Multilingual Rerankers Under-Rank Same-Language Documents; LAMAR (0.6B) Adds Language Coherence to Reranking

LAMAR: An Open Language-Aware Multilingual Alignment Reranker

Seongtae Hong, Youngjoon Jang, Jungseob Lee, Seungyoon Lee, Heuiseok Lim

cs.IR

2026-07-24

Multilingual rerankers fail to prioritize same-language documents, English worst of all. LAMAR adds English-anchored distillation plus language-coherence alignment; at 0.6B it beats 1B-4B rerankers.

What problem this solves

In multilingual RAG, the retriever returns documents in several languages, a reranker reorders them, and the result feeds the generator. The paper opens by pointing out that even when documents are semantically equivalent, the language they are written in affects generation quality, and same-language documents usually yield higher answer F1 (tested on a parallel XQuAD setup with Qwen2.5-32B and Llama-3.3-70B).

Then they show that existing rerankers ignore this. Using XQuAD across 12 languages, they build an oracle candidate set (every query has a gold document in each language) and count only the same-language one as relevant. Mainstream rerankers do not consistently rank the same-language document first. The most counterintuitive case is English: Jina-reranker-v3 puts the English document first for an English query only 27.2% of the time, Qwen3-Reranker-4B only 20.9%; Chinese and Thai documents often get ranked above it. So the task is: rank relevant above irrelevant, and among semantically equivalent candidates push the same-language document up, without hurting semantic ranking.

Method

LAMAR is a 0.6B cross-encoder (built on bge-m3-retromae) trained in two stages.

Stage one, English-anchored multilingual relevance distillation. Each sample has a semantically corresponding English anchor pair and a multilingual pair. The teacher (Qwen3-Reranker-4B) scores the English pair; the student predicts the multilingual pair, and an MSE loss pulls the student toward the teacher's English score. English is the anchor because English pairs give a stable teacher signal, placing every multilingual pair on one consistent semantic-relevance scale regardless of input language.

Stage two, preference alignment for language coherence, with a joint objective Lrank plus lambda times LLC (lambda=2). For a source-language query, construct a parallel document group (pos-source, pos-target, neg-source, neg-target) and target the order pos-source at least pos-target, then neg-source at least neg-target. Lrank is a listwise ranking loss (ADR-MSE) that preserves positive-versus-negative separation without forcing a relative order between the source and target versions; LLC applies softplus to score differences, adding the same-language-higher preference across both positive and negative groups with no hard margin.

Why two stages: stage one alone gives a consistent semantic scale but a weak language preference; LLC in stage two is the source of language coherence, and Lrank carries positive-versus-negative ranking. In ablation, removing LLC collapses the language-coherence eval (XQuAD nDCG@10 down 30 points); removing Lrank hurts general ranking (MIRACL).

Training data: stage one, 6.7M instances (MMARCO 14 languages, MIRACL 51, RLHN English); stage two, 8.6K listwise instances (MIRACL 51 languages, paired).

Results

Language coherence (parallel oracle subsets, XQuAD 12 languages and BELEBELE 14):

LAMAR takes first on nDCG@1, @10, and MRR on both XQuAD and BELEBELE, beating 1B-4B models at 0.6B, and tops every one of the 12 languages individually. The sharpest contrast is Qwen3-Reranker-4B: first on MTEB general reranking average (87.34), dead last on language coherence (33.03), yet it is LAMAR's stage-one teacher.

General reranking does not slip: across five MTEB multilingual reranking benchmarks LAMAR averages 86.84, second only to the 6.7x-larger Qwen3-4B (87.34); on MIRACL multi-monolingual (18 languages) it is second at 69.5. In practical retrieval (reranking top-20 XQuAD candidates from bge-m3) it is first on all three metrics, nDCG@1 at 92.90.

Why it matters

For anyone building multilingual RAG, this surfaces a previously underweighted thread, that document language affects generation, shows mainstream rerankers broadly miss it, and offers a 0.6B, open-weights, drop-in fix. It also punctures a reflex: bigger is not more complete. The 4B Qwen3-Reranker is first in general ability yet bottoms out on language coherence. The two-stage recipe (English-anchored distillation plus language-coherence alignment) is reusable.

Limitations

There is no Limitations section; the conclusion frames everything positively, which is itself a flag. Points that look under-supported:

• The biggest unverified claim: there is no end-to-end test of RAG answer F1 with LAMAR in the loop. The claim that language coherence improves generation rests on the motivation figure (Figure 1), not on measuring LAMAR's actual downstream answer quality.
• English anchor and Qwen3-4B teacher are fixed; no ablation over anchor language or teacher.
• The teacher is itself last on language coherence; the authors distill only its English-pair scores to sidestep its multilingual weakness, but do not justify why this is optimal.
• No analysis of when same-language priority goes wrong (if a non-query-language document is genuinely better, does LLC over-suppress it); lambda=2 is fixed with no sensitivity sweep.
• The coherence eval uses an oracle set where all candidates are semantically equivalent, the most favorable case; real candidate sets mix relevant and irrelevant (Table 6 addresses this partly, but it remains limited).

Terms

• Cross-encoder: a reranker that encodes the query and document together and emits a single relevance score.
• Reranker: the second-stage retrieval component that re-scores and reorders candidates after first-stage retrieval.
• Distillation: a student model imitates a teacher's output scores.
• Listwise ranking loss: optimizes ranking over a whole candidate list (vs. pointwise or pairwise).
• ADR-MSE: a listwise ranking loss using mean squared error over approximate ranks.
• Softplus: log(1 + e to the x), a smooth function used here to impose a soft same-language-scores-higher preference.

Source

• arXiv 2607.22042v1
• PDF
• Hugging Face

Related papers

• ColBERT Delays BERT Interaction Until MaxSim, Matching BERT Quality at 170× Lower Latency
• Off-topic docs cut BM25 nDCG 22%; independent first-stage scorers hold up
• Alipay user logs saturate past 30M users; ALGN densifies them cheaper
• Spotify: better CoT traces do not raise next-item Recall@10 under standard SFT
• SST: history-side semantic subwords lift TIGER-KM Beauty HR@5 by 10.4%
• ANDOR: multi-vector retrieval stays ~2× ahead of single-vector after fine-tuning

All paper explainers