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
Existing multilingual retrieval-augmented generation relies heavily on learned retrievers and similarity-based ranking from multilingual collections, which are vulnerable to cross-lingual retrieval errors and indexing overhead. In this paper, we explore retriever-free multilingual search, where the model directly interacts with raw corpora through executable shell commands. This paradigm shifts the bottleneck to multilingual search action generation, where exact search is sensitive to entity surface forms and outcome-only supervision provides limited guidance for multi-turn search. To address these challenges, we propose MARS, a multilingual search agent in direct corpus interaction (DCI) with reinforcement learning (RL). MARS introduces corpus-native entity grounding to recover reliable lexical anchors from corpus evidence, mitigating failure to retrieve evidence. Meanwhile, we present a redundancy-aware search reward to encourage concise and effective search trajectories, addressing training collapse in RL of multilingual scenarios. Experiments show that \ours~consistently outperforms strong retrieval-based and DCI baselines, while demonstrating robust cross-lingual generalization and substantially reducing redundant corpus interactions.
Figures & tables
Figure 1: Illustration of lexical mismatch and trajectory optimization challenges in multilingual DCI. Challenge 2 compares multi-corpus and single-corpus search, where the non-redundant ratio denotes the fraction of retrieved documents that are unique within a trajectory for each language.
Figure 2: Illustration of Mars for multilingual DCI. The policy directly interacts with language-specific corpora in URPO, while corpus-native grounding supplies reusable lexical anchors and redundancy-aware search encourages successful, non-redundant search trajectories.
Dataset
Split
# / Lang.
Lang.
MKQA
Train
1000
En, Ar, Fi, Fr, Ja, It, Ru, Zh
Test
1277
Train Lang. + De, Ko, Es, Pt, Th (Unseen)
MLQA
Train
1000
En, De, Es, Hi, Vi, Zh, Ar
Test
1000
Train Lang.
Table 1: Language coverage for seen and unseen categorization.
Methods
MKQA-Seen
MKQA-Unseen
MLQA
EM
fEM
EM
fEM
EM
fEM
Qwen3.5-4B
Direct
13.9
17.3
13.7
17.0
5.3
6.1
RAG
24.0
30.8
21.9
27.6
10.4
12.7
D-RAG
27.1
33.3
23.3
29.1
11.6
14.3
LcRL
24.0
28.9
21.0
27.1
7.6
10.1
Table 2: Main results (%) of on MKQA and MLQA. We report the average scores of categorized languages, including seen languages of MKQA and MLQA, and unseen languages of MKQA. The highest score under each base model is highlighted in bold , while † denote statistically significant improvement, compared to the best baseline (p < 0.05). Detail results of each language are provided in Appendix B .
Rgro
Rred
EM
fEM
✗
✗
37.07
53.23
✓
✗
39.79 +2.72
53.97 +0.74
✗
✓
38.97 +1.90
54.50 +1.27
✓
✓
40.64 +3.57
55.54 +2.31
Table 3: Ablation results on MKQA. Subscripts denote absolute improvements over the multilingual DCI baseline.
Search Key
Empty Hit ↓
First Hit ↓
Correct Hit ↑
EM ↑
In-Trans
32.2
1.25
54.1
37.1
Ex-Trans
39.0
1.33
54.9
38.3
Ours
24.3
1.30
56.7
40.3
Table 4: Effects of search-key sources by different strategies on multilingual evidence access. “First Hit” denotes the number of rounds required for the first hit.
Metric
Ours
w/o. Rred
Repeated Search Rate ↓
4.89%
22.33%
# Search Commands ↓
2.56
5.51
# Interaction Turns ↓
1.36
2.27
Table 5: Effects of redundancy-aware search on multilingual DCI trajectories. We report the average of all queries. # denotes the count. Lower values indicate more concise and less redundant corpus interaction.
Methods
Offline Index
↑ EM (%)
↓ Search Commands (s)
↓ Search Turns (#)
↓ Search Duplicate Rate (%)
↓ Response Tokens (Latency Proxy)
D-RAG
High (mE5)
–
–
–
–
6,920
LcRL
High (mE5)
27.23
–
2.07
6.2
3,532
Grepseek
None
37.07
4.22
1.87
16.4
4,853
Mars
None
40.64
2.56
1.36
4.9
3,447
( Δ ) Mars vs. Grepseek
–
+3.2 pt
↓ 39%
↓ 27%
↓ 70%
↓ 29%
Table 6: Efficiency comparison between dense multilingual RAG and direct corpus interaction. Response tokens are reported as a proxy for generation latency.
Figure 3: Overall EM under increasing corpus-language coverage on MKQA. Ul denotes the corpus in the user-query language, Ul+En additionally includes English, and All includes all available language-specific corpora.
HotpotQA
Search-R1
GrepSeek
Mars
F1
54.24
62.31
63.51
EM
32.02
46.27
50.00
Δ (F1-EM)
22.22
16.04
13.51
Table 7: Token-level F1 and EM in the original monolingual setting of GrepSeek.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Methods
MKQA (Seen Languages)
MKQA (Unseen Languages)
En
Fr
It
Ru
Fi
Zh
Ja
Ar
Avg.
Es
Pt
Ko
Th
De
Avg.
Qwen3.5-4B
Direct
24.3
18.6
17.2
12.7
12.3
11.9
6.8
7.1
13.9
18.6
17.4
7.1
6.3
19.2
13.7
RAG
41.7
31.6
30.5
23.8
22.2
11.7
17.8
12.9
24.0
30.1
29.6
10.4
11.8
27.8
21.9
D-RAG
43.9
33.7
28.5
24.2
29.4
19.3
20.6
17.5
27.1
35.1
14.8
13.5
10.7
42.2
23.3
LcRL
36.1
30.6
28.9
21.9
27.1
16.2
19.0
12.0
24.0
27.7
29.1
12.1
9.2
26.8
21.0
Appendix
Table 8: Main results of EM (%) on MKQA. Background colors distinguish seen languages , unseen languages , and average performance . The highest score under each base model is highlighted in bold , and the second-highest score is underlined .
Methods
MKQA (Seen Languages)
MKQA (Unseen Languages)
En
Fr
It
Ru
Fi
Zh
Ja
Ar
Avg.
Es
Pt
Ko
Th
De
Avg.
Qwen3.5-4B
Direct
29.6
22.7
20.4
15.0
14.3
17.4
10.5
8.4
17.3
23.0
21.5
9.2
7.8
23.4
17.0
RAG
53.6
39.6
36.8
29.4
28.8
16.6
24.6
16.8
30.8
37.7
37.7
13.7
14.3
34.8
27.6
D-RAG
49.7
39.5
34.8
30.3
35.7
25.5
26.9
23.7
33.3
41.2
20.2
19.7
15.9
48.4
29.1
LcRL
44.2
36.4
33.9
25.1
31.8
22.0
22.9
14.9
28.9
39.4
37.4
14.0
13.0
31.6
27.1
Appendix
Table 9: Main results of fEM (%) on MKQA. Background colors distinguish seen languages , unseen languages , and average performance . The highest score under each base model is highlighted in bold , and the second-highest score is underlined .
Methods
EM
fEM
En
De
Es
Hi
Vi
Zh
Ar
Avg.
En
De
Es
Hi
Vi
Zh
Ar
Avg.
Qwen3.5-4B
Direct
9.4
6.6
5.5
2.1
6.6
4.1
2.5
5.3
11.0
7.7
6.1
2.3
7.8
5.0
3.1
6.1
RAG
14.7
11.6
9.3
10.1
14.4
6.3
6.6
10.4
18.3
14.0
11.2
11.8
16.1
8.7
8.6
12.7
D-RAG
17.2
13.2
9.8
10.7
15.1
7.6
7.4
11.6
19.8
15.8
12.6
13.5
17.8
10.2
10.2
14.3
LcRL
12.4
8.7
8.2
5.5
7.6
7.2
3.3
7.6
14.7
11.4
13.9
6.4
10.0
9.6
4.5
10.1
Appendix
Table 10: Main results (%) on MLQA measured by EM and fEM. The highest score under each base model is highlighted in bold , and the second-highest score is underlined .
Methods
Corpora
En
Fr
It
Ru
Fi
Zh
Ja
Ar
Avg.
Grepseek
Ul
40.6
37.5
37.1
31.3
33.1
27.1
29.4
22.4
32.3
Ul+En
53.6
48.9
45.2
37.4
39.8
28.8
36.9
22.9
39.2
All
51.7
44.6
44.7
36.6
36.6
24.8
35.1
22.5
37.1
Mars
Ul
51.9
46.3
42.4
34.8
36.7
25.5
37.4
21.4
37.1
Ul+En
53.0
46.2
46.0
37.9
42.0
31.6
37.7
23.9
39.8
All
53.6
49.4
46.8
38.7
42.5
31.2
38.1
24.8
40.6
Appendix
Table 11: Results of EM (%) on MKQA for corpus-language coverage, based on Qwen3.5-9B . Ul denotes the corpus in the user-query language, Ul+En additionally includes English, and All includes all available language-specific corpora. The highest score under each base model is highlighted in bold , and the second-highest score is underlined .
Table 12: System prompt for Mars . Multilingual-specific instructions are highlighted in blue.
Chinese Information Processing Laboratory, Institute of Software, Chinese Academy of Sciences · University of Chinese Academy of Sciences · MYbank, AntGroup