acceptodds
Under review as a conference paper at ICLR 2027

Rewrite What Matters: Adaptive Multilingual Query Rewriting for Reasoning via Agentic Reinforcement Learning

Abstract

In multilingual scenarios, queries with equivalent semantics but in different languages could guide the model into different reasoning trajectories, leading to performance disparities. Previous studies typically apply a one-size-fits-all query rewriting strategy, such as translation, which overlooks the fact that different scenarios requires diverse types of semantic transformations. In this paper, we propose mRewriter-R1, an agentic multilingual query rewriting framework with reinforcement learning. Unlike single-turn rewriting, mRewriter-R1 formulates multilingual query rewriting as a multi-turn sequential decision-making process, where the model dynamically performs multi-aspect optimization through adaptive operator selection. Experimental results demonstrate that \ours outperforms all strong multilingual rewriting baselines on different large reasoning backbones. Further analyzes show that the learned policy can adaptively decide on rewriting operators according to query characteristics, exhibiting strong generalization ability across diverse reasoning tasks, and plug-and-play compatibility with heterogeneous reasoning language models.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
multilingual reasoning, multilingual query rewriting

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

This paper proposes mRewriter-R1, an adaptive multilingual query rewriting framework based on agentic reinforcement learning (RL), aiming to improve the reasoning performance of large language models across different languages.

Critical strengths

The strongest aspect of this paper is its encouraging empirical performance across multiple multilingual reasoning benchmarks. As reported in Table 1, mRewriter-R1 achieves average accuracies of 73.9% on MGSM, 77.6% on XCOPA, and 80.2% on BELEBELE, outperforming the strongest reported baselines by 4.2, 1.7, and 2.6 percentage points, respectively.

Critical weaknesses

W1. Insufficient evidence that RL-based adaptive operator selection is responsible for the performance improvements The main weakness is that the experimental design does not adequately isolate the contribution of the proposed adaptive RL mechanism.

The paper emphasizes learning when and how to apply different rewriting operators. However, according to Section 2.2 and Appendix G, the three rewriting aspects are executed in a predefined order: language, structural, and semantic. The model therefore learns operator selection within each aspect, rather than learning the overall sequence of aspects or the number of rewriting stages.

More importantly, the existing comparisons primarily involve direct inference or previously proposed rewriting approaches. Although the ablation studies show that removing operators or individual aspects degrades performance, these experiments do not establish whether reinforcement learning provides substantial advantages over simpler approaches using the same operator space.

W2. Incomplete validation of semantic faithfulness and inconsistencies in the experimental protocol A second important concern is whether the rewriting process reliably preserves the original problem semantics and whether the experimental setup is sufficiently documented. The proposed method is optimized using downstream answer correctness rather than an explicit semantic-preservation objective. Consequently, improved answer accuracy does not necessarily guarantee that the rewritten query faithfully preserves all the information in the original input.

Although the authors provide a semantic-preservation analysis, the reported results indicate that critical-number token recall is 93.28%, while only 89.07% of queries retain all original numerical tokens. These statistics do not necessarily imply semantic errors, since some numerical transformations may be harmless. Nevertheless, they suggest that numerical information is not always preserved verbatim. The paper primarily relies on token-level preservation statistics, semantic similarity, and a limited number of qualitative examples. These evaluations are insufficient to establish whether the rewriting process consistently preserves mathematical relationships, logical conditions, and the original question intent, particularly in cases where numerical information is omitted or modified.

This issue is especially important for mathematical reasoning, where small changes to numerical constraints can fundamentally alter the problem.

Questions for authors
  1. Could the authors provide controlled comparisons against simpler methods using the same rewriting operators and three-stage pipeline, such as a training-free LLM-based operator selector, a heuristic operator-selection strategy, or a fixed rewriting pipeline without RL training?
  2. Could the authors clarify the exact MGSM training and evaluation splits, particularly the treatment of Telugu?
5 comments

We sincerely thank the reviewer for the thoughtful evaluation and constructive feedback. We carefully address the concerns below and provide additional experimental evidence and clarifications to strengthen our work.

W1 & Q1.

Thank you for this insightful comment. We first clarify that the three rewriting aspects follow a predefined order (Language → Structural → Semantic), while the operator within each aspect is adaptively selected based on the current query state and preceding rewriting outcomes. Thus, our adaptivity lies in query-dependent operator selection and sequential rewriting, rather than learning the ordering or number of rewriting aspects. We will clarify this distinction in the revised manuscript.

To address your concern, we introduce two additional controlled baselines:

  • Non-RL Adaptive Pipeline: An untrained Qwen2.5-7B-Instruct model adaptively selects operators using the same three-stage workflow and operator space.
  • Fixed Operator Pipeline: A fixed combination of operators is selected on a held-out validation set and applied to all queries through the same three-stage workflow, without RL training.

These experiments help distinguish the benefit of the fixed multi-stage framework from the additional improvements associated with adaptive selection and RL optimization.

Method             En    Bn    De    Es    Fr    Ja    Ru    Sw    Th    Zh    Te    Avg
Non-RL Adaptive   86.8  72.0  82.4  83.2  82.4  71.2  86.0  18.4  80.4  84.0  35.6  71.12
Fixed Operator    88.0  68.4  76.4  79.6  76.0  69.6  81.6  20.8  74.4  82.4  32.8  68.18
Ours              92.8  74.4  83.2  84.8  84.4  78.8  86.8  24.8  79.2  84.0  40.0  73.90

As shown above, Non-RL Adaptive outperforms Fixed Operator by 2.94 percentage points, demonstrating the benefit of query-dependent operator selection over a fixed rewriting strategy. Moreover, mRewriter-R1 further improves over Non-RL Adaptive by 2.78 points, indicating that RL optimization provides additional gains beyond training-free adaptive rewriting.

W2.

Thank you for raising this important concern. We clarify that the previously reported 89.07% measures exact numerical-token preservation rather than numerical semantic correctness. Differences in numerical tokens may arise from semantically equivalent transformations rather than actual information loss.

To address this concern, we conduct two additional analyses. First, we examine queries with numerical-token changes to distinguish semantically equivalent transformations from genuine alterations of numerical information. Second, we directly evaluate the preservation of three key elements essential to mathematical reasoning: mathematical relationships, logical conditions, and the original question intent. We assess semantic equivalence by comparing each original query with its rewritten version using an independent LLM-based evaluator, without providing ground-truth answers or downstream predictions.

MetricResult (%)
Semantically Valid Numerical Transformations92.69 (279/301)
Mathematical Relation Preservation88.76 (2441/2750)
Logical Constraint Preservation87.16 (2397/2750)
Question Intent Preservation88.47 (2433/2750)

The results show that most numerical-token changes correspond to semantically equivalent transformations rather than genuine numerical errors, explaining why exact token preservation underestimates numerical faithfulness. Furthermore, the additional evaluations provide quantitative evidence that the rewriting process preserves the key mathematical relationships, logical conditions, and question intent in the majority of cases. While occasional deviations remain, these analyses provide direct evidence of semantic preservation beyond token-level matching and embedding similarity.

We will incorporate these results and the corresponding evaluation protocol into the revised manuscript.

Q2.

Thank you for pointing out this ambiguity. We clarify that the MGSM training set contains 10 languages (10,300 instances) constructed from multilingual GSM8K, whereas evaluation covers 11 languages with 250 instances each (2,750 in total). Specifically, Telugu is excluded from training and included only in evaluation, serving as an unseen-language test of cross-lingual generalization.

We also identified a reporting error in Appendix Table 4, where the MGSM test total was incorrectly listed as 2,500 instead of 2,750. We will correct this value and explicitly distinguish the training and evaluation languages in the revised manuscript.

We sincerely thank the reviewer again for the constructive feedback and hope that our clarifications and additional experiments adequately address your concerns and encourage you to reconsider your assessment.

Dear Reviewer qasp,

We sincerely thank you again for your thoughtful comments and constructive feedback. We have carefully addressed your concerns in our rebuttal and hope that our responses have helped clarify the questions you raised.

As the discussion period is approaching its deadline, we would be truly grateful if you could take a moment to review our responses and share any further thoughts. If you have any remaining questions or concerns, we would be glad to provide additional clarifications or evidence.

If our responses and additional experiments have adequately addressed your concerns, we would sincerely appreciate your reconsideration of the overall rating.

We greatly appreciate your time and effort in reviewing our work and look forward to hearing your feedback.

Best regards,

The Authors

Reviewer_6ZCCOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes mRewriter-R1, a reinforcement learning framework for multilingual query rewriting. A policy model selects predefined operators and generates rewritten queries in a fixed sequence covering language, structure, and semantics. The policy is trained with GRPO using the answer correctness of a frozen downstream model as the reward. Experiments on MGSM, XCOPA, and BELEBELE report improvements over multilingual rewriting baselines, with additional evaluations of transfer across models, tasks, and languages.

Critical strengths
  1. A clear and modular approach. The framework organizes multilingual rewriting into interpretable operations while keeping the downstream model frozen. This offers a potentially useful way to improve models whose parameters cannot be updated.

  2. Promising improvements and transferability. Table 1 reports gains over the strongest baseline of approximately 4.2, 1.7, and 2.6 percentage points on MGSM, XCOPA, and BELEBELE, respectively. Evaluations across downstream models, tasks, and unseen languages suggest practical value as a reusable preprocessing module.

Critical weaknesses
  1. The benefit over direct answerer training is unclear. In the main setting, both roles use Qwen2.5-7B, but there is no comparison against directly RL-training the answerer with the same data and comparable compute.

  2. Inference budgets and costs are unclear. CoT comparisons lack explicit budget matching, and Table 3 does not specify whether all rewriting calls are included. A MGSMv2 study https://arxiv.org/html/2606.21954v1 reports large thinking-mode effects; its revised data and grading preclude direct score comparisons here.

  3. Recent backbones are not evaluated. In Qwen3's 8-shot CoT evaluation, Qwen3-8B-Base scores 76.02% on MGSM versus 63.60% for Qwen2.5-7B-Base. This motivates testing newer, similarly sized models; these Base-model scores cannot be directly compared with this paper's Instruct/RL results.

  4. Adaptive selection is not isolated. Existing ablations support structured rewriting, but fixed-operator and non-RL three-stage controls are missing. Figure 5 also confounds query populations and rewriting depth.

  5. Operator design and coverage need justification. The inventory's rationale and sensitivity to alternatives need validation, as does adherence to the selected operators. Cross-task transfer offers some support, but coverage beyond the tested tasks remains unclear.

  6. Results and statistical tests need clarification. Figure 6 implies MGSM accuracies of 75.78%/64.29% for ours/CSICL, versus 73.9%/69.7% in Table 1. A single run versus three-run averages may explain this. Significance-test comparators and repeated-run handling also need clarification.

  7. Reproduction details are underspecified. Key GRPO hyperparameters, decoding/thinking budgets, evaluation framework/version, task templates, and answer-extraction rules need clarification.

Questions for authors
  1. How does the method compare against directly RL-trained answering with comparable training data and compute, and against direct reasoning on a recent, similarly sized multilingual backbone with explicitly controlled thinking settings and a matched end-to-end inference budget, including all rewriting calls?
  2. How do the same three-stage pipeline without RL and a validation-selected fixed operator combination perform? What motivates the operator inventory, and how sensitive are the results to a modest change in it?
  3. Does Figure 6 represent a single run rather than the three-run averages in Table 1? Please reconcile the results, report variability across runs, and clarify the significance tests' comparison targets and handling of repeated runs.
  4. Please provide the training hyperparameters, evaluation framework/version, task templates, decoding and thinking settings, and answer-extraction rules, and clarify how Table 3 accounts for all rewriting and answering calls.
8 comments

We sincerely thank the reviewer for the careful and constructive assessment of our work. We carefully address each concern below and provide additional experiments and clarifications to strengthen the evaluation of our approach.

W1: The benefit over direct answerer training is unclear. In the main setting, both roles use Qwen2.5-7B, but there is no comparison against directly RL-training the answerer with the same data and comparable compute.

Thank you for this important suggestion. Our primary goal is to develop a plug-and-play query rewriting policy that improves multiple frozen downstream reasoning models without modifying their parameters. Nevertheless, we recognize the importance of comparison with direct answerer training. We therefore additionally evaluate direct GRPO-based reinforcement learning (RL) using the same training data, GRPO hyperparameters, and number of optimization steps as mRewriter-R1.

MethodEnBnDeEsFrJaRuSwThZhTeAvg
RL92.071.681.680.483.277.285.623.276.483.238.872.1
Ours92.874.483.284.884.478.886.824.879.284.040.073.9

As shown above, mRewriter-R1 outperforms direct answerer RL, suggesting that optimizing query rewriting is an effective alternative to directly optimizing the answerer. More importantly, rather than improving a specific downstream model through parameter updates, our approach learns a transferable rewriting policy that benefits multiple frozen reasoning models, as further demonstrated in Figure 3(a) and Table 9. We will incorporate these results into the revised manuscript.

W2:Inference budgets and costs are unclear. CoT comparisons lack explicit budget matching, and Table 3 does not specify whether all rewriting calls are included. A MGSMv2 study https://arxiv.org/html/2606.21954v1 reports large thinking-mode effects; its revised data and grading preclude direct score comparisons here.

Thank you for raising this important concern. We clarify the inference budgets and computational costs from two perspectives. (1)CoT inference budget. We agree that identical decoding settings do not necessarily imply equal end-to-end costs. We therefore additionally report the actual end-to-end token consumption for the CoT comparisons, including all rewriting and answering calls.

MethodInput TokensOutput TokensTotal TokensMGSM Acc. (%)
Direct + CoT179.17531.04710.2169.0
mRewriter-R1 + CoT263.16456.24719.4074.7

(2)End-to-end inference cost. The original Table 3 did not sufficiently clarify the accounting of rewriting overhead. To ensure transparent and consistent cost comparisons, we re-evaluate the token consumption by accounting for all rewriting and answering calls. Under this updated accounting, mRewriter-R1 consumes 252.16 input tokens and 394.91 output tokens per sample, totaling 647.07 tokens.

MethodInput TokensOutput TokensTotal TokensMGSM Acc. (%)
Direct172.17507.25679.4265.2
mRewriter-R1252.16394.91647.0773.9

We found that mRewriter-R1 achieves an 8.7-point accuracy improvement while reducing total token consumption by 4.76% compared with Direct, and mRewriter-R1 + CoT achieves a 5.7-point accuracy improvement over Direct + CoT despite comparable end-to-end token consumption, demonstrating a favorable accuracy–efficiency trade-off.

Regarding thinking configurations, we agree that MGSMv2 uses revised data and grading procedures, precluding direct score comparisons. To further address the reviewer's concern about thinking-mode effects, we additionally evaluate Qwen3-8B under a consistent thinking configuration across all methods, as detailed in our response to W3. We will clarify these settings and the complete inference budgets in the revised manuscript.

W3: Recent backbones are not evaluated. In Qwen3's 8-shot CoT evaluation, Qwen3-8B-Base scores 76.02% on MGSM versus 63.60% for Qwen2.5-7B-Base. This motivates testing newer, similarly sized models; these Base-model scores cannot be directly compared with this paper's Instruct/RL results.

Thank you for this suggestion. We additionally evaluate Qwen3-8B as a frozen downstream answerer, while keeping our original Qwen2.5-7B-based rewriter unchanged, without any additional training. All methods use the same Qwen3-8B checkpoint with thinking mode enabled and identical downstream decoding configurations.

MethodEnBnDeEsFrJaRuSwThZhTeAvg
Direct95.279.683.687.684.478.888.848.887.684.866.880.5
CoT96.884.085.688.083.680.088.847.689.685.271.281.9
CSICL93.283.287.288.884.481.686.854.888.084.877.282.7
Ours94.486.488.089.686.083.690.456.489.287.279.284.6

As shown above, mRewriter-R1 consistently achieves the best overall performance, outperforming Direct, CoT, and CSICL. These results demonstrate that our learned rewriting policy remains effective on a newer and stronger reasoning backbone without additional training, further validating its plug-and-play generalizability. We will incorporate these results into the revised manuscript.

W4:Adaptive selection is not isolated. Existing ablations support structured rewriting, but fixed-operator and non-RL three-stage controls are missing. Figure 5 also confounds query populations and rewriting depth.

We appreciate this insightful concern. We introduce two additional controlled baselines:

  • Non-RL Adaptive Pipeline: An untrained Qwen2.5-7B-Instruct model adaptively selects operators using the same three-stage workflow and operator space.
  • Fixed Operator Pipeline: A fixed combination of operators is selected on a held-out validation set and applied to all queries through the same three-stage workflow, without RL training.

These experiments help distinguish the benefit of the fixed multi-stage framework from the additional improvements associated with adaptive selection and RL optimization.

Method             En    Bn    De    Es    Fr    Ja    Ru    Sw    Th    Zh    Te    Avg
Non-RL Adaptive   86.8  72.0  82.4  83.2  82.4  71.2  86.0  18.4  80.4  84.0  35.6  71.12
Fixed Operator    88.0  68.4  76.4  79.6  76.0  69.6  81.6  20.8  74.4  82.4  32.8  68.18
Ours              92.8  74.4  83.2  84.8  84.4  78.8  86.8  24.8  79.2  84.0  40.0  73.90

As shown above, Non-RL Adaptive outperforms Fixed Operator by 2.94 percentage points, demonstrating the benefit of query-dependent operator selection over a fixed rewriting strategy. Moreover, mRewriter-R1 further improves over Non-RL Adaptive by 2.78 points, indicating that RL optimization provides additional gains beyond training-free adaptive rewriting.

Regarding Figure 5, its primary purpose is to illustrate the operator selection patterns learned by our policy. We acknowledge that the accuracy comparisons involve different query subsets and rewriting depths, and therefore should not be interpreted as a controlled evaluation of adaptive selection. To address this limitation, our newly added Fixed Operator and Non-RL Adaptive baselines provide more direct evidence under the same three-stage rewriting framework. We will clarify this distinction in the revised manuscript.

W5: Operator design and coverage need justification. The inventory's rationale and sensitivity to alternatives need validation, as does adherence to the selected operator. Cross-task transfer offers some support, but coverage beyond the tested tasks remains unclear.

Thank you for raising this important question. Our operator inventory is motivated by three complementary sources of multilingual reasoning difficulties:

  1. Language aspect: Mitigating cross-lingual expression discrepancies while allowing the preservation of the original language when appropriate.
  2. Structural aspect: Making implicit relationships and scattered conditions more explicit and reasoning-friendly.
  3. Semantic aspect: Resolving ambiguities, clarifying references, and making implicit reasoning requirements explicit.

The corresponding operators instantiate these rewriting objectives, while keep_original allows the model to avoid unnecessary transformations within each aspect.

As discussed in Section 2.2 and Appendix D, unrestricted rewriting introduces a substantially larger and less structured generation space, making outcome-based RL optimization more challenging. Figure 4 further shows unstable training behavior when the operator-guided mechanism is removed. Our structured operator inventory therefore provides a tractable decision space while retaining flexibility in the actual query rewriting process.

Nevertheless, we agree that these observations do not fully establish the robustness of the specific operator inventory. We therefore conduct additional sensitivity analyses by modifying the available operator set. Specifically, we keep the trained rewriting policy fixed and modify only the available operators during inference.

ConfigurationNumber of OperatorsMGSM Accuracy (%)Δ vs. Original (pp)
Original1374.22—
Reduced1073.85−0.37
Alternative1373.82−0.40

As shown above, performance remains largely stable under both reduced and alternative operator configurations, indicating that our method is not overly sensitive to the specific operator inventory. These results suggest that the learned policy remains robust to modest changes in the available operator inventory, rather than depending critically on every individual operator.

We further examine whether the generated rewrites follow the selected operators.

AspectOperator Adherence Rate (%)
Language96.2
Structural90.1
Semantic87.7
Overall91.3

As shown above, the generated rewrites exhibit consistently high adherence to the selected operators across all three aspects, with language-level operations showing the strongest adherence. These results demonstrate that the learned policy can reliably execute its selected rewriting strategies, further supporting the interpretability and reliability of our operator-guided framework.

Regarding generalization, the three rewriting aspects are designed to address common challenges in multilingual query understanding rather than task-specific characteristics. Our existing evaluations on unseen languages (Appendix C.1), cross-task transfer (Appendix C.2), and heterogeneous downstream backbones (Figure 3(a)) further demonstrate the transferability of the learned rewriting policy across different languages, tasks, and models. These results support the broader applicability of our operator design beyond a single task or backbone.

W6: Results and statistical tests need clarification. Figure 6 implies MGSM accuracies of 75.78%/64.29% for ours/CSICL, versus 73.9%/69.7% in Table 1. A single run versus three-run averages may explain this. Significance-test comparators and repeated-run handling also need clarification.

Thank you for carefully identifying this discrepancy. We revisited our experimental records and identified two causes: (1) For CSICL, Figure 6 used an earlier answer-normalization implementation that incorrectly classified some correct predictions, leading to underestimated accuracy. (2) For mRewriter-R1, Figure 6 was based on a single inference run, whereas Table 1 reports the average over three runs, explaining the observed difference.

To further assess the reliability of our results, we conducted three additional independent inference runs and reported the corresponding mean and standard deviation below.

DatasetMethodRun 1Run 2Run 3Mean ± Std.
MGSMCSICL70.169.369.769.70 ± 0.40
MGSMmRewriter-R173.575.072.973.80 ± 1.08
XCOPACSICL71.672.471.971.97 ± 0.40
XCOPAmRewriter-R178.277.476.877.47 ± 0.70
BELEBELECSICL78.076.977.877.57 ± 0.59
BELEBELEmRewriter-R179.880.780.380.27 ± 0.45

(W6 continues in the next comment due to the length limit.)

Reviewer_Hx4MOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces mRewriter-R1, an agentic framework that use RL to tackle multilingual query rewriting for reasoning tasks. Instead of relying on static, single-turn rewriting approaches like direct translation, the authors frame the problem as a multi-turn, sequential decision-making process. The framework dynamically selects aspect-specific operators tailored to query characteristics, aiming to optimise reasoning trajectories across various language models. We validate on standard multilingual reasoning benchmarks.

Critical strengths

The conceptual shift from single-turn, monolithic query translation to a sequential, multi-turn decision-making process via RL is a compelling perspective. Applying agentic operator selection to multilingual reasoning is a conceptually sound direction that acknowledges the semantic nuances required when aligning queries across different languages.

Critical weaknesses

The paper's primary weakness is its inadequate contextualisation within the current SOTA. The authors overlook a substantial, highly relevant body of recent literature on Multilingual RAG and advanced LLM-based query-rewriting paradigms.

Moreover, the baselines are artificially weak and do not reflect the true competitive landscape of modern query optimisation. By neglecting to benchmark against—or even thoroughly discuss—contemporary multilingual RAG rewriting techniques, the empirical validity of the proposed mRewriter-R1 is significantly undermined. While the framework claims superiority over direct translation, the absence of comparison with actual SOTA makes the real interpretation gains questionable.

Questions for authors

Could you map their approach against the recent literature on Multilingual RAG and explain why they omitted contemporary LLM-based query-rewriting baselines from the empirical evaluation?

Are you able to provide supplementary experimental results that evaluate mRewriter-R1 against more modern SOTA query rewriting techniques, rather than relying predominantly on direct translation and basic prompt baselines?

3 comments

We sincerely thank the reviewer for the careful assessment and constructive suggestions. Our original evaluation included recent and representative query reformulation methods for multilingual reasoning published in ACL26(selective), COLM26(CSICL) and *SEM26(DIP), rather than relying solely on direct translation or basic prompting.

Regarding the absence of contemporary multilingual RAG rewriting baselines, we prioritized these methods because recent multilingual RAG rewriting approaches are designed to optimize retrieval quality and heavily rely on retrieval feedback as their optimization signal, making them difficult to directly apply to retrieval-free multilingual reasoning.

Nevertheless, we recognize the importance of broader comparisons and therefore adapt three recent query rewriting methods to our setting, covering both multilingual RAG and advanced LLM-based query rewriting paradigms:

  • DELTA (ACL 2026 Findings): A multilingual RAG method employing language preference-guided query fusion. We adapt its query fusion mechanism to the retrieval-free reasoning setting.
  • CLIR-QE (2025): An LLM-based cross-lingual query expansion method leveraging multilingual LLMs. We apply its query expansion strategy to multilingual reasoning without retrieval.
  • ReQueR (ACL 2026): An RL-based query refinement framework for downstream reasoning. We extend its evaluation to multilingual reasoning tasks.

We evaluate these methods under the same MGSM benchmark and frozen Qwen2.5-7B-Instruct answerer, adapting their query transformation components to the retrieval-free setting.

MethodEnBnDeEsFrJaRuSwThZhTeAvg
DELTA69.230.046.851.649.241.244.810.845.655.623.642.6
ReQueR85.666.081.284.880.474.480.424.477.278.026.068.9
clir-qe92.470.881.286.480.472.482.818.879.282.034.070.9
ours92.874.483.284.884.478.886.824.879.284.040.073.9

As shown above, mRewriter-R1 outperforms all three additional baselines. These results demonstrate the effectiveness of our approach beyond conventional translation-based methods and more recent query optimization paradigms. Unlike retrieval-oriented query expansion or general query refinement, our method explicitly learns adaptive, multi-aspect rewriting policies optimized for multilingual reasoning.

Thank you for this valuable comment again. We will expand the related work discussion and incorporate these additional comparisons into the revised manuscript. We hope these clarifications and supplementary results address your concerns and encourage you to reconsider your assessment.

Dear Reviewer Hx4M,

We sincerely thank you again for your valuable comments and constructive feedback. Following our previous response, we have further extended our evaluation of recent query rewriting methods to XCOPA and BELEBELE, providing additional evidence of the effectiveness and generalizability of mRewriter-R1 across different multilingual reasoning tasks.

Specifically, we further evaluate CLIR-QE and ReQueR on both benchmarks under the same frozen Qwen2.5-7B-Instruct answerer. Given DELTA's substantially lower performance in our previous MGSM experiments, we prioritized evaluating CLIR-QE and ReQueR on these additional benchmarks first.

(1) Results on XCOPA

MethodEtHtIdItQuSwTaThTrViZhAvg.
Translate68.465.893.292.850.053.663.882.685.889.290.275.9
CSICL60.660.085.690.053.256.856.877.879.883.287.872.0
CLIR-QE68.461.092.892.451.052.859.884.484.290.492.075.4
ReQueR63.259.289.291.650.451.458.281.882.486.290.673.1
mRewriter-R170.666.693.893.854.656.864.284.686.690.691.477.6

(2) Results on BELEBELE

MethodArZhEnFrDeHiJaRuSwThAvg.
Translate79.287.690.088.084.052.880.489.636.066.075.4
CSICL80.088.490.888.084.862.482.484.438.875.677.6
CLIR-QE84.087.690.487.286.065.682.484.834.475.277.8
ReQueR83.287.690.083.684.862.078.484.034.068.875.6
mRewriter-R186.490.091.290.087.865.684.887.440.278.280.2

As shown above, mRewriter-R1 consistently achieves the best overall performance on both benchmarks. These additional results demonstrate that our approach remains effective against recent query rewriting methods across various reasoning tasks.

We hope these additional experiments further address your concerns. As the discussion period is approaching its deadline, we would be truly grateful if you could take a moment to review our responses and share any further thoughts. If you have any remaining questions or concerns, we would be glad to provide additional clarifications or evidence.

If our responses and additional experiments have adequately addressed your concerns, we would sincerely appreciate your reconsideration of the overall rating.

We greatly appreciate your time and effort in reviewing our work and look forward to hearing your feedback.

Best regards,
The Authors

Dear Reviewer Hx4M,

We understand that you may have a busy schedule, but we would sincerely appreciate it if you could take a moment to share your thoughts on our rebuttal.

Your feedback is extremely valuable to us, and we would welcome any further questions or suggestions you may have. We remain fully committed to addressing any outstanding concerns.

Thank you again for your time and 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.