acceptodds
Under review as a conference paper at ICLR 2027

RLVR: Reinforcement Learning with Verifiable Rubric-based Ranking

Abstract

Reinforcement Learning with Verifiable Rewards (RLVR) is expanding from tasks with relatively well-defined correctness signals, such as mathematics and code, toward multifaceted quality requirements, where quality is typically specified by multi-dimensional rubrics rather than a single criterion. Since policy optimization generally consumes one scalar per rollout, rubric-based RL pipelines must map multiple criterion scores into a scalar reward. This aggregation step is often treated as score scaling, although it implicitly determines how different quality dimensions trade off during training. A prevailing practice normalizes each criterion and takes a linear aggregation, which implicitly assumes that cardinal score differences are comparable across criteria and that gains on one criterion can compensate for failures on another—assumptions that are unreliable when rubric criteria are semantically heterogeneous. We propose Reinforcement Learning with Verifiable Rubric-based Ranking (RLVR), a verifiable ranking paradigm for rubric-based RLVR. For each criterion, RLVR converts rubric scores into criterion-specific within-group ordinal outcomes and recovers a latent utility from the resulting within-group comparison matrix; these utilities are then converted into a single optimization signal for policy training. By retaining only within-group ordering and discarding raw score magnitudes, RLVR avoids calibrating heterogeneous rubric scales. RLVR further supports objective-preserving attribute adjustment: auxiliary attributes that are systematically associated with observed rankings but are not themselves training objectives can be incorporated into the estimation process without expanding the rubric or directly rewarding those attributes. Across three model scales and 16 benchmarks, RLVR consistently outperforms a broad range of representative rubric-based RLVR baselines, achieves the best overall performance on a majority of benchmarks at every scale. Further analysis shows that it controls systematic effects associated with reasoning efficiency and response formatting while preserving the underlying quality objective.

Primary area:
reinforcement learning
Keywords:
Reinforcement Learning, Large Language Models

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_pDgiOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies the heterogeneous aggregration issue in rubric-based reinforcementLearning. It proposes Reinforcement Learning with Verifiable Rubric-based Ranking (RLVR2), which converts raw rubric scores into pairwise comparisons within group and applies BT model to estimate the latent utility. The proposed method improves performance on several benchmarks.

Critical strengths
  1. The research topic is meaningful. Scores of different rubric criteria may come from binary score, discrete score or continuous score. These heterogeneous scores may not have same semantic meaning to be compared.

  2. The paper conducts extensive experiments on multiple benchmarks and achieves competitive results.

  3. Besides aggregration, the paper also investigates systematic correlations between rewards and auxiliary attributes, such as reasoning length and response formatting, and analyzes the corresponding behavioral changes during training.

Critical weaknesses
  1. The novelty is not clear compared with existing pairwise rubric reward modeling works, such as [1-3]. Although these papers are not the same as the proposed method exactly, they have covered important part of this paper's motivation. This paper should discuss the advantages and differences compared with these works clearly.

  2. The paper lacks some simple while important baselines. First, it should compare with pairwise rubric evaluation. For example, it can ask judge compare two answers for each criterion and then aggregate criterion-level preferences as RL rewards. Second, it should add simple ordinal aggregation as baselines, such as mean rank and pairwise win rate.

  3. The BT model may suffer from complete separation, particularly for binary or discrete criteria, leading to non-finite maximum likelihood estimates because the pairwise outcomes are deterministically derived from rubric scores. The paper should clarify how this issue is handled and whether the resulting utilities are sensitive to regularization or optimization choices.

  4. The necessity of the proposed attribute adjustment mechanism is not justified sufficiently. The paper argues that introducing additional rubric criteria to control reasoning length and formatting may create new optimization targets and increase the risk of reward hacking. However, these issues could potentially be addressed by explicitly incorporating criteria for reasoning efficiency, conciseness, or formatting appropriateness into the existing rubric framework. Conducting an experiment between these two approaches would be healpful.

[1] Jia R, Yang Y, Wang W, et al. Open rubric system: Scaling reinforcement learning with pairwise adaptive rubric[J]. arXiv preprint arXiv:2602.14069, 2026.

[2] Niu W, He W, Xie Z, et al. From Absolute to Relative: Rethinking Reward Shaping in Group-Based Reinforcement Learning[J]. arXiv preprint arXiv:2601.23058, 2026.

[3] Juneja G, Phal S M, She J, et al. Tournament Style RL: Stabilizing Policy Optimization on Non Verifiable Problems[C]//Forty-third International Conference on Machine Learning.

Questions for authors

refer to weakness

6 comments

We sincerely thank Reviewerfor the thoughtful feedback devoted to reviewing our work. We address each point below and will incorporate the corresponding clarifications and improvements into the manuscript and supplementary material.

❓Q1. How does RLVR$^{2}$ differ from recent pairwise-rubric and relative-reward methods?

💡 Thank you for pointing out these closely related works. They all demonstrate the value of relative signals, but they obtain and use those signals through different reward-construction mechanisms. Open Rubric System dynamically instantiates rubrics and uses an LLM judge to produce criterion-wise pairwise preferences. From Absolute to Relative trains a Ranking Reward Model to produce group-level rankings, introducing an additional learned ranking component. Tournament Style RL compares policy responses against an externally generated set of anchor responses and uses the resulting win rates as rewards.

RLVR$^{2}$ instead starts from the existing criterion-level outputs of verifiable rubric verifiers. For each query, it treats the sampled rollout group as a self-contained comparison set, deterministically converts each criterion's scores into within-group win/tie/loss outcomes, and fits a criterion-specific BT utility before fusing those utilities. RLVR$^{2}$ therefore requires no additional learned ranking model, dynamically generated rubric, or external anchor set. Its contribution is a verifiable aggregation layer for heterogeneous rubric rewards: it preserves within-criterion ordinal information while converting multiple verifier signals into a single scalar reward for policy optimization. As shown in the table below, the performance of the proposed method remains robust even when a different language model is employed as the judge.

❓Q2. Does RLVR$^{2}$ outperform simple ordinal aggregation and direct pairwise rubric evaluation?

💡Thank you for raising this question. As described in our response to Reviewer ZvvJ’s Q1, we additionally include a direct LLM-ranking baseline in the aggregation comparison. This baseline asks an LLM to rank rollout responses without access to an explicit rubric, criterion definitions, or criterion-level verifier outputs. It performs worse than the rubric-informed methods, potentially because its judging standard is underspecified and may rely more heavily on superficial or exploitable attributes, such as response length and formatting. The other baseline methods also achieve lower average performance than RLVR$^{2}$, suggesting that our method is more effective for this complex task.

Aggregation methodSysBAdvIFIFEIFBMChalSQA-vCSQALB-v2LProcHLMFrmAA-LGPQAMMLUAH-v2EQB3Average
Mean-score70.0061.5090.0066.5052.1016.0062.8052.5076.0025.0063.3068.0070.0080.4072.0077.0062.69
Pairwise win rate71.0063.5091.9067.8054.8015.8063.0052.5076.5025.0063.2069.0070.6080.4074.8078.8063.66
RLVR$^{2}$73.4863.6192.0569.6955.4011.1068.4333.8081.7528.1875.7358.6780.8180.4871.1087.1564.46
RLVR$^{2}$ with Qwen2.5-32B judge72.8063.3091.9069.2056.8011.8067.8034.0081.8028.0075.2061.5077.5080.7070.7087.2064.39
LLM-based aggregator71.5063.0092.2069.0055.8015.7062.9053.0076.2025.2062.4069.1070.5080.4075.5079.7063.88
LLM ranking69.0060.0088.0066.0054.5015.2061.8050.5073.0024.0060.5066.5068.5078.5070.0076.0061.38

❓Q3. How does the BT implementation handle complete separation, and how sensitive are the results to regularization and optimization choices?

💡 Thank you for raising this question. We clarify that the BT output is an intermediate per-criterion, per-rollout signal rather than the final policy reward. Each criterion is fitted separately from its within-group comparison outcomes, with ties represented as half wins. The resulting criterion-level signals are aligned by rollout, summed across the active dimensions, and min--max normalized within the group before being written back as the GRPO reward. Thus, the final reward is bounded in [0,1][0,1] under the normal fusion path, and policy optimization depends on the resulting group-relative ordering and advantages rather than on the absolute scale of an individual BT utility. We also shared core code of our methods for you to refer https://anonymous.4open.science/r/rlvrr-B001/.

❓Q4. Is attribute adjustment necessary, or could length and formatting simply be added as ordinary rubric criteria?

💡Thank you for raising this important distinction. We chose reasoning length and formatting as representative attributes because they are two observable properties that may influence quality judgments without necessarily reflecting substantive task quality. They also allow us to demonstrate that the proposed adjustment can address different types of response behavior. The framework can be extended to other measurable attributes according to the needs of a particular task; these two examples are not intended to be an exhaustive or optimized attribute set. Our existing reasoning-efficiency and formatting-control experiments demonstrate this extensibility while preserving the capability-oriented rubric objective.

Adding these attributes as ordinary reward criteria is possible, but it changes them into direct optimization targets. A policy may then obtain higher reward by changing response length or formatting even when the underlying task quality does not improve, creating an additional shortcut for optimization. Our attribute-adjustment mechanism provides an alternative for post-training requirements that concern such observable response properties: the property can be introduced as a covariate, and its association with the observed ranking can be adjusted without adding a new reward dimension or redesigning the underlying rubric. This allows the training process to accommodate additional behavioral requirements while preserving the original quality objective. We will clarify this objective-preserving role in the revised manuscript.

Comments, Suggestions And Typos: We sincerely thank Reviewer pDgi again for the helpful feedback, which has greatly helped us improve the paper. In the revision, we will make the following changes:

  • Clarify the distinction from relative-reward methods in Section 5 related work (Q1).
  • Add additional baseline into Appendix.
  • Expand the discussion of objective-preserving attribute adjustment in Section 3.2.

We are very happy to answer any further questions. We sincerely hope that these improvements will address the concerns and that the reviewer may kindly reconsider the score and support the paper’s presentation at the conference.

Thank you for the detailed rebuttal and additional experiments. While the revisions address some concerns, several key issues remain unresolved, particularly the sensitivity of BT utility estimation to regularization, the lack of a direct rubric-aware pairwise comparison baseline, and the absence of a comparison between attribute adjustment and explicit rubric criteria.

I appreciate the authors' efforts but will maintain my original rating.

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

Common rubric training methods use binary (or partial credit) rewards per rubric item, and then combine them all into a final scalar reward (usually through linear aggregation) individually for each rollout, and then combine them into advantage. This paper, however, takes advantage of the within-rubric rankings and fits a bradley-terry utility per item. They run exhaustive experiments covering multiple datasets and different model families and show on average this method is better than traditional approaches.

Critical strengths

The idea is very intuitive (RL amplifies differences between rollouts, so we need to account for this across within each criteria more explicitly). Using a bradley-terry and the within-metric comparison metrix makes sense and is well executed.

One great advantage of this method is that it's not limited to binary or weighted criteria. It can adapt to any kind of measurement (rankings, pass/fail, partial credit, etc) because they use rankings not individual scores.

The experiments are also robust, enough coverage of benchmarks, and also present a reasonable ablation study. I think this is a good paper.

Critical weaknesses

I think it would be nice to see some comparisons with some basic baselines. For example, using an LLM to aggregate the rankings instead of fitting a bradley-terry. Even if this scores higher, the paper's approach should be more efficient cost/compute wise, but it would make the paper more comprehensive. Or even simpler baselines like a monotone function of that criterion’s rank within the group.

It's not clear what models are used in Table 1 and average across different model families, if that's the case, is not great and not explained. Figure 3 (Reasonign Efficiency Control) is also missing capability scores. It's very easy to train a model that is concise but underperforms. The more interesting contribution would be having same performance at the same or shorter length. I think overall some of the objective attribute adjustments feel arbitrary and I didn't see enough justifications on why these choices were made and why they are optimal.

Questions for authors

none

2 comments

We sincerely thank you for the positive feedback to our work. We are especially grateful for the recognition of the motivation, broad experimental coverage, and practical value of the proposed reward-construction interface. We address the requested comparisons and clarifications below, and will incorporate the corresponding changes into the paper.

❓Q1. How does RLVR$^{2}$ compare with the LLM-based and rank-based methods?

💡 Thank you for suggesting these comparisons. We compare RLVR$^{2}$ against four representative aggregation baselines under matched 9B training settings: (1) mean-score aggregation, which averages the normalized scores across rubric criteria; (2) pairwise win-rate aggregation, a deterministic ordinal rule that converts each criterion into within-group win/tie/loss outcomes and averages the resulting per-response win rates; (3) an LLM-based aggregator (Qwen2.5-32B), which produces an overall ranking for each rollout group from the rubric definitions and criterion-level scores; and (4) an LLM-judge variant of RLVR$^{2},whichusesQwen2.5−32Btoproducecriterion−leveljudgmentswhileretainingRLVR, which uses Qwen2.5-32B to produce criterion-level judgments while retaining RLVR^{2}$'s ordinal conversion and criterion-wise utility fusion. All methods use the same criterion-scoring pipeline, rollout groups, policy-training budget, and evaluation settings.

Aggregation methodSysBAdvIFIFEIFBMChalSQA-vCSQALB-v2LProcHLMFrmAA-LGPQAMMLUAH-v2EQB3Average
Mean-score70.0061.5090.0066.5052.1016.0062.8052.5076.0025.0063.3068.0070.0080.4072.0077.0062.69
Pairwise win rate71.0063.5091.9067.8054.8015.8063.0052.5076.5025.0063.2069.0070.6080.4074.8078.8063.66
RLVR$^{2}$73.4863.6192.0569.6955.4011.1068.4333.8081.7528.1875.7358.6780.8180.4871.1087.1564.46
RLVR$^{2}$ with Qwen2.5-32B as judger72.8063.3091.9069.2056.8011.8067.8034.0081.8028.0075.2061.5077.5080.7070.7087.2064.39
LLM-based aggregator71.5063.0092.2069.0055.8015.7062.9053.0076.2025.2062.4069.1070.5080.4075.5079.7063.88

The results show that our methods outperform the other baselines. Using Qwen2.5-32B for criterion-level judging while retaining RLVR$^{2}′sfusionyieldsanaveragescoreonly0.07pointsbelowtheoriginalRLVR's fusion yields an average score only 0.07 points below the original RLVR^{2}$--BT setting. Although performance declines on some benchmarks, the overall result remains strong. Due to the limit of the time we only did 9b experiment, we will expand to all model size and add to paper later.

❓Q2. Which models are used in Table 1, and how is the reported average computed?

💡Table 1 reports separate blocks for Qwen3.5-4B, Qwen3.5-9B, and GLM-4.5-Air (106B-A12B). The Average column is computed over the 16 listed benchmarks. All methods within a block use the same backbone, training queries, active rubric criteria, verifiers, judges, and optimization settings, and differ only in reward aggregation. We make this statistical convention explicit in the Table 1 caption.

❓Q3. Does reasoning-efficiency control preserve task capability, or does it simply make responses shorter?

💡Thank you for raising this point. Our results show that the adjustment does more than simply shorten responses: it reduces reasoning length without a consistent drop in overall task performance. Figure 3 shows the change in thinking-token usage, while Table 2(b) reports the corresponding capability scores across instruction following, factuality, long-context understanding, reasoning, knowledge, and open-ended generation. For λ∈{0,0.25,0.50,0.75}\lambda \in \{0, 0.25, 0.50, 0.75\}, the overall scores are 67.92, 68.30, 67.98, and 68.57, respectively, compared with 68.32 for the RLVR$^{2}$+ reference. We make this relationship explicit in the revised figure caption and discussion.

❓Q4. What motivates the choice of auxiliary attributes, and are these choices optimal or exhaustive?

💡As stated at the beginning of Section 3.2, reasoning length and formatting are representative examples used to demonstrate the extensibility of the attribute-adjustment mechanism, rather than an exhaustive or optimized set of attributes. We selected them because they are directly measurable and capture two practical sources of reward variation: extended deliberation that may not improve substantive quality, and structural formatting that may attract favorable judgments without adding information. The formulation is not restricted to these examples: other measurable auxiliary attributes can enter the same comparison model. For example, when response conciseness is desired as an auxiliary property, response length can be controlled analogously to reasoning-efficiency adjustment through the proposed framework. Given the page limit, we focus on these two examples in the paper to make the mechanism concrete.

Comments, Suggestions And Typos: We again thank you for the high score and constructive suggestions. We will add the aggregation comparison and per-benchmark results to the experimental section.

  • Adding further baseline result to Appendix .
  • Revising Section 3 for expand the explanation of representative auxiliary attributes

We really appreciate the reviewer’s support, which has helped us make the paper clearer and more complete.

Reviewer_JHszOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates multi-dimensional rubric score reward aggregation in RLVR. The authors argue that while existing methods typically convert multi-dimensional scores into a single reward via normalization and weighted summation, different rubric dimensions often vary significantly in scale and semantics. To address this, the paper proposes converting rubric scores into within-group pairwise rankings and estimating latent utility via the Bradley–Terry model to generate the final reward for reinforcement learning. Additionally, the paper introduces attribute adjustment to mitigate the influence of non-target attributes (e.g., length and formatting) on reward estimation, along with quality-level gating to handle key quality constraints. Experiments across three model scales and multiple benchmarks validate the effectiveness of the proposed approach.

Critical strengths
  1. The paper addresses multi-dimensional rubric reward aggregation in RLVR, which carries solid research significance and practical value.

  2. The evaluation is extensive, spanning three model scales—Qwen3.5-4B, Qwen3.5-9B, and GLM-4.5-Air (106B)—and comparing against existing baselines across numerous benchmarks.

  3. The experimental analysis is comprehensive and includes thorough ablation studies.

Critical weaknesses

1.The core methodology relies on converting rubric scores to pairwise rankings and estimating latent utility via the Bradley–Terry model. However, both the Bradley–Terry model and pairwise preference learning are well-established methods, making the paper's algorithmic or theoretical novelty somewhat limited.

  1. The rationale behind the weighting mechanism requires further clarification and justification.

  2. The full method incorporates ranking-based aggregation, attribute adjustment, and quality-level gating. Experimental results show that RLVR²+ significantly outperforms RLVR² across multiple model scales (especially on the 106B model). Detailed ablation studies isolating these components should be provided.

  3. The paper uses ~120K training queries containing substantial proprietary internal data, providing only category-level counts and statistics. Moreover, rubric design, dimension weights, and verifier quality can heavily impact reward aggregation performance. Despite evaluations across model scales, further validation is needed to demonstrate robustness against varying rubric designs, scoring noise, and evaluation criteria.

  4. The code does not appear to be fully open-sourced, as the supplementary material contains only a single file.

Questions for authors

see the weaknesses

6 comments

We thank the reviewer for recognizing the significance of heterogeneous rubric aggregation, the breadth of our evaluation, and the extensive ablation analysis. We address the concerns below.

❓Q1. What is the methodological novelty beyond combining within-group ranking with the classical Bradley--Terry model?

💡  Our novelty lies in formulating and solving a distinct reward-aggregation problem in rubric-based RLVR. Specifically, existing rubric-based RLVR methods typically normalize and combine criterion-level scores in cardinal space, implicitly treating score differences from binary, discrete, and continuous criteria as comparable. RLVR$^{2}$ addresses the heterogeneous aggregation problem by converting the output of each criterion into within-rollout-group ordinal outcomes and recovering a criterion-specific latent utility before fusion. This changes the reward-construction interface from aggregating heterogeneous cardinal scores to aggregating criterion-wise ordinal utilities, while allowing the fusion weights to represent criterion importance rather than compensate for arbitrary score scales.

BT serves as the latent-utility estimator within this formulation; the methodological novelty lies in the aggregation framework built around it. Whereas adaptive weighting methods such as Focal address imbalanced optimization across criteria, RLVR$^{2}$ addresses the comparability of the criterion signals being aggregated. These are complementary aspects of reward construction. The framework further supports attribute adjustment to control systematic associations with reasoning length and formatting without introducing them as additional reward objectives. We will make this distinction explicit in the paper and supplement the evaluation with simple ordinal baselines to isolate the contribution of BT utility estimation.

We clarify this distinction in the related-work and add ordinal aggregation baselines to appendix C.1.

❓Q2. Could you clarify and justify the rationale behind the weighting mechanism?

💡 Our weighting design separates quality constraints from criterion-level discrimination. In our experiments, ordinary rubric criteria use uniform fusion weights. Critical quality violations are handled separately by a hard gate, while other low-quality conditions reduce the fused reward through a soft penalty. This gives essential quality requirements explicit priority without manually assigning different importance weights to every ordinary criterion. Converting scores into ordinal comparisons before utility estimation also prevents arbitrary score ranges from determining a criterion’s contribution.

With uniform weights, a criterion’s effective contribution is determined by the utility differences it produces within the current rollout group. A criterion on which all responses tie contributes no utility difference, whereas one that distinguishes responses provides a relative optimization signal. These contributions change as the policy and its sampled responses evolve, yielding group-adaptive effective contributions without explicitly changing the criterion weights. The design therefore combines explicit quality control with adaptation to the comparative information available in each group.

❓Q3. Could you provide ablation studies to isolate the contributions of ranking-based aggregation, attribute adjustment, and quality-level gating, particularly on the 106B model?

💡 The main table already compares RLVR$^{2}$ with RLVR$^{2}$+,where “+” denotes quality-level gating, across all three model scales. The paper also includes separate experiments on reasoning-efficiency control and formatting control, together with capability evaluations. To make the contribution of each component explicit, we summarize the controlled ablations below.

We have added controlled ablations on the 106B model to evaluate reasoning-length adjustment, formatting adjustment, and their combination, each with and without quality-level gating. All settings use the same ranking-based aggregation. Given the limited rebuttal period, we prioritized the 106B model, which the reviewer specifically highlighted, and will extend these ablations to the 4B and 9B models.

VersionLength adj.Format adj.GateSysBAdvIFIFEIFBMChalSQA-vCSQALB-v2LProcHLMFrmAA-LGPQAMMLUAH-v2EQB3Average
AOffOffOff69.1264.6493.1670.7153.1913.8053.9763.8584.0234.3681.1964.0077.2073.5566.7486.0465.60
BOffOffOn73.2067.3294.6473.1360.0415.2056.0765.7784.1235.3181.1967.6777.8680.4367.1487.4267.91
COnOffOff68.7264.3492.7670.3153.3914.1054.5063.6583.7234.0480.8763.5076.9073.8066.4485.7665.42
DOffOnOff68.8264.4492.8670.3153.2914.0554.3263.5583.8234.1680.9963.7077.0073.7066.5485.8665.46
EOnOnOn72.7066.9294.1472.6360.3415.5556.5565.4783.7234.8981.4967.1777.5680.4866.7487.0467.71
VersionMean thinking lengthHeading (%)Bold (%)Ordered-list (%)Unordered-list (%)
A1051711.828.457.1116.42
B1146312.128.977.6516.45
C890811.548.417.0816.33
D1004610.717.666.7716.24
E941310.197.386.4516.36

The result suggests that length adjustment, formatting adjustment, and their combination have only a small effect on the main capability score when gating is disabled, whereas quality-level gating provides the larger improvement by suppressing responses that violate critical requirements. The behavioral estimates are consistent with the intended mechanisms: reasoning-length adjustment reduces mean thinking length by roughly 1.2K tokens, formatting adjustment primarily reduces formatting densities, and gating changes these behavioral quantities only slightly. Thus, the components play different roles: attribute adjustment controls systematic nuisance variation, while gating enforces non-compensatory quality constraints.

❓Q4. How do you address data transparency and robustness to rubric design, scoring noise, and verifier quality??

💡 Thank you for raising these questions. Our anonymous repository (https://anonymous.4open.science/r/rlvrr-B001/) provides sanitized data examples together with their rubric definitions, score domains, standard fusion strategy, and evaluation-model configuration. It also provides the core code of RLVR$^{2}$which can be easily integrated as a plugin into common RL training systems such as verl. The training setup uses the same fixed rollout groups, criterion-level verifier outputs, criterion weights, tie margins, and aggregation settings across the compared methods.

The appendix subsection C already evaluates different aggregators on identical logged rollout groups under three order-preserving score transformations. RLVR$^{2}$ preserves the fused reward ordering and advantage signs in all three tested settings. This invariance holds when the transformation preserves the complete pairwise outcome matrix, including tie decisions, with all other aggregation settings fixed.

❓Q5. Will the code be publicly available?

💡 Thank you for pointing out the limited scope of the original supplementary artifact. We have now provided an anonymous repository containing code and dataset examples: https://anonymous.4open.science/r/rlvrr-B001/. We have also expanded the appendix with domain-specific aggregation strategies to clarify how the method is applied across the training domains.

Comments, Suggestions And Typos: We greatly appreciate your detailed feedback and constructive suggestions. We have sincerely followed your advice to revise the paper and supplement the experiments. To address the concerns, we have:

  • Revising Section 5 Related Work for Q1.
  • Checking and correcting grammar and spelling errors throughout the paper.

We will always willing to solve your concern. Please consider improving our score and allowing us to present at conferences.

Comments on the paper

2 comments

Dear Area Chair and Reviewers,

We sincerely thank you all reviewers for the time and effort devoted to evaluating our work. We also appreciate the constructive and encouraging feedback, which helped us clarify the method and strengthen the manuscript. In response to the reviewers' comments, we revised the paper and marked the newly added or substantially revised text in purple. The main revisions include:

  • revising the methodology to explain the proposed component roles and objective-preserving attribute adjustment more clearly;
  • adding component ablations to the appendix to analyze the effects of length adjustment, formatting adjustment, and quality-level gating; and
  • adding more aggregation baselines to provide a stronger comparison in the appendix.

Detailed responses to each concern are provided in the corresponding sections below for Reviewers.

Dear Reviewers,

As the discussion period is coming to an end, I kindly encourage you to review and respond to the authors' rebuttal as soon as possible.

Thank you very much for your time and valuable contributions to the review process.

Best regards,
Area Chair

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.