acceptodds
Under review as a conference paper at ICLR 2027

Abstract-State Grouping Policy Optimization for CLI Agents

Abstract

Recent methods for step-level credit assignment in agentic reinforcement learning, such as Group-in-Group Policy Optimization (GiGPO), compare actions taken from common states across trajectories. However, states of command-line interface (CLI) agents are represented by terminal transcripts that vary across runs. Run-specific paths, hashes, and logs make these transcripts differ even when they describe similar execution situations. Because common-state comparisons require identical transcripts, these differences make common states rare across trajectories. In this paper, we observe that up to 98% of the groups formed by common states contain only one sampled step, and such groups yield zero step-level advantage. To address this problem, we propose Abstract-State Grouping Policy Optimization (ASGPO), a critic-free algorithm that estimates step-level advantages by grouping steps according to abstract states, characterized by recent execution feedback and interaction progress. Specifically, ASGPO maps each preaction transcript to an abstract state, and then groups steps sharing this abstract state and computes their relative advantages using bounded progress-aware returns. These returns combine the final task reward with execution-feedback scores accumulated from each step onward. Positive and negative progress contributions are capped separately to control their scale. Comprehensive experimental results demonstrate the effectiveness of our proposed ASGPO.

Primary area:
reinforcement learning
Keywords:
Reinforcement Learning, Large Language Model Agents, Step-Level Credit Assignment, State Abstraction, Group Relative Policy Optimization

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_a1EWOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses state-matching for GRPO training of CLI agents, where varying terminal transcripts make exact-state matches rare. It groups steps by pre-action feedback type and interaction depth, then compares returns combining final task rewards with capped scores accumulated from each step onward. A theorem is proposed to quantify the trade-off between larger comparison groups and greater within-group value heterogeneity. Experiments across 4B/7B/8B on OpenThoughts Terminal-Bench-format tasks and SETA shows improved performance compared to exact-state step-grouping baseline and greater zero-shot transfer to Terminal-Bench 2.0.

Critical strengths

The abstract comparison trade-off between comparison-group size, within-group value differences, and baseline estimation error provides insights on the state-grouping mechanism for RL training. It also explains well why abstract-state grouping can provide more learning signals for step-by-step progression, and the extent to which such grouping is appropriate.

Critical weaknesses
  1. The feedback mode and execution-feedback scores (process rewards) are defined heuristically, particularly for CLI agents. Terminal-Bench transfer is useful evidence within CLI domain, but broader applicability to web, GUI, or other tool environments remains unknown. So I have concerns about the generality of the proposed method.
  2. Insufficient comparison with the closest prior work.
  • No comparison with full GiGPO or its similarity-based grouping variant.
  • HGPO[1], ProxMO[2], and RTMC[3] are discussed as alternatives that relax raw-state identity but receive no quantitative comparison. BiPACE is evaluated only at 4B using the authors’ reimplementation, with different training/evaluation settings and checkpoint coverage. The externally reported PPO result also uses a different setup. These results provide limited evidence that ASGPO improves over competitive existing credit-assignment methods.

Suggestions for the presentation: Better to add benchmark and baseline details in Section 6.1. Also, Figure 3 is overlapping with table 1-4 and Figure 4 is overlapping with table 3. They can be deleted or removed to the appendix.

References:

  1. He, Shuo, et al. "Hierarchy-of-groups policy optimization for long-horizon agentic tasks." International Conference on Learning Representations. 2026.
  2. Fang, Yangyi, et al. "Proximity-based multi-turn optimization: Practical credit assignment for llm agent training." ACL 2026.
  3. Wang, Tao, Suhang Zheng, and Xiaoxiao Xu. "RTMC: Step-Level Credit Assignment via Rollout Trees." arXiv preprint arXiv:2604.11037 (2026).
Questions for authors
  1. What is the value of Γt\Gamma_t in Eq.6? Is it assigned to each step in a short failed episode or a failed episode whose prompts exceed the context budget?
  2. Since the premise of Theorem 1 is leave-in centering, does it still apply to the leave-one out setting, such as the RLOO algorithm?
  3. Are the comparable samples within a group drawn from the entire training batch, or only from trajectories that originated from the same initial state?
  4. What do the cells in Table 1 with no values signify, such as 7B-ASGPO sing., and BiPACE using 7B/8B agent?
5 comments

Thank you for identifying the questions about the comparison protocol and the implementation. We address them below. The revised appendix highlights the corresponding clarifications.

W1. Scope of the estimator. ASGPO centers progress-aware step returns in groups defined by pre-action feedback and interaction depth. The same estimator applies when an environment exposes feedback for selecting peers and observable events for constructing progress-aware returns. In a web or GUI environment, a form validation error could define a feedback label, and a verified application-state change could contribute to the step return. Our transfer evidence covers CLI tasks. ASGPO reaches 12.7% on 79 scored Terminal-Bench 2.0 tasks (Appendix B.6).

W2. Full GiGPO and other baselines. The original Exact baseline tests how concrete-state matching supplies peers for step credit. Given the limited rebuttal window, we prioritized completing the full GiGPO comparison at 7B, the middle scale in Table 1. We used default exact matching and separate episode and step advantages. Both runs use the same SFT initialization, OpenThoughts RL split, and evaluation tasks, with four rollouts per task. The original 7B ASGPO leads standard GiGPO in both shared-checkpoint summaries. The original Table 1 entry is 34.59%, and standard GiGPO's checkpoint mean is 28.15% (Appendix B.2).

When all sampled trajectories for a hard task receive zero verifier reward, episode centering gives zero throughout the group. The episode term also leaves exact step groups unchanged. Abstract grouping supplies peers across transcripts, and bounded progress in the step return distinguishes execution outcomes within those groups. This mechanism applies at any model size under these reward and grouping conditions. Table 1 shows gains over Exact at 4B, 7B, and 8B across two CLI task suites.

Appendix B.4 compares the other estimators. GiGPO's similarity variant uses a text threshold. HGPO's nested histories refine current-state groups and therefore cannot add peers to a singleton at the coarsest level. ProxMO's soft TF–IDF baseline can be dominated by the target's own return. RTMC estimates values for state–action signatures, and we could not verify a complete public release of its original training implementation. These comparisons concern the estimators. Table 1 reports our paper-based 4B BiPACE implementation with thinking on during training and off during evaluation. Its checkpoint coverage differs from the 4B ASGPO run. Table 2 reports the published PPO score under the Endless Terminals setup (Appendices A and B.7).

Q1. Return adjustment in Eq. (6). In the verified original 4B configuration, each retained step receives the same Γt\Gamma_t. A failed trajectory incurs −0.01-0.01 per recorded action short of three. The adjustment is −0.02-0.02 for a one-action failure and −0.01-0.01 for a two-action failure. A successful short rollout has no short-episode penalty. A recorded PROMPT_TRUNCATION termination adds −0.10-0.10. The terms add when both conditions hold and the guard is zero when neither applies (Appendix B.7).

Q2. Leave-one-out. For n>1n > 1, let Rˉ\bar R be the group mean and Rˉ−i\bar R_{-i} the mean of the other n−1n-1 returns. Then

Ri−Rˉ=(1−1/n)(Ri−Rˉ−i).R_i-\bar R=(1-1/n)(R_i-\bar R_{-i}).

Leave-one-out removes the self-inclusion factor. Under the conditional action-law and independent-peer assumptions, the peer baseline has zero expected product with the target score. The peer-count and value-aliasing trade-off still applies with n−1n-1 independent peers. A singleton has no leave-one-out peer. Appendix B.5 distinguishes strict centering from the archived implementation's singleton fallback.

Q3. Comparison pool. The original ASGPO pools steps sharing a label across the current update batch, including across tasks. Both branches of the same-size shuffle in Appendix B.3 group within each task. Task conditioning removes the between-task component of value heterogeneity. Batch pooling can supply additional peers (Appendix B.6).

Q4. Missing entries. The 7B ASGPO dash denotes an unreported singleton diagnostic. The 7B and 8B BiPACE dashes denote unreported scores (Appendix B.7).

Presentation. Section 6.1 now consolidates the benchmark and baseline setup. We removed the original Figure 3 and moved the original Figure 4 to Appendix D, where it appears as Figure 3. Appendix B.7 records these changes.

Thanks for the authors' responses. However, without experiment results on environments besides CLI domains and mentioned baselines (e.g., similarity-based GiGPO & HGPO, etc.), my concerns still remained. So I maintain the original score.

Dear Reviewer a1EW,

Thank you again for your careful review and for following up on transfer beyond CLI and the additional baselines.

Our earlier response reported the completed 7B CLI comparison against full GiGPO. It used GiGPO’s default exact matching and both episode-level and step-level advantages.

We also ran an initial zero-shot test on 100 synthetic WebShop tasks using the CLI-trained 7B policies. Two-level ASGPO succeeded on 23 tasks, and standard GiGPO succeeded on 22. Training ASGPO on WebShop requires defining comparison groups and progress-aware returns from shopping feedback.

The 7B HGPO runs are still training, and the CLI run may take longer than the remaining discussion period. If the held-out evaluation finishes in time, we will include the HGPO results in a final update.

As the discussion period draws to a close, we would welcome any further questions about the transfer setup or the baseline comparisons. Thank you again for your time and consideration.

Best regards, The Authors

Reviewer_hup9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper pays attention to the problem of ineffective step-level credit assignment in reinforcement learning for command-line interface (CLI) agents. Existing methods, such as GiGPO, rely on matching identical states across trajectories to estimate step-level advantages. However, CLI states often contain run-specific information, making exact state matches rare and resulting in many groups with zero step-level advantages. To address this issue, the authors propose Abstract-State Grouping Policy Optimization (ASGPO), which groups steps based on execution feedback and interaction progress rather than exact transcript matching. The method also introduces bounded progress-aware returns to distinguish the contributions of individual actions, even when trajectories receive the same final reward. In addition, the authors provide a theoretical analysis of the trade-off between grouping granularity and baseline estimation error. Experiments across different model scales demonstrate that ASGPO improves task success rates over existing baselines, including on Terminal-Bench 2.0.

Critical strengths
  1. Clear motivation and a well-identified problem. The paper identifies a practical limitation of exact-state grouping in CLI environments, where run-specific information makes identical states difficult to match across trajectories. The authors provide empirical evidence showing that a large proportion of groups contain only one sample, resulting in zero step-level advantages. This observation provides a clear motivation for the proposed method.

  2. Consistent experimental improvements. The proposed method achieves improvements over GRPO and exact-state grouping across different model scales (4B, 7B, and 8B). The results on Terminal-Bench 2.0 further demonstrate its potential to generalize to unseen tasks. The ablation studies also provide useful evidence for the contributions of abstract-state grouping and progress-aware returns.

Critical weaknesses
  1. Limited novelty. The proposed method closely follows GiGPO's step-level credit assignment framework with incremental improvements. The main difference is the replacement of exact-state grouping with a heuristic abstraction based on execution feedback and interaction progress. Therefore, it is unclear how much methodological novelty ASGPO offers beyond adapting existing grouping strategies to CLI environments.

  2. Incomplete comparisons with existing methods. (1) The main Exact baseline only uses GiGPO's exact-state grouping mechanism, rather than its complete advantage estimator that combines episode-level and step-level advantages. While this comparison helps isolate the effect of state grouping, it does not establish whether ASGPO outperforms the original GiGPO. (2) There are some other step-level credit assignment methods, including grouping actions from shared states, incorporating interaction histories into state grouping, using state similarity to construct comparison groups, and clustering states based on learned representations. These methods are closely related to ASGPO's core idea of grouping steps based on abstract state information rather than exact transcript matching.

Questions for authors

Please see the weakness.

2 comments

Thank you for recognizing the evidence for fragmentation under exact state matching. We clarify the estimator and add the requested comparison with full GiGPO.

W1. Methodological contribution. ASGPO centers each step's progress-aware return within an execution-defined group. Pre-action feedback and depth let distinct terminal transcripts share a baseline when their execution conditions match. The policy still receives the full transcript. Bounded progress is part of the step return and can distinguish two failed rollouts when one makes measurable execution progress. Equal returns give zero centered credit at any group size, and strict self-centering cancels the return in a singleton. The estimator addresses both conditions for a useful step comparison.

The 4B ablation scores 31.4% with the full estimator, 22.9% with concrete-state grouping and progress-aware returns, and 13.5% with abstract grouping and sparse task returns. In Appendix B.3, shuffling the labels preserves each batch's group sizes, the step-return definition, and the two-level estimator. Execution labels lead by 4.97 percentage points. This tests who is compared at fixed group sizes (Appendix B.4).

W2(a). Exact and full GiGPO. The original Exact row uses concrete-state step groups and sparse task returns to expose fragmentation. Table 3 includes a separate concrete-state control that retains progress-aware returns. Given the limited rebuttal window, we prioritized completing the full GiGPO comparison at 7B, the middle scale in Table 1. The full baseline uses default exact grouping and both episode and step advantages. The 7B runs share the SFT initialization and OpenThoughts RL split. The original ASGPO leads in both shared-checkpoint summaries. The original Table 1 entry is 34.59%, and standard GiGPO's checkpoint mean is 28.15% (Appendix B.2).

When all sampled trajectories for a task fail with zero verifier reward, the centered episode term is zero, and the exact step groups still have the same peers. ASGPO can recover step credit when distinct transcripts supply peers and execution progress differs within a group. These conditions depend on the rollouts and the estimator. They apply across model sizes. Table 1 shows gains over Exact at 4B, 7B, and 8B across two CLI task suites.

W2(b). Related group constructions. HGPO uses shared histories to refine common-state groups. Refinement cannot add peers to a singleton. GiGPO's similarity variant uses a text threshold, and its first-matching-representative rule can depend on input order. ProxMO uses target-specific TF–IDF weights, so a textually isolated sample can nearly subtract its own return. BiPACE clusters actor representations and averages by action key. Its MICo-based guarantee requires a diameter condition beyond the cosine threshold used by the clustering rule. RTMC averages returns within state–action signatures and can use execution outcomes in action keys. ASGPO gives each pre-action label one shared baseline and keeps each sample's own progress-aware return. Appendix B.4 gives the assumptions and concrete examples behind these distinctions.

Dear Reviewer hup9,

Thank you again for your careful review and for raising questions about ASGPO’s contribution and the choice of baselines. We appreciate that this is a busy period and would be grateful if you had a chance to read our response before the discussion closes.

Our response explains how ASGPO uses pre-action feedback and interaction depth to define which steps are compared, then centers each step’s progress-aware return within its group. The label-shuffle control holds group sizes fixed while changing which steps are compared. We also reported a 7B comparison against full GiGPO with both episode-level and step-level advantages.

If you have any remaining questions about the estimator or the comparisons, we would be happy to address them. Thank you again for your time and consideration.

Best regards, The Authors

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

This paper proposes Abstract-State Grouping Policy Optimization (ASGPO), a critic-free reinforcement learning method for step-level credit assignment in command-line interface (CLI) agents. The paper identifies a limitation of exact-state grouping methods, where differences in terminal transcripts lead to a large number of singleton groups and consequently zero step-level advantages. To address this issue, ASGPO constructs abstract comparison states based on execution feedback and interaction progress, allowing actions from different trajectories to share comparison groups. In addition, the method introduces bounded progress-aware returns that combine terminal task rewards with intermediate execution feedback. The resulting group-relative advantages are used for policy optimization. The paper also presents a theoretical analysis of the trade-off between group size and value aliasing. Experiments on several CLI agent benchmarks demonstrate improved task success rates over exact-state grouping and GRPO baselines.

Critical strengths
  1. The paper addresses a practical limitation of step-level credit assignment in CLI agents. The observation that exact-state grouping produces a large proportion of singleton groups is interesting and provides a clear motivation for the proposed method.

  2. ASGPO constructs abstract comparison states using execution feedback and interaction progress, without requiring an additional critic network or learned state encoder. The method can be integrated into existing group-relative policy optimization frameworks with relatively low computational overhead.

  3. The proposed method consistently outperforms the matched exact-state grouping baseline across different model scales. The ablation studies also provide useful evidence regarding the contributions of abstract-state grouping and progress-aware returns.

Critical weaknesses
  1. The main components of ASGPO, including state abstraction, group-relative advantage estimation, and intermediate reward shaping, have been extensively studied in reinforcement learning. The proposed method mainly combines heuristic state grouping with manually designed progress rewards within the existing GRPO framework. Moreover, previous methods, such as GiGPO and ProxMO, have already explored similarity-based grouping or relaxed state matching. The paper does not sufficiently clarify the fundamental methodological differences between ASGPO and these existing approaches. It is therefore unclear whether the proposed method constitutes a substantial algorithmic contribution beyond a task-specific implementation of existing ideas.

  2. Theorem 1 characterizes the trade-off between group size and value aliasing under state abstraction. However, the established monotonicity properties largely follow from standard partition refinement and conditional variance arguments. More importantly, the theorem does not establish that the particular abstraction adopted by ASGPO leads to improved advantage estimation or policy optimization. The analysis also relies on conditional independence assumptions and primarily considers baseline estimation errors, whereas the actual algorithm involves normalized advantages, correlated multi-step trajectories, and clipped policy updates. The connection between the theoretical results and the empirical performance improvements remains insufficiently established.

  3. The abstract states are constructed using manually specified feedback categories and interaction-depth thresholds. However, states with the same feedback type and interaction stage may have substantially different continuation values and require different actions. The paper does not provide sufficient justification for why these two features are adequate for identifying meaningful comparison groups. Similarly, the progress-aware rewards rely on manually selected event weights and clipping thresholds. Such intermediate rewards may not always reflect actual task progress and could introduce undesirable optimization biases. The robustness of the proposed method to these design choices remains unclear.

Questions for authors
  1. What are the essential differences between ASGPO and existing similarity-based grouping methods, particularly the similarity-based variant of GiGPO and ProxMO? Could the authors provide a more detailed comparison and clarify which component represents the primary methodological contribution?

  2. Theorem 1 establishes a general trade-off between grouping granularity and value aliasing. Can the authors provide a stronger theoretical result that directly characterizes the advantage estimation error, gradient variance, or optimization performance of ASGPO? In particular, under what conditions would the proposed abstract-state grouping outperform exact-state grouping?

  3. How sensitive is the performance to the manually defined feedback categories and progress-stage thresholds? Can the authors provide a systematic sensitivity analysis? Moreover, how does the method handle states that share the same abstract label but have substantially different continuation values?

  4. How are the intermediate reward weights and clipping thresholds determined? Are these parameters tuned separately for different environments? Can the authors evaluate the robustness of ASGPO to alternative reward configurations and investigate whether the additional progress rewards introduce optimization bias?

4 comments

Thank you for asking what the estimator changes and when the added comparison peers improve credit assignment. The revised appendix gives the following details.

Q1. Estimator and related work. ASGPO centers each sample's progress-aware return within a group defined by pre-action feedback and depth. These labels can put two early tracebacks in one group despite different paths and exception strings. Bounded execution progress can then distinguish their returns when final verifier outcomes tie. The full transcript remains available to the policy.

Similarity-based GiGPO uses the first representative that passes its text threshold. If A matches B and B matches C, while A and C fall below threshold, starting with B can merge all three and starting with A separates C. Equality of ASGPO labels gives a fixed partition. ProxMO's Eqs. (10)–(11) give the target return coefficient 1−wii1-w_{ii} after soft centering. For eight candidates, temperature 0.1, and zero similarity to the seven peers, this coefficient is about 0.032%. A uniformly weighted ASGPO group has coefficient 1−1/n1-1/n, at least 1/21/2 for n≥2n\geq2, before standardization.

BiPACE's Proposition 1 requires MICo-Lipschitz state values and a bounded MICo cluster diameter. Its actor-space cosine threshold alone does not establish that diameter. RTMC test keys can condition on the observed outcome. A command that passes half the time can give outcome-conditioned values 1 and 0, while its pre-execution value is 0.5. ASGPO retains the sampled step return without averaging by action key. Appendix B.4 gives these examples. In Appendix B.3, execution labels lead shuffled labels at fixed group sizes, 30.48% versus 25.51%.

Q2. Advantage error and assumptions. With the same target return R0R_0, the difference between R0−b^R_0-\hat b and R0−V(S0)R_0-V(S_0) is V(S0)−b^V(S_0)-\hat b. The unscaled advantage MSE therefore equals the baseline MSE. For independent reference sampling, fix the return definition and target-state distribution within cell zz. Let HzH_z be the within-cell continuation-value variance and σˉz2\bar\sigma_z^2 the average within-state return variance. With mAm_A abstract peers and mEm_E exact-state peers, the MSEs are Hz+(Hz+σˉz2)/mAH_z+(H_z+\bar\sigma_z^2)/m_A and σˉz2/mE\bar\sigma_z^2/m_E, respectively. For mA>mE≥1m_A > m_E\geq1, abstraction lowers this error when

Hz<σˉz2(mA−mE)mE(mA+1).H_z < \frac{\bar\sigma_z^2(m_A-m_E)}{m_E(m_A+1)}.

Correlated peers contribute covariance terms to the baseline variance. If their returns depend on the target action, score–peer cross-terms also matter. Empirical normalization and clipping change the update. Appendix B.5 states the result at the level of unscaled advantage error. The 8B singleton rate drops from 98% to 10%, which measures comparison support. The fixed-size shuffle tests the choice of peers at equal group sizes.

Q3. Labels, thresholds, and value aliasing. The feedback categories distinguish observed execution conditions. Depth buckets at 0, 1–2, 3–6, and 7 or more interactions separate stages of the rollout. Two states with one label can have different values. This variation enters HzH_z, and the full transcript still lets the policy select different actions. Removing feedback or depth lowers the 4B score from 31.4% to 24.5% or 21.9%. These ablations and the shuffle support the coordinate and peer choices. Sensitivity to category merges or individual cutoffs remains open in the reported experiments. Appendix B.6 specifies a controlled test from the same checkpoint.

Q4. Progress weights, caps, and optimization bias. The reported OpenThoughts and SETA configurations use the same weights and caps. A successful ordinary command contributes 0.001, and a passing local test contributes 0.060. Errors, repeats, and timeouts incur penalties. The two 0.060 caps bound the progress suffix to 6% of a unit verifier reward. Local checks can pass in a rollout that ultimately fails. Progress feedback can thus change relative credit among tied final outcomes, and the cap bounds the change without guaranteeing policy invariance. The sparse-return ablation scores 13.5% versus 31.4% for the full estimator. Appendix B.6 explains how alternative weights and caps can be tested by measuring verifier pass rates, progress scores, repeat rates, and cap activation.

Dear Reviewer vMQy,

Thank you again for your careful review and for asking when abstract grouping can improve step-level credit assignment. We appreciate that this is a busy period and would be grateful if you had a chance to read our response before the discussion closes.

Our response gives a condition under which adding independent peers reduces the mean squared error of the unscaled advantage estimate when states in a comparison group have different continuation values. We also explain how feedback labels and depth thresholds define comparison groups, and how bounded progress returns distinguish steps whose terminal rewards tie.

If any questions remain about the assumptions in the analysis or the choice of feedback labels and progress weights, we would be glad to discuss them. Thank you again for your time and consideration.

Best regards, The Authors

Thank you for your response. I have carefully reviewed the additional information provided regarding the theoretical and experimental aspects of the paper.

However, given the current state of the theory and the experimental results, I remain concerned about the theoretical and experimental improvements reported in this study; therefore, I have decided to maintain my initial rating (2).

Dear Reviewer vMQy,

Thank you for taking the time to review our additional analysis. We have since completed the 7B HGPO comparison. At checkpoint 50, the original ASGPO achieved 34.59% success versus 26.71% for HGPO on the same 73 held-out CLI tasks, with four rollouts per task.

The fixed-group-size control addresses whether the choice of comparison peers matters. With group sizes, step returns, and the two-level estimator held fixed, execution-based labels achieved 30.48% success versus 25.51% for shuffled labels. The 4B ablation also supports the joint design. Full ASGPO achieved 31.4%, compared with 22.9% when using concrete-state groups and 13.5% when using sparse step returns.

The theoretical analysis gives a condition for lower unnormalized advantage estimation error. Under independent peer sampling and a fixed return definition, abstraction helps when the variance reduction from additional peers exceeds the error introduced by grouping states with different continuation values. The fixed-size control complements this result by testing whether the choice of peers improves task success at the same group sizes.

Thank you again for your careful consideration.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.