Multilingual dense retrievers place all languages in one shared space so that a question finds relevant documents, regardless of which language they are written in. We show that the same question, asked in different languages, retrieves substantially different documents, and this disagreement is not because of weak language understanding. Each language performs well on its own and ranks candidates consistently with the other languages; divergence appears only when languages share an index. We trace the mechanism: each language keeps its own ranking, but in the shared index, its correct answers are pushed far down, buried under non-relevant documents that are either in the query's own language or in English, depending on the model. We introduce disagreement-guided steering: for each language, we contrast the documents that other languages agree on for a question against those that language retrieves idiosyncratically, and take the difference of their representations as a language-specific direction, added to that language's queries at inference. Against baselines of language-mean removal, subspace removal, and per-language rotation, it most strongly reduces same-language crowding while remaining competitive on recall, trading off against mean removal, which better preserves ranking quality.
Figures & tables
Belebele
XQuAD
Model
nDCG all
Rec all
Agr
nDCG all
Rec all
Agr
mE5-large
95.0
99.0
53.6
97.2
99.9
66.3
mE5-base
93.5
98.6
50.7
96.1
99.8
63.1
bge-m3
91.1
97.6
53.8
90.3
98.0
62.5
Qwen3-0.6B
87.7
95.5
50.8
88.4
97.6
59.2
Qwen3-8B
92.7
98.1
52.5
90.8
99.0
59.9
Table 1: Isolation: monolingual nDCG@20, Recall@20, and among-language agreement (Agr) when each language retrieves only within its own documents.
Belebele
XQuAD
Model
nDCG all
Rec all
Gold
Ord
Agr
nDCG all
Rec all
Gold
Ord
Agr
bge-m3
87.5
96.0
55
61.0
51.9
84.9
96.5
25
67.2
60.9
mE5-base
69.0
84.3
144
62.1
30.4
72.6
91.2
42
67.8
39.9
mE5-large
73.7
88.9
96
70.2
36.4
76.7
94.0
31
75.6
45.4
Qwen3-0.6B
80.7
91.3
108
55.7
41.5
79.6
94.5
34
61.5
53.7
Qwen3-8B
88.7
96.0
50
60.4
36.8
85.0
97.1
20
67.0
49.8
Table 2: Mixing: mixed-index nDCG@20, Recall@20, foreign-gold rank (Gold), preservation of each language’s monolingual order (Ord), and among-language agreement inside the shared index (Agr).
Belebele
XQuAD
Model
Same
En
Other
Hub
Drop
Same
En
Other
Hub
Drop
mE5-large
87.7
0.1
3.8
8.5
23/24
89.6
0.2
5.1
8.7
22/24
mE5-base
77.7
0.5
9.7
7.3
11/12
79.5
1.1
11.2
7.1
11/12
e5-mistral
91.6
0.0
5.6
3.4
29/32
97.3
0.0
1.6
2.6
29/32
bge-m3
8.7
6.6
72.8
8.5
22/24
13.9
11.7
67.7
8.1
23/24
Qwen3-8B
38.4
9.7
33.3
7.5
34/36
43.5
12.4
31.3
5.8
30/36
Table 3: Composition of the non-relevant documents that replace a demoted cross-lingual gold, as a percentage of all entrants: same-language (Same), English (En), and other foreign languages (Other), with the hub share (Hub) and the average drop layer (Drop).
Belebele
XQuAD
Multi-EuP
Method
Rec xl
nDCG xl
Own
En
Rec xl
nDCG xl
Own
En
Rec xl
nDCG xl
Own
En
mE5-large
baseline
85.0
45.7
69.3
8.5
87.7
47.0
71.3
14.0
76.8
23.0
45.2
35.3
random
85.0 ± 0.2
45.7 ± 0.2
69.3 ± 0.2
8.5 ± 0.1
87.7 ± 0.2
47.1 ± 0.1
71.3 ± 0.2
14.1 ± 0.1
76.7 ± 0.2
22.9 ± 0.1
45.4 ± 0.3
35.0 ± 0.3
permuted
33.2 ± 0.7
6.5 ± 0.2
96.1 ± 0.1
10.0
17.4 ± 0.3
2.8 ± 0.1
98.6 ± 0.0
16.4 ± 0.0
9.5 ± 0.1
1.3 ± 0.0
98.2 ± 0.0
5.4 ± 0.0
subspace
93.5 ± 0.7
63.4 ± 0.8
44.9 ± 1.9
8.0 ± 0.4
96.1 ± 0.3
59.4 ± 0.6
55.4 ± 1.0
11.8 ± 0.2
79.8 ± 1.1
27.2 ± 1.3
37.6 ± 4.0
29.1 ± 2.8
rotate
83.5 ± 2.2
45.3 ± 3.7
22.2 ± 7.6
19.4 ± 8.9
97.9 ± 0.3
68.5 ± 1.9
21.9 ± 3.8
34.6 ± 6.0
68.0 ± 2.3
15.4 ± 1.4
8.4 ± 1.3
70.4 ± 9.8
Table 4: Interventions: cross-lingual Recall@20, nDCG@20, own-language share (Own) and English share (En), averaged over query languages and five estimation seeds ( ± std; omitted when 0.0). Belebele directions are estimated on TED2020, XQuAD and Multi-EuP on FLORES-200. random and permuted are controls. The best per model and dataset is bold (highest Rec/nDCG, lowest Own/En). The final block averages across models.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
Model
Source
Params
Architecture
Multiling.
License
mE5-base
intfloat/multilingual-e5-base
278M
encoder
yes
MIT
mE5-large
intfloat/multilingual-e5-large
560M
encoder
yes
MIT
bge-m3
BAAI/bge-m3
568M
encoder
yes
MIT
Qwen3-0.6B
Qwen/Qwen3-Embedding-0.6B
0.6B
decoder
yes
Apache-2.0
Qwen3-8B
Qwen/Qwen3-Embedding-8B
8B
decoder
yes
Apache-2.0
e5-mistral
intfloat/e5-mistral-7b-instruct
7B
decoder
(en)
MIT
Appendix
Table 5: Embedding models evaluated, all obtained from Hugging Face and used in inference only. “Multiling.” indicates whether the model is explicitly trained and released for multilingual use; e5-mistral is primarily English-trained but applied multilingually. All listed licenses permit research use.
Query (it): Quale dei seguenti aspetti continuerà nonostante la presenza di stimoli esterni?
Rank
Lang
Gold
Document snippet
Baseline
1
it
✓
In molti hanno osservato che i ritmi fisiologici e il comportamento dipendono in maniera…
2
it
La stazione non ha cambiato approccio, nonostante la precedente perdita di un giroscopio…
3
it
L’avidità e l’egoismo faranno sempre parte di noi. È insito nella natura della cooperazione…
4
it
Internet riunisce in sé alcuni elementi sia della comunicazione di massa…
Appendix
Table 6: Top-5 retrievals for an Italian query under mE5-large. The baseline returns only Italian passages and no cross-lingual gold; steering keeps the Italian answer and surfaces the Russian and Hindi copies.
Query (pt): …qual dos seguintes é considerado uma solução de longo prazo viável para questões ambientais?
Rank
Lang
Gold
Document snippet
Baseline
1
pt
✓
Podemos começar a viver de maneira mais amigável com o ambiente, podemos nos juntar ao movimento…
2
pt
Outros assuntos na agenda em Bali incluem salvar as florestas remanescentes do mundo…
3
pt
Na savana, é difícil para um primata com um sistema digestivo como o dos humanos…
4
pt
Peter Costello, um tesoureiro da Austrália e o homem com mais chance de suceder…
Appendix
Table 7: Top-5 retrievals for a Portuguese query under mE5-large. The baseline returns only Portuguese passages; steering retains the Portuguese answer and surfaces the Spanish parallel copy at rank 5.
Belebele
XQuAD
Model
R@20
R@100
R@20
R@100
mE5-base
56
77
71
90
mE5-large
63
82
78
93
Qwen3-0.6B
75
88
82
94
Qwen3-8B
84
94
88
97
Appendix
Table 8: Cross-lingual Recall at K=20 vs. K=100 on the parallel collections. Demoted golds reappear at the larger cutoff, confirming demotion rather than loss.
Model
RRF=20
RRF=60
RRF=100
no wℓ
mE5-base
58.80
58.80
58.80
58.80
mE5-large
66.79
66.79
66.79
66.79
Qwen3-0.6B
69.22
69.22
69.21
69.22
Qwen3-8B
83.66
83.67
83.66
83.67
Appendix
Table 9: Steering nDCG@20 on Belebele under three rank-fusion constants and with the agreement weight removed. Values move by at most 0.01 across all conditions; cross-lingual recall is identical to within 0.01 . The main results use RRF=60.
Metric
Steering
Mean rem.
p
Own-lang share ( ↓ )
11.3
21.0
<10−19
Recall
87.4
84.7
0.28
nDCG (Belebele)
65.8
67.1
0.004
nDCG (XQuAD)
67.2
64.3
0.007
Appendix
Table 10: Paired Wilcoxon signed-rank tests, steering vs. mean removal, over model × dataset × language cells ( n=184 ). Steering’s reduction in own-language crowding is highly significant; recall is not significant; nDCG is significant in opposite directions by collection.
Figure 1: Per-language Recall@20 underbaseline, mean removal (centering), and disagreement-guided steering (caa), for mE5-large and Qwen3-8B on Belebele (all ten languages) and a subset of Multi-EuP. The Multi-EuP languages shown are a resource-balanced subset chosen to include cases where each correction wins, not to favor ours.
Steering
Residual
Model
Rec
nDCG
Rec
nDCG
cos
mE5-large
97.2
66.8
92.1
50.1
0.64
mE5-base
94.3
58.8
89.6
45.3
0.62
bge-m3
94.5
74.8
94.9
72.7
0.73
Qwen3-0.6B
92.3
69.2
91.2
63.4
0.65
Qwen3-8B
99.0
83.7
98.4
76.4
0.66
Appendix
Table 11: Steering vs. its mean-removed residual, cross-lingual Recall@20 / nDCG@20 on Belebele, and the cosine between the steering direction and the language-mean offset. Removing the mean-aligned component degrades steering for every model but bge-m3, confirming the orthogonal component carries signal.
A1 (set)
A2 (rank)
Model
Base
Cent
Steer
Base
Cent
Steer
mE5-base
30.4
28.2
33.1
57.9
59.8
60.5
mE5-large
36.4
37.9
40.1
59.7
65.2
63.4
bge-m3
51.9
51.9
48.6
73.6
75.0
71.1
Qwen3-0.6B
39.1
43.7
39.6
64.9
70.6
62.2
Qwen3-8B
33.6
44.5
41.1
67.9
74.9
68.6
Appendix
Table 12: Cross-lingual agreement under interventions on Belebele, averaged over languages ( ×100 ): set agreement A1 and rank agreement A2 for baseline, mean removal (Cent), and steering (Steer).
While mixed-language querying is ubiquitous in multilingual communities, the sensitivity of dense retrievers to such queries remains poorly understood. We present a ratio-controlled study on mMARCO that systematically evaluates retrieval performance by varying the mixing proportion of parallel query translations via embedding-level mixing -- constructing mixed queries as an interpolation of monolingual embeddings. Experiments with BGE-M3 demonstrate that an optimal mixing ratio outperforms the best monolingual endpoint in 88/105 cases. We uncover a distinct asymmetry driven by English dominance: mixing is uniformly beneficial when retrieving from non-English document indices, whereas indices containing English are best served by pure English queries. Furthermore, English acts as the strongest mixing partner for every non-English document language. Finally, when controlling for English dominance, mixing gains correlate negatively with typological distance. We conclude that language-mix sensitivity is structured and predictable, and we validate the robustness of these patterns across model families and scales.
Multilingual dense retrieval aims to handle queries and documents across different languages based on a unified retriever model. The challenge lies in enabling robust retrieval transfer to low-resource languages where annotated retrieval data is often scarce. Although previous studies transfer high-resource supervision to low-resource languages in multilingual semantic representation learning, the shared representation often entangles semantic and linguistic features, which may interfere with optimizing semantic relevance for retrieval. Different from existing methods that focus on learning language-agnostic semantic features under such entanglement, we propose a disentangled contrastive learning~(DCL) method for multilingual dense retrieval by separating multilingual representations into semantic and linguistic subspaces. Specifically, we design disentangled optimization objectives based on hierarchical semantic alignment and language debiasing contrastive learning. By aligning retrieval-relevant semantics across languages at both sentence and token levels while capturing language-specific variations in the linguistic subspace, these objectives reduce language-induced interference in semantic matching. We jointly optimize them with the retrieval objective to facilitate stable zero-shot transfer from English supervision to multilingual dense retrieval. Extensive experiments on mMARCO and MIRACL show that our method consistently outperforms several strong baselines, demonstrating its effectiveness and generalization ability.
Chao Huang, Yufeng Chen, Changhao Guan +3
Key Laboratory of Big Data & Artificial Intelligence in Transportation (Beijing Jiaotong University), Ministry of Education · School of Computer Science and Technology, Beijing Jiaotong University
Dense retrieval systems increasingly rely on multi-billion-parameter language models, whose memory and computational requirements make large-scale indexing, frequent corpus updates, and low-latency serving costly. We present a three-stage training pipeline for developing compact and efficient retrievers that remain competitive with substantially larger models. The pipeline combines cross-lingual alignment, relational knowledge distillation, and contrastive fine-tuning. It requires no original ground-truth relevance labels, relying exclusively on supervision generated by strong embedding models and rerankers utilised as teachers. Using this pipeline, we develop PolDense and EuroDense, both supporting contexts of up to 8,192 tokens. PolDense is a family of six Polish retrievers ranging from 17M to 1B parameters. EuroDense is a 435M-parameter retriever supporting nine European languages. We conduct an extensive evaluation covering 41 Polish and 150 multilingual retrieval tasks. The results demonstrate strong quality-efficiency trade-offs. PolDense-1B outperforms the evaluated retrievers with up to 9B parameters, while the PolDense family forms the Pareto frontier across model sizes. Among the evaluated models below 1B parameters, EuroDense ranks first in both task-averaged and language-averaged performance and leads in seven of nine languages. We release all models publicly.