acceptodds
Under review as a conference paper at ICLR 2027

MASCraft: Multi-Agent Specialization through Role–Skill Composition and Self-Evolving Orchestration

Abstract

Effective multi-agent systems (MAS) require agents specialized for different types of work. However, role prompting primarily specifies who an agent should act as, without necessarily providing sufficient procedural guidance on how to perform its work. We introduce MASCraft, a framework that combines role–skill composition with self-evolving orchestration for multi-agent specialization. By pairing independently defined roles with reusable skills, MASCraft equips specialists with both professional identities and explicit procedural guidance. An orchestrator agent plans subtasks, selects role–skill pairs from separate libraries, and organizes collaboration according to information dependencies. To better meet the specialized demands of MAS orchestration, we provide an explicit policy that evolves through execution experience. This policy comprises two textual modules: Environment Knowledge, which records task characteristics and recurring failures, and Team Strategy, which guides specialist configuration and coordination. MASCraft generates multiple rollouts per task, selects tasks with contrasting results for within-task reflection, and aggregates lessons across tasks to update the policy, guiding subsequent team construction. Extensive experiments on HLE, DeepResearch Bench, and BrowseComp-Plus show that MASCraft consistently outperforms competing MAS methods across different LLM backbones. Code is included in the supplementary material.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-Agent Systems, Multi-Agent Specialization, Role–Skill Composition, Self-Evolving Orchestration

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

This paper targets a limitation of MAS: existing approaches typically specify only the roles agents should play (who to act as), without providing concrete guidance on how to carry out their work (how to do the work), which limits overall system performance. To address this, the paper uses a strong model to construct role and skill libraries based on the validation set, which are then provided to an orchestrator. The orchestrator decomposes each input task into subtasks, assigns an appropriate role and skill to each subtask, and establishes the dependencies among subtasks, thereby organizing the execution of the entire MAS. To further improve orchestration, the paper self-evolves two natural-language experience modules, Environment Knowledge and Team Strategy, on the validation set. In each round, both modules are updated through contrastive reflection on multiple executions of the same task; after a fixed number of rounds, the version that performs best on the validation set is frozen for testing. Experiments show that the method improves MAS performance across multiple benchmarks, suggesting that both Role–Skill Composition and Self-Evolving Orchestration contribute to better task performance.

Critical strengths
  • The framework constructs role and skill libraries in advance, and the orchestrator dynamically selects appropriate role–skill combinations for each task. This addresses a limitation of existing MAS, which typically assign roles to agents without providing specific guidance on how to carry out their work.

  • The self-evolving Environment Knowledge and Team Strategy modules iteratively distill task characteristics and orchestration experience, providing additional guidance to the orchestrator and thereby enhancing its orchestration capabilities.

Critical weaknesses
  • Both the role and skill libraries are generated by GPT-5.6-sol, a model stronger than the three backbone models used in the MAS framework. Consequently, it is difficult to determine whether the observed performance gains stem from the Role–Skill Composition and Self-Evolving Orchestration mechanisms themselves, or simply from the additional knowledge provided by a more powerful model. It would be helpful to conduct supplementary experiments where these libraries are generated by weaker models, or directly by the backbone models themselves.

  • Both Environment Knowledge and Team Strategy are provided to the orchestrator in their entirety as natural-language documents, without any filtering. Since some of this information may be irrelevant to the current task, the method could benefit from an appropriate retrieval mechanism that filters out such extraneous content.

Questions for authors
  • Could you report the performance of MASCraft when the role and skill libraries are generated by weaker models?

  • As shown in Figure 3, the improvements from self-evolution vary considerably in stability across benchmarks. What do you consider to be the primary reason? Furthermore, the validation curves for DRB and BCP do not appear to have converged. Beyond running a fixed number of rounds and selecting the best-performing one on the validation set, are there more principled stopping criteria?

3 comments
AuthorsOpenReview ↗

Response to Reviewer GPFV

Thank you for the positive assessment and helpful suggestions. We address library construction, policy relevance, and evolution stability below.

R1. Dependence on the Library-Construction Model (W1, Q1)

Using the execution backbone to construct the libraries led to modest score decreases across the three benchmarks relative to GPT-5.6-sol. We evaluated four library-construction models while keeping GLM-5.2-Fast-Preview fixed as the execution backbone.

Library-construction modelHLEDRBBCP
GPT-5.6-sol19.6451.6036.90
Qwen3.8-Max19.6451.5235.71
GLM-5.2-Fast-Preview (self)18.4551.0935.71
Qwen3.8-27B16.0749.2329.76

These results suggest that a sufficiently capable construction model is important, but a stronger external model is not essential for comparable performance in these settings.

R2. Task Relevance of Policy Context (W2)

We include both policy modules in full for two reasons:

  • Compact size: The modules impose limited context overhead.
  • Generalizable content: The update rules consolidate recurring lessons across tasks while discouraging instance-specific details.

We will explore task-based retrieval and filtering to address potentially irrelevant guidance as the policies grow.

R3. Evolution Stability and Stopping Criteria (Q2)

(1) Stability across benchmarks. Performance may fluctuate across rounds because textual policy updates offer limited control over update direction and magnitude, which gradients and step sizes help specify in parameter optimization. We observed similar fluctuations in our reproductions of Skill-MAS and other baselines, consistent with reports in OPRO (p. 4, §2.3) [1], SIPDO (p. 6, Table 1) [2], and STEVE (p. 2, Figure 1(b)) [3]. Our comparison of successive policies suggests that overly aggressive updates may have contributed to some of the observed declines (see our response to Reviewer LWES, R2). DRB’s continuous, multidimensional scoring may help explain its smoother trajectory by capturing incremental improvements in coverage and organization.

(2) Additional rounds and stopping criteria. We extended evolution to 20 rounds on DRB and BCP but observed no substantial further gains on the held-out test sets. We will explore validation-based early stopping with a predefined improvement threshold and patience as an alternative to a fixed round limit.

Thank you again for your constructive feedback. We will incorporate the additional results and clarifications into the revised paper and welcome any further questions or concerns.

References

[1] Large Language Models as Optimizers. ICLR 2024.

[2] SIPDO: Closed-Loop Prompt Optimization via Synthetic Data Feedback. ICLR 2026.

[3] STEVE: Stabilizing Textual Gradient-Based Prompt Optimization via Error-Driven Refinement and Regularized Verification. Findings of IJCNLP–AACL 2026.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer,

Thank you again for taking the time to review our paper and share your feedback. Is there anything we could clarify in our response or any remaining concerns we could address?

Reviewer_GPFVOpenReview ↗

Response to the Authors

Thank you for your response. The additional experiments and clarifications have addressed most of my concerns. However, I believe my current score is already appropriate, so I will keep it unchanged.

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

This paper proposes MASCraft, a multi-agent framework combining role–skill composition with self-evolving orchestration. An orchestrator dynamically constructs task-specific teams by selecting roles, assigning reusable procedural skills, and defining information dependencies.

Critical strengths
  • Separating roles (who) from procedural skills (how) provides a practical and composable mechanism for configuring specialized agents. The ablations in Table 2 (p. 7) demonstrate that combining role and skill guidance generally outperforms removing either component.
Critical weaknesses
  • W1. The role–skill matching mechanism is not convincingly validated. Table 2(a) (p. 7) shows that Random Role achieves 19.05% versus MASCraft's 20.24% on HLE, while Random Skill drops to 13.69%. This suggests that performance may be driven primarily by providing useful procedural skills rather than learning effective role–skill composition; a skill-only baseline with matched orchestration and inference budgets is needed.

  • W2. The claimed self-evolution lacks evidence of effective policy learning. Section 2.3.2 selects tasks with contrasting rollout scores and uses LLM reflection to update textual policies, but Figure 3 (p. 8) shows fluctuating validation performance across evolution rounds. Since the best policy is selected retrospectively, the paper does not demonstrate consistent policy improvement or establish that the identified policy revisions cause better outcomes.

  • W3. The performance gains come with substantial computational overhead. Appendix C.1 (p. 27) reports 13.97× the token consumption of a single agent, versus 9.64× for the MAS baselines on average. Table 1 does not provide compute-matched comparisons, making it unclear whether MASCraft's improvements justify its additional computational cost.

Questions for authors
  • Q1. Given the small degradation under random role assignment (Table 2a), how much improvement can be attributed specifically to role–skill matching rather than skill availability and selection?

  • Q2. Can the authors report compute-matched performance and evaluate whether individual policy revisions consistently improve outcomes across independent rollouts?

6 comments
AuthorsOpenReview ↗

Response to Reviewer LWES

Thank you for raising these questions about roles, policy learning, and inference cost.

R1. Contribution of Role–Skill Composition (W1, Q1)

We include the skill-only baseline, “w/o Role,” in Table 2(a), with unchanged orchestration, tools, and budget limits. We apologize for not making this clear. Results below use GLM-5.2-Fast-Preview; Appendix Table 7 covers all backbones.

VariantHLEDRBBCP
w/o Role (skill-only)16.6751.5935.71
Random Role19.0551.4137.50
Random Skill13.6951.6333.33
MASCraft20.2452.2138.10
  • Role contribution: MASCraft outperforms skill-only on all three benchmarks.
  • Matching contribution: Task-matched roles outperform Random Role on all three benchmarks. Random Role still exceeds skill-only on HLE and BCP: high-level guidance may remain useful without precise matching. Skills specify execution procedures, so mismatches may be more disruptive, consistent with the larger Random Skill drops on HLE and BCP.

R2. Effectiveness and Stability of Policy Updates (W2, Q2)

(1) Independent Runs. Beyond Table 2(b), we froze one run’s policies at uniformly spaced rounds 0, 3, 6, and 9 for five independent test runs each. Mean ± sample SD:

BenchmarkRound 0Round 3Round 6Round 9
HLE14.88 ± 0.6016.19 ± 0.5020.12 ± 0.5015.24 ± 0.33
DRB51.05 ± 0.0451.82 ± 0.0351.94 ± 0.0352.23 ± 0.03
BCP34.52 ± 0.7335.59 ± 0.7833.57 ± 0.6837.98 ± 0.89

Across all three benchmarks, the mean test scores followed the same checkpoint-level trends as the validation scores in Figure 3. See our reply to Reviewer wvHM, R2, for selection sensitivity.

(2) Learning trajectory and declines. We compared successive policies, validation feedback, and execution plans across improving and declining rounds.

  • The Team Strategy examples below illustrate a consistent pattern of cumulative refinement rather than arbitrary perturbations. Successive revisions build on earlier guidance and make it more concrete (Figure 4 and Appendix C.3).

    BenchmarkCumulative refinement
    HLECheck answers → specify target quantities and independent solution methods → require calculation steps or classification tables
    BCPSearch again when evidence is missing → identify relevant clues → specify queries and require retrieval results
    DRBAssign research tasks → specify analytical methods → require supporting evidence and conditions for conclusions
  • Declining rounds tended to overly strengthen requirements or broaden their scope. BCP round 1 mandates independent search by the final agent regardless of upstream results, potentially duplicating work. By contrast, round 3, which improved performance, distinguishes searchable clues from verification-only constraints. Regressions may thus reflect excessive update strength rather than an ineffective learning direction.

(3) Learning need not be monotonic. We do not claim every revision improves performance. Textual updates lack numerical direction and step-size control. Similar fluctuations occur in our reproduced Skill-MAS and other baselines, OPRO (p. 4, §2.3) [1], SIPDO (p. 6, Table 1) [2], and STEVE (p. 2, Figure 1(b)) [3]. Following Skill-MAS [4] and metaTextGrad [5], we select policies on validation data and freeze them before testing.

R3. Performance Gains and Computational Cost (W3, Q2)

On BCP, MASCraft achieves higher accuracy with comparable or lower token usage. Below are accuracy (%) / test-time tokens (M) from Tables 1 and 6.

MethodDeepSeekQwenGLM
Self-Refine13.10 / 39.0114.88 / 41.5216.67 / 35.48
Skill-MAS23.21 / 26.9627.98 / 27.8535.12 / 35.43
MASCraft29.76 / 27.4737.50 / 25.5038.10 / 29.90
  • Single-agent comparison: Across all three backbones, MASCraft is more accurate and uses fewer tokens than Self-Refine. This suggests that multi-agent coordination is an effective way to translate test-time compute into performance gains.
  • MAS comparison: MASCraft outperforms Skill-MAS on all three backbones, using fewer tokens on Qwen and GLM and about 2% more on DeepSeek.

Thank you again. We will include these detailed results in the revision.

References

[1] Large Language Models as Optimizers. ICLR 2024.

[2] SIPDO: Closed-Loop Prompt Optimization via Synthetic Data Feedback. ICLR 2026.

[3] STEVE: Stabilizing Textual Gradient-Based Prompt Optimization via Error-Driven Refinement and Regularized Verification. Findings of IJCNLP–AACL 2026.

[4] Skill-MAS: Evolving Meta-Skill for Automatic Multi-Agent Systems. Findings of EMNLP 2026.

[5] metaTextGrad: Automatically Optimizing Language Model Optimizers. NeurIPS 2025.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer,

Thank you again for taking the time to review our paper and share your feedback. Is there anything we could clarify in our response or any remaining concerns we could address?

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer,

As the discussion deadline approaches, we would appreciate hearing whether our response has adequately addressed your concerns. We look forward to hearing from you and would be happy to address any remaining concerns.

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

This paper proposes MASCraft, a framework for LLM-based multi-agent systems that combines role–skill composition with self-evolving orchestration. It maintains separate role and skill libraries, assigns role–skill pairs to task-specific specialists, and organizes their information dependencies as a directed acyclic graph (DAG). Its orchestration policy consists of two natural-language modules, Environment Knowledge and Team Strategy, which are updated through within-task reflection and cross-task aggregation of execution feedback.

The authors evaluate MASCraft on HLE-Math, DeepResearch Bench (DRB), and BrowseComp-Plus (BCP) using three LLM backbones. MASCraft achieves the highest reported score in all nine backbone–benchmark combinations, with a reported mean relative improvement of 15.06% over the strongest baseline in each setting.

Critical strengths

1.A practical framework for agent specialization. Separating role descriptions from procedural skills is intuitive and potentially useful. Explicitly representing task dependencies and validating the resulting DAG also provides a structured approach to team construction.

2.Evaluation across multiple benchmarks and backbones. The paper evaluates mathematical reasoning, deep research, and multi-hop question answering using three LLM backbones. It includes several relevant baselines, including AFlow, EvoMAS, TacoMAS, and Skill-MAS.

3.Component ablations provide preliminary evidence of effectiveness. The results generally show that removing role–skill guidance or reverting evolved orchestration policies reduces performance. The baseline-augmentation experiment also suggests that role–skill guidance can improve Skill-MAS, providing some evidence that these components may be complementary.

Critical weaknesses

1.The methodological novelty over existing self-evolving orchestration methods remains insufficiently established. MASCraft's process of generating multiple executions, contrasting successful and unsuccessful trajectories, extracting reusable lessons, and revising natural-language orchestration guidance is closely related to Skill-MAS. Other baselines, including EvoMAS and TacoMAS, also use execution experience to improve multi-agent configurations or coordination.

The potentially distinctive aspects of MASCraft are its explicit role–skill composition and its separation of orchestration guidance into Environment Knowledge and Team Strategy. However, the paper does not sufficiently establish what new methodological insight these choices introduce beyond combining existing ideas in a unified framework. The component ablations demonstrate that the proposed components can be useful, but do not by themselves establish substantial methodological novelty.

A more direct analysis of the differences from Skill-MAS, including which mechanisms are unique and which experimental results isolate their respective contributions, would strengthen the paper.

2.The reliability and generalizability of the reported performance improvements are not sufficiently established. The test sets contain 168 examples for HLE-Math, 84 for DRB, and 168 for BCP. MASCraft evaluates the selected policy with one rollout per test task and does not report confidence intervals, statistical significance tests, or repeated-test-run variability.

The gains also vary considerably across settings. For example, MASCraft improves over the strongest baseline by only 0.21 points on Qwen's DRB setting, compared with 9.52 percentage points on Qwen's BCP setting. Thus, achieving the highest score in all nine settings is encouraging, but does not establish that every improvement is statistically reliable or robust to execution randomness and task sampling.

Furthermore, the authors select the best policy from ten rounds using relatively small validation sets, including only 16 DRB tasks. The paper would benefit from a sensitivity analysis of policy selection and additional uncertainty estimates, particularly for DRB and other settings with small performance differences.

3.The comparability and reproducibility of the Skill-MAS baseline require clarification. The reported Skill-MAS results differ from those in its original paper. For DeepSeek-V4-Flash, MASCraft reports 20.24% on HLE-Math, 47.68 on DRB, and 23.21% on BCP, whereas the original Skill-MAS paper reports 26.79%, 51.69, and 22.62%, respectively.

These differences should not be interpreted as evidence of an incorrect reproduction without further investigation. The exact evaluation subsets and implementation configurations have not been established as identical across the two papers. Moreover, the original Skill-MAS paper uses Gemini-3.1-Flash as its default LLM judge, whereas MASCraft uses GPT-5-mini-0807 for DRB and BCP. Differences in evaluation protocols can therefore affect cross-paper comparability.

MASCraft states that its baselines share the same tool settings and evaluation splits within its own experiments. However, the paper does not provide sufficient information to establish that the official Skill-MAS implementation, original configuration, and exact evaluation subsets have been reproduced. The authors should clarify these details and, where feasible, compare the methods under matched evaluation conditions. This would help determine whether the observed improvements reflect the proposed method rather than differences in experimental configuration.

4.The baseline comparison could be more comprehensive, and the cost–performance trade-off remains unclear. The paper includes several meaningful baselines, so the comparison is not fundamentally weak. Nevertheless, MAS-Orchestra is cited but is not included in the main experimental comparison. Given its focus on holistic multi-agent orchestration, the authors should justify its omission or provide a comparison if a fair evaluation is feasible. Other relevant approaches may also merit discussion, although not all are necessarily directly comparable under the current experimental setup.

The computational cost is also substantial. The paper reports that MASCraft consumes approximately 13.97 times the test-time tokens of a single-agent method on average, compared with 9.64 times for the multi-agent baselines. These figures exclude offline library construction and validation-time policy evolution. Although such costs may be amortized across future tasks, the paper does not establish the complete cost–performance trade-off of the proposed framework.

Questions for authors

1.What is the precise methodological distinction between MASCraft and Skill-MAS beyond the role–skill interface and the separation of orchestration guidance into Environment Knowledge and Team Strategy? Which experiments most directly establish the contribution of these design choices?

2.Can the authors provide the exact Skill-MAS implementation and configuration used in their experiments, together with the evaluation split identifiers and relevant scoring settings? Can the reported differences from the original Skill-MAS results be investigated under matched conditions?

3.Why is MAS-Orchestra not included in the main comparison despite being cited in the paper? If a fair comparison is not feasible, could the authors clarify the relevant differences in experimental scope?

4.Can the authors provide task-level results and uncertainty estimates, including paired statistical tests where appropriate, and evaluate the sensitivity of the conclusions to repeated execution and policy selection?

5.Can the authors report a more complete cost–performance analysis that accounts for policy evolution, library construction, model calls, and test-time inference?

4 comments
AuthorsOpenReview ↗

Response to Reviewer wvHM

Thank you for your detailed feedback.

R1. Methodological Insight (W1, Q1)

We acknowledge the need to clarify the scope of our contribution. Our methodological contribution is modular MAS specialization at the modeling level, rather than a new self-evolution algorithm at the technical level.

MASCraft builds task-specific specialists from independently reusable roles (responsibilities) and skills (execution procedures). The orchestration policy can assign several skills to a role and reuse a skill across roles. This allows the same libraries to support different tasks and lays the groundwork for independent evolution of roles, skills, and orchestration policies.

DistinctionSkill-MASMASCraftSupporting evidence
Specialist construction (main)Role-based specializationTask-specific role–skill compositionSkill-MAS + R–S (Table 2(c)), Table 1
Evolution scope (secondary)Execution (orchestration meta-skill)Perception and execution (Environment Knowledge + Team Strategy)Policy ablations (Table 2(b))

R2. Reliability of Performance Gains (W2, Q4)

Independent Runs

MASCraft significantly outperforms the baseline with the highest mean score in all nine model–benchmark settings (two-sided t-tests, all p < 0.05). The gains are 0.120 points over EvoMAS on Qwen–DRB (95% CI [0.001, 0.239], p = 0.0481). Each test uses five run-level scores per method (one from Table 1 and four from additional runs). Tables report mean ± sample SD.

DeepSeekHLEDRBBCP
AFlow10.00±0.7849.53±0.1019.16±0.78
EvoMAS13.57±0.7849.06±0.1317.86±0.73
Meta-Team12.02±0.6545.97±0.0417.38±0.65
TacoMAS11.07±0.5349.50±0.2123.45±0.68
Skill-MAS20.24±1.1947.61±0.0523.21±1.19
MASCraft26.78±0.8450.84±0.0730.35±0.84
QwenHLEDRBBCP
AFlow20.00±1.0047.56±0.2521.07±1.00
EvoMAS21.90±1.0749.17±0.0922.98±1.00
Meta-Team17.50±0.8045.86±0.0727.02±0.80
TacoMAS22.86±1.0046.85±0.1125.36±1.00
Skill-MAS22.50±0.2748.51±0.0527.86±0.27
MASCraft26.79±0.6049.29±0.0638.21±0.65
GLMHLEDRBBCP
AFlow10.00±1.0744.16±0.1916.31±0.90
EvoMAS17.98±0.5046.62±0.1525.12±0.50
Meta-Team11.31±0.7344.31±0.0628.33±0.33
TacoMAS8.93±0.7350.05±0.0430.36±0.73
Skill-MAS15.96±0.7850.49±0.0535.24±0.78
MASCraft20.33±0.5152.16±0.0537.59±0.64

Policy Selection Sensitivity

The two checkpoints with the highest validation scores both exceeded round 0 in mean test score on all three benchmarks. For each benchmark, we evaluated the three frozen policies from one GLM-5.2-Fast-Preview evolution run five times each on the same test split (mean ± sample SD). You can find a more detailed analysis of policy evolution in our response to Reviewer LWES (R2).

BenchmarkInitial policy (round 0)Validation rank 1Validation rank 2
HLE14.88 ± 0.6020.12 ± 0.50 (r6)19.64 ± 0.60 (r4)
DRB51.05 ± 0.0452.23 ± 0.03 (r9)52.11 ± 0.06 (r5)
BCP34.52 ± 0.7337.98 ± 0.89 (r9)36.31 ± 0.94 (r7)

R3. Skill-MAS Reproduction and Comparability (W3, Q2)

We use official Skill-MAS code (commit b55d47e) with aligned generation settings and validation/test instances. We retain official programmatic HLE scoring and use GPT-5-mini-0807 instead to judge DRB and BCP.

We tried rejudging the original DeepSeek-V4-Flash predictions with Gemini-3.1-Flash-Image-Preview in response to your comment about the judge. Skill-MAS’s BCP accuracy remained at 23.21%, and its DRB score rose from 47.68 to 49.03. MASCraft scored 52.34 on DRB, a 3.31-point lead. The remaining gaps may reflect differences in API providers or versions, which the Skill-MAS paper does not specify. All methods use DashScope, controlling for cross-paper API differences.

R4. Comparison with MAS-Orchestra (W4, Q3)

Our comparison covers harness-level adaptation with frozen, API-accessed LLMs. MAS-Orchestra’s parameter-level GRPO adaptation requires trainable model access, training data, and compute beyond this setting.

R5. Cost–Performance Trade-off (W4, Q5)

  • Policy evolution uses about 10× the test-time tokens across rounds and rollouts. Averaged across backbones and summed over benchmarks, MASCraft uses 391.8M evolution tokens versus about 455.0M for Skill-MAS, a 13.9% reduction alongside higher test scores.

  • Library construction adds about 2.9M tokens across benchmarks (<1% of evolution usage). This cost can be amortized across tasks, and deployments can reuse established business libraries.

We will incorporate these results and clarifications in the revision.

AuthorsOpenReview ↗

Follow-up on our response

Dear Reviewer,

Thank you again for taking the time to review our paper and share your feedback. Is there anything we could clarify in our response or any remaining concerns we could address?

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer,

As the discussion deadline approaches, we would appreciate hearing whether our response has adequately addressed your concerns. We look forward to hearing from you and would be happy to address any remaining concerns.

AuthorsOpenReview ↗

Final Follow-up on Our Response

Dear Reviewer,

As the discussion period closes today, could you take a look at our response and let us know if any concerns remain? We would be happy to provide any further clarification.

Thank you for your time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.