Pith. sign in

REVIEW 3 major objections 5 minor 45 references

Multilingual rerankers should prefer documents in the query language — and LAMAR shows how to train them to do it.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 05:56 UTC pith:ESK6JMOG

load-bearing objection A well-built reranker for query-language preference, but the paper's reason for preferring that language rests on one thin figure and no end-to-end RAG test. the 3 major comments →

arxiv 2607.22042 v2 pith:ESK6JMOG submitted 2026-07-24 cs.IR

LAMAR: An Open Language-Aware Multilingual Alignment Reranker

Seongtae Hong , Youngjoon Jang , Jungseob Lee , Seungyoon Lee , Heuiseok Lim This is my paper
classification cs.IR
keywords multilingual rerankinglanguage coherenceretrieval-augmented generationcross-encoderrelevance distillationpreference alignmentmultilingual retrievalcross-lingual RAG
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that multilingual rerankers—models that reorder retrieved documents before answer generation—need to consider not just semantic relevance but also "language coherence": whether a document is written in the same language as the query. The authors first show that widely used rerankers often rank non-query-language documents above same-language ones, especially for English queries, even when the candidates are semantically equivalent. They then introduce LAMAR, a 0.6B cross-encoder trained in two stages: English-anchored relevance distillation to align relevance scores across languages, and preference alignment that explicitly pushes same-language documents higher. In controlled evaluations with oracle candidate sets, LAMAR places the query-language document first far more reliably than 13 prior rerankers, while remaining competitive on standard multilingual reranking benchmarks and achieving the best results when reranking real retrieved candidate sets.

Core claim

LAMAR demonstrates that language coherence can be learned as an explicit, trainable reranking objective without sacrificing general semantic reranking ability. On parallel oracle candidate sets where every document is semantically correct, LAMAR ranks the query-language document first at 96.89 nDCG@1 on XQuAD and 94.66 on BELEBELE—the highest among all compared rerankers—and it maintains competitive nDCG@10 averages (86.84) across five standard multilingual reranking benchmarks. The key evidence is that existing rerankers, even large ones, do not consistently prioritize the query-language document (e.g., one strong baseline ranks it first only 27.2% of the time for English queries), confirmi

What carries the argument

LAMAR's central mechanism is a two-stage training procedure on a cross-encoder. Stage 1, English-anchored relevance distillation, uses a Mean Squared Error objective to make the student's score for a multilingual query–document pair approximate the teacher's score for the corresponding English pair, putting all languages on a common relevance scale. Stage 2, preference alignment for language coherence, combines a listwise group-ranking loss (Approx Discounted Rank MSE) that separates positives from negatives with a language-coherence loss that applies softplus to the score differences between source- and target-language documents, gently rewarding higher scores for same-language documents. T

Load-bearing premise

The claim's load-bearing premise is that documents written in the query language genuinely improve RAG answer quality over equally relevant documents in other languages—supported only by Figure 1's F1 measurements with two LLMs on one dataset, and if this benefit varies by domain, model, or document quality, optimizing for language coherence could demote better evidence.

What would settle it

Conduct a multilingual RAG study where, for the same queries, non-query-language gold documents are systematically longer or contain additional supporting details, while query-language documents are shorter. If a RAG system using LAMAR's reranking produces lower answer F1 than a relevance-only reranker—because it promotes the thinner same-language documents—then the language-coherence premise fails for that setting.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If language coherence is a real and trainable capability, multilingual RAG systems can adopt LAMAR to boost answer quality without trading away general reranking accuracy.
  • Existing rerankers' inconsistent behavior on English queries suggests that language-aware training should be a first-class consideration in multilingual retrieval pipelines, not an afterthought.
  • LAMAR's 0.6B size achieving competitive or better results than 1B-4B models indicates that explicit language preference can substitute for raw model scale in this setting.
  • The ablation results indicate the group-ranking and language-coherence losses play complementary roles—coherence loss drives oracle performance, while ranking loss maintains relevance discrimination.
  • Because LAMAR also performs best on candidate sets produced by real first-stage retrievers, the benefit persists under noisy retrieval conditions, not just controlled oracle sets.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The language-coherence loss might be transferable to first-stage retrieval: one could apply the same softplus-style preference to embedding training to make retrievers themselves prioritize query-language documents.
  • The paper's premise—that same-language documents yield better RAG answers—is tested only on two instruction-tuned LLMs and one QA dataset; extending the diagnostic to other generation tasks, domains, and models could reveal where language coherence matters most or backfires.
  • A practical extension would be a confidence-weighted version of the language preference: when same-language and other-language documents are near-equivalent, the preference is strong; when a non-query-language document is clearly more detailed or authoritative, the model could be trained to override the preference.
  • The finding that English queries are particularly prone to language incoherence suggests that English-centric distillation data may inadvertently teach models to devalue English documents; LAMAR's English-anchored distillation may specifically counteract this artifact.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes LAMAR, a 0.6B multilingual cross-encoder reranker trained in two stages: English-anchored relevance distillation (Eq. 1) followed by preference alignment that combines a group ranking loss (Eq. 5) with a language-coherence loss (Eq. 6). The central claim is that existing multilingual rerankers do not consistently rank documents in the query language first when semantically equivalent candidates exist, and that enforcing this 'language coherence' improves multilingual RAG while preserving general reranking performance. LAMAR is evaluated on controlled oracle candidate sets from XQuAD and BELEBELE (Table 3), on standard MTEB multilingual reranking benchmarks (Table 5), and on retrieved top-20 candidate sets (Table 6). The ablations (Fig. 4) show both losses contribute. The paper also makes the broader claim that language coherence improves RAG answer quality, supported mainly by Figure 1.

Significance. If the central claim holds, LAMAR is a useful contribution: it is the first reranker explicitly trained for language coherence, it is substantially smaller than several baselines while achieving the best scores on the controlled oracle evaluation, and it remains competitive on standard multilingual reranking benchmarks. The two-stage training recipe, including the English-anchored distillation and the preference-alignment objective, is clearly described and the ablations are informative. The release of such a model would be valuable for multilingual RAG pipelines. However, the significance is tempered by the fact that the headline language-coherence evaluation is closely aligned with the training objective, and the evidence that language coherence actually improves RAG answer quality is thin and confounded.

major comments (3)
  1. [§4.3, Figure 1] The premise motivating Eq. (6) and the oracle labeling — that same-language documents yield better RAG answers — is supported only by Figure 1, which reports F1 for two LLMs on XQuAD across 16 query/document language pairs. This evidence is confounded: the gold answers are in the query language, so a model given a document in another language may answer in that language and receive lower F1 even when the content is fully understood. No control for answer-language mismatch, no variance estimates, and no other domains/models are provided. Because this premise is load-bearing for the paper's central value claim, the authors should either add an end-to-end RAG experiment that controls for answer language and spans more domains/models, or explicitly limit the claimed benefit to the ranking behavior rather than downstream answer quality.
  2. [§5.1, Table 3; §5.3, Table 6] The language-coherence evaluation labels the document in the query language as the only relevant one, which is exactly the preference encoded by L_LC. LAMAR's large gains in Table 3 therefore demonstrate that it implements the trained preference, but they do not independently validate that this preference is beneficial. The retrieved-candidate setting in Table 6 uses XQuAD, whose gold documents are in the query language, so it inherits the same bias. To support the claim that language coherence improves practical retrieval, the evaluation should include candidate sets where query-language documents are not always the gold, or where a non-query-language document is more informative, and measure whether LAMAR can balance language preference against content quality. Without such a test, the practical advantage of LAMAR beyond its training objective is not established.
  3. [Eq. (6), §3.2] The language-coherence loss is content-blind: it penalizes the score difference s+_tgt − s+_src for any parallel pair, regardless of which document is more informative or better written. During training, source and target documents are translation-equivalent, so the loss is well-defined. At inference, however, LAMAR has no mechanism to override the language preference when a target-language document is more informative. This is a genuine limitation given that real retrieved sets contain candidates of varying quality. The paper should either demonstrate empirically that the learned preference does not harm retrieval when non-query-language documents are superior, or modify L_LC to incorporate a content-based override (e.g., a margin that is diminished when the target document is more relevant).
minor comments (5)
  1. [§5.1] No error bars, confidence intervals, or significance tests are reported for the language-coherence results in Table 3, even though differences between some models (e.g., LAMAR vs. llama-nemotron on nDCG@1) are a few points. At least a per-language variance or a paired significance test would strengthen the comparison.
  2. [§4.1] MIRACL is used in both Stage 1 and Stage 2 training, and also appears in the MTEB evaluation (Table 5). The paper should explicitly state that the train and test splits are disjoint, or explain how train/test contamination is avoided. This is important for interpreting the MIRACL results in Table 4.
  3. [Abstract/§1] The model is called 'open' but no URL, HuggingFace identifier, or code release is provided. If the model is truly open, include a link; otherwise remove the word 'open' from the title and abstract.
  4. [§5.3, Table 6] The table header uses 'R@20' while the text says 'R, N, and M denote Recall, nDCG, and MRR.' Clarify that Recall is computed at the retrieved-candidate level (R@20) and that N@1/N@10 and M@10 are reranking metrics on the top-20 set. The current notation is somewhat confusing.
  5. [§4.2] The maximum sequence length of 8,192 is stated, but no memory or inference cost comparison with baselines is provided. Given that LAMAR uses a cross-encoder with long-context capacity, a brief cost analysis would help practitioners assess the trade-off.

Circularity Check

0 steps flagged

No significant circularity: language-coherence training and evaluation are deliberately aligned, but the paper's central results also rest on independent benchmarks and held-out data.

full rationale

The paper does not derive its main claims from its own inputs. The closest candidate is the alignment between the language-coherence loss in Eq. (6), which prefers source-language documents over target-language counterparts (s+_src > s+_tgt), and the oracle evaluation in Section 4.3, which designates the document in the query language as relevant. This is an intentional consistency between a training objective and a capability test, not a circular derivation: the evaluation sets (XQuAD, BELEBELE) are held out, and LAMAR is not fitted to them. The paper does not use Table 3 to establish that same-language documents improve answer generation; that premise is supported separately by Figure 1, an empirical observation (whose limitations are a validity concern rather than a circularity). LAMAR's general reranking quality is also checked on independent MTEB benchmarks (MIRACL, XGLUE, HUME, MLDR, Wikipedia), where it remains competitive (86.84 average), so the central contribution does not reduce to the language-coherence objective alone. The self-citations in Related Work (Jang et al. 2026; Hong et al. 2026) are contextual and not load-bearing. Overall, no step in the claimed derivation chain is equivalent to its inputs by construction.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The central claim rests on one empirical premise (same-language documents help generation), one dataset equivalence assumption, and one distillation assumption. The main free hyperparameters are λ and R. No new entities are introduced.

free parameters (2)
  • λ (language-coherence loss weight) = 2
    Chosen for L_LC in Eq. (2); no sensitivity analysis or selection procedure is reported.
  • R (Stage-2 repetition factor) = 3
    Controls number of language-pair assignments per MIRACL triplet (App. A.3); hand-chosen.
axioms (4)
  • domain assumption Same-language documents improve RAG answer generation
    Section 1/Figure 1 with Qwen2.5-32B and Llama-3.3-70B on XQuAD; generalized to all 12/14 languages and other RAG settings.
  • domain assumption Parallel oracle candidates in XQuAD/BELEBELE are semantically equivalent
    Section 2/4.3 assumes all translations preserve answer content, so only language distinguishes candidates.
  • domain assumption English teacher scores are valid multilingual relevance targets
    Stage 1 Eq. (1) regresses multilingual scores to English-pair teacher scores; assumes cross-lingual semantic identity preserves relevance.
  • domain assumption The preference order in Eq. (3) is correct
    s+_src ≥ s+_tgt > s-_src ≥ s-_tgt encodes the paper's value judgment that language coherence should only break ties among relevant/irrelevant documents.

pith-pipeline@v1.3.0-alltime-deepseek · 26269 in / 11318 out tokens · 117520 ms · 2026-08-01T05:56:40.279966+00:00 · methodology

0 comments
read the original abstract

In multilingual retrieval augmented generation pipelines, an embedding model can retrieve relevant documents written in multiple languages, which are subsequently reranked before answer generation. However, it remains unclear whether existing multilingual rerankers consider document language when ordering semantically relevant candidates. Our analysis shows that these rerankers do not consistently prioritize documents written in the same language as the query when semantically equivalent documents are available across languages, even though document language can affect answer generation. We release LAMAR, a language aware multilingual cross encoder trained to account for both semantic relevance and language coherence. LAMAR first uses English anchored relevance distillation to establish consistent relevance scoring across multilingual inputs and then applies preference alignment for language coherence to encourage documents written in the same language as the query to receive higher rankings while retaining semantic relevance. In a controlled experiment designed to assess language coherence, LAMAR achieves the best performance overall and across all languages examined individually. LAMAR also remains competitive on established multilingual reranking benchmarks. In practical retrieval settings, LAMAR achieves the best results across all reported metrics when reranking candidates retrieved in the first stage. These results demonstrate that LAMAR accounts for language coherence while achieving strong performance on general multilingual reranking benchmarks.

Figures

Figures reproduced from arXiv: 2607.22042 by Heuiseok Lim, Jungseob Lee, Seongtae Hong, Seungyoon Lee, Youngjoon Jang.

Figure 1
Figure 1. Figure 1: Multilingual RAG performance on XQuAD across [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Language coherence of Jina-reranker-v3 (Wang, Li, and Xiao 2025), Qwen3-Reranker-4B (Zhang et al. 2025), and [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: nDCG@1 for the six query languages common to XQuAD and BELEBELE under the parallel oracle setting. While [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: nDCG@10 ablation results across LAMAR train [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 1 canonical work pages

  1. [1]

    2025 , eprint=

    Qwen2.5 Technical Report , author=. 2025 , eprint=

    2025

  2. [2]

    arXiv preprint arXiv:2407.21783 , year=

    The llama 3 herd of models , author=. arXiv preprint arXiv:2407.21783 , year=

    Pith/arXiv arXiv

  3. [3]

    2021 , eprint=

    Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks , author=. 2021 , eprint=

    2021

  4. [4]

    arXiv preprint arXiv:2312.10997 , year=

    Retrieval-augmented generation for large language models: A survey , author=. arXiv preprint arXiv:2312.10997 , year=

    Pith/arXiv arXiv

  5. [5]

    2023 , eprint=

    Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection , author=. 2023 , eprint=

    2023

  6. [6]

    R e2 G : Retrieve, Rerank, Generate

    Glass, Michael and Rossiello, Gaetano and Chowdhury, Md Faisal Mahbub and Naik, Ankita Rajaram and Cai, Pengshan and Gliozzo, Alfio. R e2 G : Retrieve, Rerank, Generate. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022. doi:10.18653/v1/2022.naacl-main.194

    doi:10.18653/v1/2022.naacl-main.194 2022

  7. [7]

    2020 , eprint=

    Passage Re-ranking with BERT , author=. 2020 , eprint=

    2020

  8. [8]

    2020 , eprint=

    Document Ranking with a Pretrained Sequence-to-Sequence Model , author=. 2020 , eprint=

    2020

  9. [9]

    Retrieval-augmented generation in multilingual settings

    Chirkova, Nadezhda and Rau, David and D \'e jean, Herv \'e and Formal, Thibault and Clinchant, St \'e phane and Nikoulina, Vassilina. Retrieval-augmented generation in multilingual settings. Proceedings of the 1st Workshop on Towards Knowledgeable Language Models (KnowLLM 2024). 2024. doi:10.18653/v1/2024.knowllm-1.15

    doi:10.18653/v1/2024.knowllm-1.15 2024

  10. [10]

    XOR QA : Cross-lingual Open-Retrieval Question Answering

    Asai, Akari and Kasai, Jungo and Clark, Jonathan and Lee, Kenton and Choi, Eunsol and Hajishirzi, Hannaneh. XOR QA : Cross-lingual Open-Retrieval Question Answering. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021. doi:10.18653/v1/2021.naacl-main.46

    doi:10.18653/v1/2021.naacl-main.46 2021

  11. [11]

    Multilingual Retrieval-Augmented Generation for Knowledge-Intensive Question Answering Task

    Ranaldi, Leonardo and Haddow, Barry and Birch, Alexandra. Multilingual Retrieval-Augmented Generation for Knowledge-Intensive Question Answering Task. Findings of the A ssociation for C omputational L inguistics: EACL 2026. 2026. doi:10.18653/v1/2026.findings-eacl.35

    doi:10.18653/v1/2026.findings-eacl.35 2026

  12. [12]

    arXiv preprint arXiv:2509.13930 , year=

    Linguistic Nepotism: Trading-off Quality for Language Preference in Multilingual RAG , author=. arXiv preprint arXiv:2509.13930 , year=

    Pith/arXiv arXiv

  13. [13]

    Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    All Languages Matter: Understanding and Mitigating Language Bias in Multilingual RAG , author=. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  14. [14]

    Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks

    Reimers, Nils and Gurevych, Iryna. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. 2019

    2019

  15. [15]

    Findings of the association for computational linguistics: ACL 2024 , pages=

    M3-embedding: Multi-linguality, multi-functionality, multi-granularity text embeddings through self-knowledge distillation , author=. Findings of the association for computational linguistics: ACL 2024 , pages=

    2024

  16. [16]

    arXiv preprint arXiv:2509.25085 , year=

    Jina-reranker-v3: Last but Not Late Interaction for Listwise Document Reranking , author=. arXiv preprint arXiv:2509.25085 , year=

    arXiv

  17. [17]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track , pages=

    mgte: Generalized long-context text representation and reranking models for multilingual text retrieval , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing: Industry Track , pages=

    2024

  18. [18]

    arXiv preprint arXiv:2506.05176 , year=

    Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models , author=. arXiv preprint arXiv:2506.05176 , year=

    Pith/arXiv arXiv

  19. [19]

    2025 , url=

    Contextual AI Reranker v2 , author=. 2025 , url=

    2025

  20. [20]

    arXiv preprint arXiv:2509.12541 , year=

    zELO: ELO-inspired Training Method for Rerankers and Embedding Models , author=. arXiv preprint arXiv:2509.12541 , year=

    arXiv

  21. [21]

    arXiv preprint arXiv:2604.23734 , year=

    Prism-Reranker: Beyond Relevance Scoring--Jointly Producing Contributions and Evidence for Agentic Retrieval , author=. arXiv preprint arXiv:2604.23734 , year=

    Pith/arXiv arXiv

  22. [22]

    CoRR , volume =

    Luiz Bonifacio and Israel Campiotti and Roberto de Alencar Lotufo and Rodrigo Frassetto Nogueira , title =. CoRR , volume =. 2021 , url =. 2108.13897 , timestamp =

    Pith/arXiv arXiv 2021

  23. [23]

    2025 , eprint=

    Fixing Data That Hurts Performance: Cascading LLMs to Relabel Hard Negatives for Robust Information Retrieval , author=. 2025 , eprint=

    2025

  24. [24]

    CoRR , volume =

    Mikel Artetxe and Sebastian Ruder and Dani Yogatama , title =. CoRR , volume =. 2019 , archivePrefix =

    2019

  25. [25]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    The belebele benchmark: a parallel reading comprehension dataset in 122 language variants , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  26. [26]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    XGLUE: A new benchmark dataset for cross-lingual pre-training, understanding and generation , author=. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages=

    2020

  27. [27]

    arXiv preprint arXiv:2210.09984 , year=

    Making a miracl: Multilingual information retrieval across a continuum of languages , author=. arXiv preprint arXiv:2210.09984 , year=

    Pith/arXiv arXiv

  28. [28]

    Wikimedia Foundation , title =

  29. [29]

    arXiv preprint arXiv:2510.10062 , year=

    HUME: Measuring the Human-Model Performance Gap in Text Embedding Tasks , author=. arXiv preprint arXiv:2510.10062 , year=

    arXiv

  30. [30]

    arXiv preprint arXiv:2502.13595 , year=

    Mmteb: Massive multilingual text embedding benchmark , author=. arXiv preprint arXiv:2502.13595 , year=

    arXiv

  31. [31]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    Investigating language preference of multilingual rag systems , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  32. [32]

    Liu, Wei and Trenous, Sony and Ribeiro, Leonardo F. R. and Byrne, Bill and Hieber, Felix. XRAG : Cross-lingual Retrieval-Augmented Generation. Findings of the Association for Computational Linguistics: EMNLP 2025. 2025. doi:10.18653/v1/2025.findings-emnlp.849

    doi:10.18653/v1/2025.findings-emnlp.849 2025

  33. [33]

    2026 , eprint=

    MLAIRE: Multilingual Language-Aware Information Retrieval Evaluation Protocal , author=. 2026 , eprint=

    2026

  34. [34]

    Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP) , pages=

    Making monolingual sentence embeddings multilingual using knowledge distillation , author=. Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP) , pages=

    2020

  35. [35]

    arXiv preprint arXiv:2604.05684 , year=

    Improving semantic proximity in information retrieval through cross-lingual alignment , author=. arXiv preprint arXiv:2604.05684 , year=

    Pith/arXiv arXiv

  36. [36]

    2026 , eprint=

    Enhancing Multilingual RAG Systems with Debiased Language Preference-Guided Query Fusion , author=. 2026 , eprint=

    2026

  37. [37]

    Proceedings of the 5th Workshop on Multilingual Representation Learning (MRL 2025) , pages=

    On the consistency of multilingual context utilization in retrieval-augmented generation , author=. Proceedings of the 5th Workshop on Multilingual Representation Learning (MRL 2025) , pages=

    2025

  38. [38]

    Proceedings of the 29th International Conference on Computational Linguistics , pages=

    Parameter-efficient neural reranking for cross-lingual and multilingual retrieval , author=. Proceedings of the 29th International Conference on Computational Linguistics , pages=

  39. [39]

    Zero-Shot Cross-Lingual Reranking with Large Language Models for Low-Resource Languages

    Adeyemi, Mofetoluwa and Oladipo, Akintunde and Pradeep, Ronak and Lin, Jimmy. Zero-Shot Cross-Lingual Reranking with Large Language Models for Low-Resource Languages. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers). 2024. doi:10.18653/v1/2024.acl-short.59

    doi:10.18653/v1/2024.acl-short.59 2024

  40. [40]

    Training of LLM -Based List-Wise Multilingual Reranker

    Yu, Hao and Adelani, David Ifeoluwa. Training of LLM -Based List-Wise Multilingual Reranker. Proceedings of the 5th Workshop on Multilingual Representation Learning (MRL 2025). 2025. doi:10.18653/v1/2025.mrl-main.42

    work page doi:10.18653/v1/2025.mrl-main.42 2025

  41. [41]

    Evaluating Large Language Models for Cross-Lingual Retrieval

    Zuo, Longfei and Hong, Pingjun and Kraus, Oliver and Plank, Barbara and Litschko, Robert. Evaluating Large Language Models for Cross-Lingual Retrieval. Findings of the Association for Computational Linguistics: EMNLP 2025. 2025. doi:10.18653/v1/2025.findings-emnlp.612

    doi:10.18653/v1/2025.findings-emnlp.612 2025

  42. [42]

    ArXiv , year=

    Improving Efficient Neural Ranking Models with Cross-Architecture Knowledge Distillation , author=. ArXiv , year=

  43. [43]

    European Conference on Information Retrieval , pages=

    Rank-distillm: Closing the effectiveness gap between cross-encoders and llms for passage re-ranking , author=. European Conference on Information Retrieval , pages=. 2025 , organization=

    2025

  44. [44]

    2023 , eprint=

    Cross-lingual Knowledge Transfer via Distillation for Multilingual Information Retrieval , author=. 2023 , eprint=

    2023

  45. [45]

    arXiv preprint arXiv:2601.09012 , year=

    Translategemma technical report , author=. arXiv preprint arXiv:2601.09012 , year=

    arXiv