0945.35

At λ=32, the single model is 16.22 points below DAPO on math and 19.63 below Base on MMLU-Pro, while the ExpDis student stays within 0.28 points of DAPO and above Base.

Duration (Figure 24). At λ=0.5 the single model's MMLU-Pro deficit goes from 1.91 points at 300 updates to 3.24 at 600 and 5.05 at 1,200. The ExpDis student stays above Base throughout.

Q1. What is the failure mode when the bonus is applied to all rollouts?

When wrong answers also earn the novelty reward, the explorer is rewarded for being unusual, whether or not it is right (Figure 26). Its responses hit the 32,768-token budget within about 50 updates, so 99% are cut off before giving an answer and correctness falls from 0.73 to 0.01. The filter then finds 259 traces instead of the full 500, almost all from early in training. The explorer drifts toward novel but incorrect behavior, and too few trajectories survive rejection sampling. We have updated our draft to discuss this.

Q2. How well should ExpDis generalize to code, tool use or long-horizon agentic tasks?

In principle, we believe that these domains are just as compatible with ExpDis as math, but we chose math due to its relatively affordable computational costs. We agree that extending our approach to other domains is a valuable objective for future work. Prompted by your feedback, we have updated the discussion section of our draft accordingly.

The revision also extends pass@k to 1,024 samples, where ExpDis remains above the base model at every k on all three models (Figure 3). We also compare ExpDis with R²PO and ICE, two recent exploration methods for RLVR. A single round of ExpDis outperforms both on all three models (Table 1).

Thanks again for your feedback. Motivated by your points, we have run additional experiments and made numerous paper edits, which we feel have improved the quality of our draft. We would greatly appreciate it if you would consider raising your score accordingly. Do you have any other questions we can address?

Reviewer_u73qOpenReview ↗
Rating
2(+1)
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Exploration-Distillation (ExpDis), a framework for decoupling exploration from optimization in reinforcement learning with verifiable rewards (RLVR). It trains a separate explorer with a novelty-augmented reward, filters its successful trajectories, and distills them into a student that is subsequently optimized with standard correctness-only RLVR. ExpDis improves both accuracy and pass@k over standard RLVR baselines.

Critical strengths
  • The writing is fine, with illustrations aiding understanding.
Critical weaknesses
  • The paper is obviously under-prepared for a conference submission. The current content is not enough. More experiments should be done and placed in Section 5 instead of leaving half of page 9 blank.

  • Novelty of the proposed method is limited. The main idea of exploration-deployment decoupling and the key methodological components are largely proposed in prior works, e.g. in R$^2$PO and ICE.

  • The most closely related methods discussed in the paper, including R$^2$PO and ICE, are not compared experimentally.

  • The sensitivity figure of novelty weight λ\lambda does not even have a caption.

Questions for authors
  • How does ExpDis compare with R$^2$PO and ICE empirically?

  • The rejection-sampling stage discards all incorrect explorer trajectories to prevent the student from imitating degraded behavior. Have the authors tried using them as negative samples instead?

4 comments
AuthorsOpenReview ↗

Response to Reviewer u73q

We thank the reviewer for the detailed and constructive feedback. We address your points below, including additional experiments and paper edits found in our revised draft. Figure, table, and appendix numbers below refer to the revised PDF.

W1. The paper is under-prepared; more experiments should be in Section 5.

Motivated by your feedback, our revised version adds the following to Section 5 and Appendix G: pass@k up to 1,024 samples, wall-clock matching including SFT, temperature and entropy baselines, the likelihood of distilled traces under the pre-SFT student, prior-capability (non-math) results on five benchmarks (MMLU-Pro, MMLU-Redux, GPQA-Diamond, ZebraLogic, IFEval) by category, across seeds, across novelty weights and across training length, and an analysis of how the explorer and student policies evolve across rounds (entropy, semantic diversity and reasoning faithfulness).

W2, W3, and Q1. R²PO and ICE.

Exploration is a hard fundamental problem with space for numerous approaches attempting to solve it. Thanks for suggesting a comparison to R²PO and ICE. We have added discussion of these methods to our draft, and we now compare the performance of these approaches directly to our own. For R²PO, we follow the authors' released code. ICE has no public code, so we follow the setup in the paper.

Mean pass@1 over the five math benchmarks in the paper's setting (Table 1):

MethodQwen3-1.7BQwen3-4BMinistral-3-3B
DAPO45.6362.5829.63
R²PO46.7964.1730.38
ICE46.4161.6129.63
ExpDis (single-round)48.9364.6932.16
ExpDis51.8166.5034.76

A single round of ExpDis outperforms both methods on every model. R²PO improves on DAPO on all three models, whereas ICE improves on DAPO only on Qwen3-1.7B.

W4. The sensitivity figure for the novelty weight has no caption. Thanks for catching this. We have added a caption to the novelty-weight figure (Figure 11).

Q2. Have the authors tried using incorrect explorer trajectories as negative samples?

Yes. We trained the student on correct traces with reward +1 and incorrect traces with −1, with an importance weight against the student:

MethodAIME24AIME25AIME26MATH500MinervaMean
ExpDis, λ=0.5 (paper)52.7639.1746.2575.8030.6948.93
ExpDis + negatives (±1)51.8438.4143.3375.2730.2147.81

Accuracy falls by 1.12 points relative to distilling correct traces only. The ablation is now in Appendix F (Table 9).

The revision also extends pass@k to 1,024 samples, where ExpDis remains above the base model at every k on all three models (Figure 3), and reports five prior-capability benchmarks, on which ExpDis improves while DAPO with a novelty bonus degrades (Figure 12 and Appendix G.4).

Thanks again for your feedback. Motivated by your points, we have run additional experiments and made numerous paper edits, which we feel have improved the quality of our draft. We would greatly appreciate it if you would consider raising your score accordingly. Do you have any other questions we can address?

AuthorsOpenReview ↗

Follow-up on our response to Reviewer u73q

Thanks again for your feedback. We made a strong effort to address all your points, including substantial edits to our manuscript, and we would greatly appreciate it if you would consider increasing your score accordingly. Do you have any other questions?

AuthorsOpenReview ↗

Final follow-up before the discussion deadline

As the discussion period closes soon, we wanted to briefly follow up. Prompted by your review, we added direct comparisons to R²PO and ICE, the negative-samples ablation, several new experiments in Section 5 and Appendix G, and the missing caption. We think your feedback and these edits improved our draft. Please let us know if you have any remaining questions before the deadline. If our updates addressed your feedback, we would appreciate it if you would consider increasing your score.

Thank you for the detailed rebuttal and substantial additional experiments. The new experiments have significantly strengthen the paper.

However, my main concern regarding methodological novelty remains. The proposed framework still largely combines existing exploration, filtering, and distillation techniques, and its incremental contribution over prior work is not fully convincing.

Given the above reasons, I have raised my rating from 1 to 2.

Reviewer_p4ngOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ExpDis, which trains an explorer with a novelty reward and uses its filtered solutions to fine-tune a separate student. The student then undergoes standard RL without the novelty reward.

Critical strengths

The paper studies a useful way to encourage exploration without hurting the final model, and reports improvements across two model families.

Critical weaknesses
  1. The method requires explorer RL, trajectory filtering, student SFT, and student RL. This makes the training pipeline fairly complicated, especially when using multiple explorers and rounds.

  2. Evaluation stops at pass@64. Prior work [1] shows that base models can overtake RL-trained models at larger sampling budgets. Testing beyond this crossover would help establish whether ExpDis improves reasoning coverage rather than mainly sampling efficiency.

  3. The budget is matched by RL steps, but ExpDis also includes student SFT. Since this extra cost is not matched in the baselines, the gains may partly come from additional training compute.

    [1] Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?

Questions for authors
  1. Could you provide a comparison that matches the full training cost, including student SFT?
  2. At what sampling budget does the base model overtake GRPO/DAPO, and how does ExpDis perform in that range?

Minor Suggestion: There is still some unused space in the main paper. Consider adjusting the layout and moving some of the more important results or details from the appendix into the main text. This is only a presentation suggestion and does not affect my score.

2 comments
AuthorsOpenReview ↗

Response to Reviewer p4ng

We thank the reviewer for the detailed and constructive feedback. We address your points below, including additional experiments and paper edits found in our revised draft. Figure, table, and appendix numbers below refer to the revised PDF.

W1. The pipeline is complicated, especially with multiple explorers and rounds.

The simplified variant with a single round containing one explorer already outperforms DAPO trained for four times as many steps (Figure 3). In practice, only the novelty bonus weight λ requires tuning, since the filter can rely on correctness alone, and ExpDis outperforms DAPO for every number of explorers and rounds we evaluate (Figure 4). Broadly, ExpDis is robust to hyperparameters (λ, number of rounds, number of explorers), and we feel that the large performance benefits will be decisively worth the complexity in many settings.

W2 and Q2. Evaluation stops at pass@64. At what sampling budget does the base model overtake GRPO/DAPO, and how does ExpDis perform in that range?

Motivated by your feedback, we extended evaluation to 1,024 samples per problem on all three models (Figure 3 and Table 5). Mean pass@k over the five math benchmarks:

ModelMethod@1@64@128@256@512@1024
Qwen3-1.7BBase43.4273.8276.5078.7380.6182.21
GRPO43.9473.6376.1878.3180.1181.65
Dr. GRPO44.5573.9976.4178.4580.1881.66
DAPO45.6374.6476.8478.7080.3081.67
DAPO (4× steps)47.2876.2478.4080.2381.7983.13
ExpDis (single-round)48.9377.8379.9681.7583.2684.56
ExpDis51.8183.0485.2887.0888.5689.77
Qwen3-4BBase61.0679.8680.9081.8382.6483.37
GRPO61.4279.8780.9081.8182.6283.33
Dr. GRPO61.8480.3381.4082.3583.1883.91
DAPO62.5881.1482.2983.2984.1784.94
DAPO (4× steps)63.6482.8284.1385.2586.2287.07
ExpDis (single-round)64.6984.5085.9687.1888.2389.12
ExpDis66.5087.6389.0890.2691.2592.07
Ministral-3-3BBase27.9556.5059.4662.0364.2666.22
GRPO28.3556.7159.5762.0564.2066.08
Dr. GRPO28.8157.9460.7663.2165.3467.20
DAPO29.6360.1262.8865.2767.3569.17
DAPO (4× steps)30.9062.3665.0867.4369.4771.25
ExpDis (single-round)32.1664.5967.2869.5971.5973.33
ExpDis34.7672.5675.6878.2680.4082.19

On Qwen3-1.7B, the base model overtakes GRPO at k=64, Dr. GRPO at k=128 and DAPO at k=256. ExpDis is not overtaken at any point up to 1,024 on any model, and its margin over Base at k=1024 is 7.56 points on Qwen3-1.7B, 8.70 on Qwen3-4B and 15.97 on Ministral-3-3B, continuing the same trend we observed before at smaller values of k.

W3 and Q1. The budget is matched by RL steps but not by the student SFT. Could you match the full training cost?

Thanks for pointing out that our budget matching was confusing in the submitted draft. We want to clarify that runs are wall-clock matched, including the SFT cost and not only the RL cost. Because our filtering scheme selects shorter generations, our models tend to generate marginally shorter traces, which makes up for the small added SFT cost, meaning ExpDis can perform the same number of RL iterations and also the SFT stages in the same time that DAPO performs just the RL iterations. Thanks for this feedback, we have updated our draft to better explain the full training cost and how to compare the runs.

We also report five prior-capability benchmarks (MMLU-Pro, MMLU-Redux, GPQA-Diamond, IFEval and ZebraLogic). Adding a novelty bonus directly to DAPO lowers accuracy on four of them, whereas ExpDis improves on all five (Figure 12 and Appendix G.4). We also compare ExpDis with R²PO and ICE, two recent exploration methods for RLVR. A single round of ExpDis outperforms both on all three models (Table 1).

Thanks again for your feedback. Motivated by your points, we have run additional experiments, made numerous paper edits, and used these to fill the remaining space, which we feel have improved the quality of our draft. We would greatly appreciate it if you would consider raising your score accordingly. Do you have any other questions we can address?

Thanks for the follow-up. Most of the points I raised in my review have been addressed, so I will update my score to 3.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.