Organizations: Key Laboratory of Big Data & Artificial Intelligence in Transportation (Beijing Jiaotong University), Ministry of Education · School of Computer Science and Technology, Beijing Jiaotong University · Pattern Recognition Center, WeChat AI, Tencent Inc, China · University of Edinburgh
In multilingual scenarios, queries with equivalent semantics but in different languages could guide the model into different reasoning trajectories, leading to performance disparities. To mitigate this gap, previous studies typically apply a one-size-fits-all query rewriting strategy, such as translation, which overlooks the fact that different scenarios require 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 mRewriter-R1 outperforms all strong multilingual rewriting baselines on different large reasoning backbones. Further analyses 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.
Figures & tables
Figure 1: Illustration of the impact of query variations for multilingual reasoning.
Figure 2: The workflow of mRewriter-R1, where the policy LLM is trainable while the reference LLM is frozen. The action follows a multi-turn iteration in a sequential decision-making process with multi-aspect operator space.
MGSM
Method
En
Bn
De
Es
Fr
Ja
Ru
Sw
Th
Zh
Te
Avg.
Direct
87.2
60.0
73.2
77.2
74.8
64.8
72.0
18.0
76.0
76.0
37.6
65.2
Translate
86.0
63.2
78.4
84.0
77.2
75.2
84.4
20.8
74.0
79.2
33.2
68.7
Selective
88.0
59.6
74.8
80.8
74.8
67.2
72.0
20.8
76.4
76.8
37.2
66.2
DIP
85.2
65.4
72.8
78.8
74.6
70.2
78.6
24.0
74.2
76.6
34.2
66.8
CSICL
91.2
62.0
78.4
83.6
76.8
70.4
83.2
20.8
77.6
82.4
40.4
69.7
Table 1: Accuracy (%) on three benchmarks across different languages based on Qwen2.5-7B-Instruct. † denotes significant improvements with McNemar test at p<0.05 between our methods with the second best method. The best results in each column within the same benchmark are bold , and the second-best is underlined .
Figure 3: Performance of plug-and-play and CoT generalization across backbones. Our rewriter is trained based on Qwen2.5-7B-Instruct . Detailed results are in Appendix E .
Table 5
Figure 5: Operator distribution and reasoning performance across operators in each aspect. The donut charts (top) show the usage frequency of rewriting operators selected by the policy LLM. The lower plot compares accuracy between mRewriter-R1 (red line), where operators are adaptively selected conditioned on the query, and standalone one-step rewriting of individual operators (blue line).
Figure 6: Correctness transitions from Direct to rewriting methods on 2,750 MGSM test instances.
Metric
Direct
Translate
Selective
DIP
CSICL
mRewriter-R1
Input Tokens
172.17
189.17
174.17
213.80
174.17
193.42
Output Tokens
507.25
512.00
512.00
439.13
241.42
217.76
Total Tokens
679.42
701.17
686.17
652.93
415.59
411.18
Time (s)
0.7194
0.7291
0.7281
0.8106
0.7233
0.5390
Table 3: Comparison of token consumption and inference latency per sample across different methods. The values are the average over all samples.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Dataset
Training Set
Test Set
# Lang
# / Lang
Total
# / Lang
Total
MGSM
En, Bn, De, Es, Fr, Ja, Ru, Sw, Th, Zh (10)
V*
10,300
250
2,500
BELEBELE
Ar, Zh, En, Fr, De, Hi, Ja, Ru, Sw, Th (10)
650
6,500
250
2,500
XCOPA
Et, Ht, Id, It, Qu, Sw, Ta, Th, Tr, Vi, Zh (11)
100
1,100
500
5,500
Appendix
Table 4: Number of training and evaluation instances across datasets and languages used in each dataset.
Language
Word Order
Writing System
Resource Level
English
SVO
Latin
High
Bengali
SOV
Non-Latin
Medium
German
Flexible
Latin
High
Spanish
SVO
Latin
High
French
SVO
Latin
High
Japanese
SOV
Non-Latin
High
Appendix
Table 5: Linguistic grouping of the MGSM languages used in the typology-aware analysis.
Word Order
Writing System
Resource Level
SVO
SOV
Flexible
Latin
Non-Latin
High
Medium
Low
Direct
68.20
54.13
72.60
66.08
64.40
75.03
68.00
27.80
mRewriter-R1
75.00
64.40
85.00
74.00
73.87
84.97
76.80
32.40
Δ
+6.80
+10.27
+12.40
+7.92
+9.47
+9.94
+8.80
+4.60
Appendix
Table 6: Performance of mRewriter-R1 across linguistic groups on MGSM. Languages are grouped by word order, writing system, and resource level. Δ denotes the absolute accuracy improvement over Direct.
Table 12
Method
En
Bn
De
Es
Fr
Ja
Ru
Sw
Th
Zh
Te
Avg.
DeepSeek-R1-Distill-Qwen-7B
Direct
34.0
10.8
50.8
58.4
52.0
49.6
56.0
3.2
15.6
67.6
6.0
36.7
Translate
87.6
58.8
74.8
75.6
72.0
66.0
74.4
4.4
63.2
76.4
32.0
62.3
Selective
49.6
22.0
55.6
63.6
57.6
52.8
58.8
4.4
25.6
67.2
11.6
42.6
DIP
66.0
27.8
50.2
57.2
49.8
52.4
55.6
6.2
35.8
66.6
20.8
44.4
CSICL
67.2
38.4
52.0
58.0
50.8
52.8
57.2
6.4
36.0
68.0
21.6
46.2
Appendix
Table 9: Accuracy (%) on MGSM using various backbones. † denotes significant improvements with McNemar test at p<0.05 between our methods with the second best method. The best result in each column is shown in bold , and the second-best is underlined .
Method
En
Bn
De
Es
Fr
Ja
Ru
Sw
Th
Zh
Te
Avg.
Direct
86.4
66.0
77.2
80.0
76.8
74.8
80.8
22.4
76.4
81.2
37.2
69.0
Translate
84.0
68.0
77.2
80.4
82.0
76.4
84.8
24.4
82.4
78.0
36.8
70.4
Selective
86.4
65.6
78.0
81.2
78.0
74.8
82.0
24.0
76.8
81.6
38.8
69.8
DIP
87.8
65.2
76.0
80.2
77.8
73.4
80.8
24.6
77.4
79.6
37.6
69.1
CSICL
89.6
66.0
80.0
84.0
80.4
76.4
84.4
21.2
76.0
81.2
38.0
70.7
mRewriter-R1
92.8 †
74.0 †
86.8 †
86.4 †
83.2 †
77.2 †
86.4 †
26.8 †
83.6 †
83.6 †
40.4 †
74.7 †
Appendix
Table 10: Accuracy (%) on MGSM using Qwen2.5-7B-Instruct under the CoT prompting setting. † denotes significant improvements with McNemar test at p<0.05 between our methods with the second best method. The best result in each column is shown in bold , and the second-best is underlined .
Metric
Setting
Value
Critical-number token recall
How many of the key numbers from
93.28%
the original problem can be found after rewriting it?
Queries with no numeric loss
How many samples showed no digital loss?
89.07%
Currency marker preservation
If the original problem contains currency information,
97.42%
is the currency constraint retained after rewriting?
Percentage marker preservation
When the original problem involves percentages,
99.71%
Appendix
Table 11: Semantic-preservation analysis on the MGSM test set.