Cross-Lingual Reusability of LLM Pipelines: Financial Extraction, Multilingual Reasoning, and Semantic Screening

Joong Ho; Choi; Sherine George; Jiayang Zhao; Jayesh Rathi; Ashwin Kandath; Ananya Sane; Vishvak Srinivasan

cs.CLOct 19, 2026

Cross-Lingual Reusability of LLM Pipelines: Financial Extraction, Multilingual Reasoning, and Semantic Screening

Authors: Joong Ho, Choi, Sherine George, Jiayang Zhao, Jayesh Rathi, Ashwin Kandath, Ananya Sane, Vishvak Srinivasan

Organizations: Joshua · BNY AI Hub · BNY Responsible AI Office

Abstract

Financial institutions increasingly deploy large language models (LLMs) in multilingual markets, raising a practical question: can pipelines developed and validated in English be reused across languages, or must they be rebuilt per locale? We test cross-lingual reusability along three dimensions--accuracy, reasoning-effort transfer, and behavioural stability--through three complementary tasks. Task 1 (cheque extraction) and Task 2 (Mintaka QA with pinned evidence across nine languages and nine model-effort configurations) test whether accuracy and effort settings transfer; Task 3 (FinanceBench pilot, 50 items over SEC 10-K filings in four languages) validates whether these patterns generalise to financial content. Accuracy reusability is broadly confirmed: Gemini 3.5 Flash achieves 95.2-97.6% extraction accuracy across prompt languages, and FinanceBench per-language means span 5.8pp (77.1-82.9%), though the language-difficulty ranking reverses between domains (Hindi is worst on Mintaka, best on FinanceBench at 82.9%). Effort reusability is domain-dependent: higher reasoning effort yields marginal gains on general-knowledge QA but up to 8pp on financial content, and effort configurations that help in English help proportionally in other languages. Behavioural reusability holds: citation-set overlap with English averages 77-79% Jaccard, with no single language consistently worse. Together, these results support qualified reuse of English-built pipelines, but per-domain, per-language validation remains necessary--neither semantic similarity nor self-reported confidence reliably predicts cross-lingual accuracy.

Figures & tables

Figure 1: Example cheque image from the Handwritten Cheque VQA dataset Verma (2024) . The model must extract the courtesy amount (numeric, from the amount box) and the legal amount (written in words, converted to a number). All cheque images are in English; only the prompt language varies across experimental conditions.

Model	Prompt	Courtesy	Legal	Joint
GPT-5.1	English	63.7%	83.1%	49.2%
	Spanish	70.2%	83.1%	54.8%
	Japanese	65.3%	84.7%	54.0%
	Hindi	67.7%	83.9%	52.4%
	Korean	63.7%	84.7%	50.8%
Claude Opus 4.6	English	85.5%	97.6%	83.9%

Table 1: Cheque amount extraction accuracy by prompt language and model family on 124 evaluation items. Courtesy : numeric amount from the amount box; Legal : amount in words converted to number; Joint : both correct. All cheque images are in English; only the prompt language varies. Gemini achieves the highest baseline (96.0% joint), Claude 83.9%, and GPT 49.2%. The multilingual effect is model-dependent: GPT-5.1 benefits, Gemini is near-ceiling, and Claude shows mild degradation with non-English prompts.

		Task 2: Mintaka (9 languages)									Task 3: FinanceBench
Family	Level	en	ar	de	es	fr	hi	it	ja	pt	en	hi	es	de
GPT-5.1	low	78.8	76.5	76.8	76.8	77.2	76.2	77.5	75.5	78.2	76.0	80.0	78.0	74.0
	medium	78.5	77.0	77.2	77.8	77.9	77.5	77.8	76.2	78.0	78.0	82.0	82.0	78.0
	high	79.0	77.2	77.2	78.2	77.7	77.8	78.2	77.0	77.8	84.0	88.0	84.0	80.0
Claude Opus 4.6	standard	80.2	78.0	79.2	80.0	79.2	78.8	78.8	77.2	79.2	78.0	84.0	78.0	82.0
	thinking	80.8	78.8	80.5	80.5	79.4	79.8	79.2	78.5	80.2	84.0	86.0	86.0	76.0

Table 2: Cross-lingual QA accuracy (%) by model configuration and language on Task 2 (Mintaka, 400 items per cell) and Task 3 (FinanceBench, 50 items per cell). On Mintaka, Claude at thinking/adaptive dominates; on FinanceBench, GPT-5.1 at high leads. Hindi is the weakest language on Mintaka but the strongest on FinanceBench, confirming that language difficulty is domain-dependent. Bold marks each column’s maximum.

Figure 2: Mean output tokens by language and reasoning effort, counted with each family’s native tokenizer (not comparable across panels). Each panel shows one model family with bars grouped by effort level. GPT-5.1 has the smallest relative cross-language variation ( ${\sim}12\%$ ); Claude Opus 4.6 the largest ( $>$ 20%), with Hindi as the most expensive language and standard mode producing more tokens than thinking or adaptive modes.

Task 2: Mintaka (prediction vs ground truth)
		Ada		MPNet
Language	$n$	$\Delta$	95% interval	$\Delta$	95% interval
Hindi	243	$-0.0790$	$[-0.0880,-0.0701]$	$-0.1167$	$[-0.1382,-0.0964]$
Arabic	240	$-0.0323$	$[-0.0384,-0.0263]$	$-0.0314$	$[-0.0470,-0.0159]$
Japanese	238	$-0.0215$	$[-0.0267,-0.0162]$	$-0.0166$	$[-0.0321,-0.0014]$
Portuguese	245	$-0.0095$	$[-0.0128,-0.0066]$	$-0.0108$	$[-0.0199,-0.0021]$

Table 3: Answer semantic-similarity gaps by target language under Ada and MPNet. Each

\Delta

is target minus English cosine similarity, paired by item and model condition. On Mintaka, Hindi shows the largest gap under both encoders; on FinanceBench, gaps are an order of magnitude smaller and German is the only language with a significant MPNet deficit.

n

is the number of paired comparisons.

Dimension	Mintaka	FinBench	Xfer?
Lang. diff. rank	en $>$ es $>$ de $>$ hi	hi $>$ es $>$ en $>$ de	No
Cross-ling. gap	$+$ 0.9 pp	$+$ 0.2 pp (mix)	Partial
Depth lift (lo $\to$ hi)	$+$ 0.5 pp	$+$ 2.0 pp	Partial
Model fam. rank	Cl $>$ GPT $>$ Gem	Cl $>$ GPT $>$ Gem	Yes ∗
Numer. robust.	$+$ 1.0 pp	$+$ 2.0 pp	Yes
Verbose $=$ incorrect	All longer ( $1.95{\times}$ )	Mixed ( $0.88{\times}$ )	No

Table 4: Cross-task reusability: do Mintaka patterns predict FinanceBench? Language ranking and the verbosity signal do not transfer; cross-lingual gap magnitude and reasoning-depth lift transfer only partially; model ranking, numerical robustness, and poor confidence calibration transfer. ∗ Overall ranking is preserved, but at high effort GPT-5.1 leads on FinanceBench while Claude leads on Mintaka. Gaps are mean en

-

target accuracy differences across the three model families at high effort, computed over the three non-English languages shared between benchmarks (Hindi, Spanish, German).

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Lang.	Cited	Cov.	Blanket	Prec.	Recall	Jaccard
en	97.6%	64.0%	44.7%	78.2%	69.5%	—
ar	98.7	63.1	43.9	79.0	68.0	78.2%
de	98.3	62.2	43.7	78.6	67.2	78.8
es	98.6	64.0	46.1	78.2	68.8	79.1
fr	98.8	63.7	45.8	78.1	68.5	79.0
hi	99.0	61.3	41.5	79.5	66.6	78.2

Appendix

Table 5: Citation and source-selection metrics averaged over the nine model–effort configurations. Cited : answers with

\geq

1 citation; Cov. : share of available references cited; Blanket : citing every reference; Prec./Recall : proxy alias-coattributed precision and recall (boolean items excluded); Jaccard : citation-set overlap with English. All target-language comparisons are localised-condition effects, not pure language effects.

Panel of Figure 3 — Figure 3: Output characteristics split by answer correctness and reasoning effort across three model families.

Cross-Lingual Transfer

Cross-lingual transfer enables models to apply knowledge or capabilities learned in one language to tasks in another.

See all 256 papers

Jul 24, 2026

Language-Routed RAG and Direct Option Scoring for Multilingual Financial QA: DS@GT at FinMMEval

We present DS@GT's submission to FinMMEval 2026 Task 1, a multilingual financial exam question answering benchmark spanning English, Spanish, Greek, Chinese, and Hindi. Financial certification exams such as the CFA, EFPA, and CPA demand structured domain reasoning that standard NLP benchmarks do not capture, and this challenge compounds across languages where retrieval and representation infrastructure is underdeveloped. We build a retrieval-augmented pipeline on LangGraph that detects query language and retrieves semantically relevant exemplars from a 30,209-entry multilingual knowledge base using BGE-M3 embeddings and FAISS indexing. The system then scores answers via Retrieval-Augmented Direct Scoring (RADS), reading next-token log-probabilities over candidate option letters rather than generating free-form output. For low-resource languages, we fuse per-language and cross-lingual retrieval indices using weighted Reciprocal Rank Fusion. Model selection is language-routed: Qwen3-14B for Arabic, Chinese, and Hindi; Qwen2.5-14B for English; and Llama-3.1-8B for Greek, a routing derived from empirical ablations that reveal substantial language-asymmetric performance gaps. Notably, chain-of-thought prompting significantly degrades Greek accuracy (90.7% to 20.9%), and enabling Qwen3's default thinking mode collapses Arabic RADS performance to near-chance levels. Our results indicate that effective multilingual financial reasoning requires language-aware retrieval, model routing, and deliberate scoring strategy selection.

Justice Ayela, Kabir Sahni

Georgia Institute of Technology, North Ave NW, Atlanta, GA 30332

Finance Multilingual Language Model Evaluation Retrieval-Augmented Generation Multilingual QA LLM Routing Multiple-Choice Question Answering+3

Ask PDF

Apr 14, 2026

ReasonXL: A Large-Scale Corpus for Language-Specific Reasoning in LLMs

Despite advances in multilingual capabilities, most large language models (LLMs) remain English-centric in their training and, crucially, in the reasoning traces they produce: even for non-English problems, they predominantly reason in English, leaving non-English users with intermediate reasoning they cannot follow. We address this disparity with two contributions. (a) We introduce

\texttt{ReasonXL}

, the first large-scale parallel corpus of cross-domain reasoning traces in five European languages (English, German, French, Italian, and Spanish): over 2.5M aligned samples (11B tokens) per language, each comprising prompt, reasoning trace, and final output, enabling direct supervision of language-specific reasoning. (b) Using

\texttt{ReasonXL}

, we demonstrate that LLMs can be adapted to reason entirely in a target language through a simple two-stage pipeline of supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR), matching or exceeding the base models' English-reasoning performance in two of four languages and recovering most of it in the remaining two.

Daniil Gurgurov, Tom Röhr, Sebastian von Rohrscheidt +3

German Research Center for Artificial Intelligence (DFKI), Saarland University · Berliner Hochschule für Technik

Cross-Lingual Reasoning in Language Models Multilingual Language Models

Ask PDF

Jun 25, 2026

Causal Connections: Leveraging Multilingual Fine-Tuning for Financial QA@FinCausal 2026

This paper describes team HSA_CORAL's submission to the FinCausal 2026 shared task on extracting cause-effect relations from financial narratives via extractive question answering in English and Spanish. We compare three modeling families: (i) encoder-only token tagging with multilingual BERT, (ii) encoder-decoder generation with multilingual BART, and (iii) decoder-only LLMs (Llama 3.1 and GPT variants) using prompt refinement, few-shot demonstrations, and supervised fine-tuning. Across settings, prompting and few-shot examples yield competitive performance, while supervised fine-tuning provides the largest gains. Our best system, GPT-4.1 Mini fine-tuned on combined English and Spanish training data, achieves a tied highest score on the English subtask (score 4.8140) and ranks third on Spanish (score 4.7753) under the shared task's LLM-as-a-judge metric. Overall, the results highlight the value of task-specific adaptation and multilingual fine-tuning for cross-lingual transfer in financial causality QA.

Akash Kumar Gautam, Serhii Hamotskyi, Christian Hänig

Anhalt University of Applied Sciences

Relation Extraction LLM Fine-Tuning Multilingual QA Finance Cross-Lingual Transfer

Ask PDF

Abstract

Cross-Lingual Transfer

Explore similar work

Language-Routed RAG and Direct Option Scoring for Multilingual Financial QA: DS@GT at FinMMEval

ReasonXL: A Large-Scale Corpus for Language-Specific Reasoning in LLMs

Causal Connections: Leveraging Multilingual Fine-Tuning for Financial QA@FinCausal 2026