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

Introduction

Financial institutions are rapidly adopting large language models (LLMs) for tasks ranging from document processing to knowledge retrieval. As these systems are deployed globally, a key question is whether solutions developed and validated in English—such as prompts, evaluation pipelines, and evidence corpora—transfer reliably to other languages. If translation alone preserves performance, institutions can reuse AI pipelines across markets rather than redeveloping them language by language.

We investigate this question through three tasks. Task 1 examines document information extraction in a financial setting: given a handwritten cheque image, an LLM extracts the numeric courtesy amount and the legal amount written in words under prompts in different languages. Task 2 examines cross-lingual question answering with deep research, where an LLM reasons over retrieved evidence to produce an answer. We evaluate this setting using the nine-language Mintaka benchmark [Sen et al., 2022] with a fixed evidence corpus, allowing language and reasoning effects to be studied independently of retrieval variance. Task 3 then validates whether these patterns hold on domain-specific financial content through a pilot study on the FinanceBench benchmark Islam et al. [2023], evaluating 50 questions over SEC 10-K filings in four languages and directly testing the generalisability of the RQ1–RQ2 findings beyond general-knowledge content.

These three tasks are deliberately chosen to span distinct dimensions of the LLM pipeline architectures that financial institutions deploy. Task 1 is perception-dominated: the answer is visible on the cheque image, and the challenge is accurate extraction under varied instruction languages—a low-reasoning, high-accuracy setting typical of back-office operations (cheque clearing, payment processing) where errors carry direct financial cost. Task 2 is reasoning-dominated: the model must plan, synthesise, and judge across multi-source evidence—a high-reasoning setting representative of front-office knowledge work (research, compliance, due diligence). Task 3 is domain-validation-focused: it applies the same cross-lingual QA design to financial filings, testing whether patterns observed on general-knowledge questions extend to domain-specific financial content. The three tasks further differ in modality (vision+text vs. text-only), in how language enters (instruction-only vs. full cross-lingual context including translated evidence), in domain, and in evaluation objectivity (exact numeric match vs. rule+judge grading). By holding model families constant across all three tasks, we can test whether findings—such as “model choice dominates language”—generalise across differences in difficulty, domain, and modality rather than being artefacts of a single task type.

Concretely, we test cross-lingual pipeline reusability along three dimensions, each operationalised as a research question:

RQ1 (Accuracy reusability) Can an English-validated LLM pipeline be reused in other languages without significant accuracy loss? We test this on cheque extraction (Task 1), general-knowledge QA (Task 2), and financial QA (Task 3) across up to nine languages.

RQ2 (Effort reusability) Do reasoning-effort configurations tuned in English transfer to non-English settings? We test this using 400 items × 9 languages × 9 model–effort configurations with a pinned evidence corpus that eliminates retrieval variance (Task 2), plus a financial QA pilot (Task 3).

RQ3 (Behavioural reusability) Does citation and source-selection behaviour remain stable when the prompt language changes? We measure citation-set overlap and proxy precision/recall across languages under fixed evidence (Task 2).

Related Work

Several benchmarks evaluate LLM capabilities on financial tasks. FLUE [Shah et al., 2022] provides five English financial NLP tasks—sentiment analysis, named entity recognition, structure boundary detection, question answering, and stock movement prediction—and introduces FLANG, a pre-trained language model for finance. BloombergGPT [Wu et al., 2023] trains a 50B-parameter model on a mixed finance–general corpus and evaluates it on standard financial NLP benchmarks, demonstrating that domain-specific pre-training improves financial task performance without degrading general capabilities. FinBen [Xie et al., 2024] offers a holistic evaluation across 35 datasets spanning seven task categories including information extraction, question answering, and risk management. FinanceBench [Islam et al., 2023] focuses on open-book financial QA over public filings, testing retrieval-augmented generation in a setting closer to real analyst workflows. Despite their breadth, these benchmarks are English-only or English-dominant. Recent work has begun to address this gap: MultiFinBen [Peng et al., 2026] introduces a multilingual financial benchmark across five languages, but evaluates models on static task sets rather than testing whether an English-built pipeline can be reused cross-lingually.

In document understanding, OCR-free vision–language models such as Donut [Kim et al., 2022] have enabled end-to-end extraction and have been applied to cheque processing [Zhao et al., 2024]. Multimodal LLMs also perform strongly on printed documents [Ma et al., 2024, Pfitzmann et al., 2022], but cross-lingual prompting effects on handwritten financial document extraction remain largely unexplored [Mondshine et al., 2025].

Cross-lingual performance gaps in LLMs are well documented but poorly explained. Sapenov and Savelka [2026] evaluate two proprietary LLMs across 43 languages and show that token usage varies systematically by language—some languages are simultaneously more expensive and less accurate—suggesting that generation behaviour is not language-neutral even when accuracy is controlled. Chen et al. [2025b] find that non-English reasoning traces produce higher output diversity, with diversity increasing as the thinking language diverges from English in embedding space. Wang et al. [2025] propose on-policy thinking-language selection and show that multilingual chain-of-thought expands the exploration space, achieving higher pass@$k$ under a fixed training budget. Most relevant to our work, Lu et al. [2026] evaluate cross-lingual deep research agents and find substantial degradation when evidence is translated: accuracy remains lower even when all gold evidence is supplied directly. Deep research agents combine retrieval, planning, and multi-step reasoning into end-to-end workflows, and Wang et al. [2026] propose REFLECT for meta-evaluating such agents, finding that even the best judges remain below 55% accuracy on evidence verification. Crucially, because Lu et al.’s benchmark does not control retrieval—each run searches anew—it cannot fully separate retrieval degradation from reasoning degradation.

Translation quality is itself a potential confound in multilingual evaluation. Thellmann et al. [2026] find that target-side translation errors are associated with 6–8 percentage-point reductions in accuracy, increasing to 6–11 points among items answered correctly in English, while automated error-span detection achieves only moderate agreement with expert annotations. These findings motivate interpreting differences under translated inputs as end-to-end outcomes of a translation-mediated pipeline rather than pure measures of intrinsic language capability.

The effect of reasoning depth on LLM accuracy has been studied primarily in English. Qin et al. [2025] provide a comprehensive multilingual evaluation and find that (i) thinking language affects accuracy, (ii) reasoning traces in different languages are not simple translations, and (iii) prompt language affects the final answer even when the question is fixed. Deng et al. [2026] show that low-resource languages degrade disproportionately as chain-of-thought length increases, suggesting a resource-dependent ceiling on the benefit of deeper reasoning. Gao et al. [2026] frame language as a latent variable in reasoning and argue that English as a default thinking language is a root cause of cross-lingual bias. These studies all manipulate prompt-induced chain-of-thought; none examines the provider-exposed reasoning-effort controls—model-id suffixes, thinking budgets—that production systems actually offer.

Finally, LLM-based evaluation itself introduces methodological risk. Haldar and Hockenmaier [2025] find intra-rater Krippendorff’s $\alpha$ [2011] below the 0.80 threshold on multiple benchmarks. Chen et al. [2025a] show that pairwise evaluation amplifies superficial preferences; Ye et al. [2024] mitigate this with PRePair, a pointwise-then-compare protocol. In the multilingual setting, Hada et al. [2025] report cross-language $\kappa\approx0.3$, and Lee et al. [2026] show that answer language has more impact on judge bias than question language.

Our work addresses these gaps jointly. We systematically vary prompt language across a financial extraction task and a general-knowledge reasoning task, covering three model families, and introduce a pinned-evidence methodology that isolates reasoning quality from retrieval quality in cross-lingual evaluation. We study provider-exposed reasoning-effort controls as deployed in production systems, where the mechanism is opaque and family-specific. To mitigate judge unreliability, we use deterministic rule-based grading as the primary path—reserving the LLM judge for cases the rules cannot settle. Comparing provider-controlled retrieval in our native deep-research track against the shared-evidence track reveals how much accuracy depends on retrieval versus reasoning.

Data

Handwritten Cheque VQA Dataset (Task 1)

For Task 1 we use the Handwritten Cheque VQA dataset Verma [2024], containing 3,312 question–answer pairs over cheque images (Figure 1). Each item provides the amount in words (legal amount) and in figures (courtesy amount). We apply three successive filters: retain only items where both amounts are annotated (914), then only items where the two amounts differ so the model cannot copy one to the other (252), and finally only items where both fields parse to numeric values for exact-match evaluation (124).

Each of the 124 items is evaluated across three model families and five prompt languages (1,860 API calls total). All cheque images are in English; the multilingual dimension enters only through the prompt language.

Mintaka Benchmark (Task 2)

Mintaka Sen et al. [2022] is a multilingual question-answering dataset built on Wikidata. Each of its 14,000 items contains a question in English plus translations into eight languages (Arabic, German, Spanish, French, Hindi, Italian, Japanese, Portuguese), typed answers (entity, numerical, boolean, date, or string), eight topic categories (e.g. geography, movies, sports), and nine complexity types including multihop, superlative, and comparative questions. This combination of nine-language coverage and verifiable answers makes it well-suited for controlled cross-lingual evaluation.

Although Mintaka’s content spans general knowledge rather than finance, its value here is as a controlled testbed for the pipeline architecture that financial institutions deploy: retrieval, multi-step reasoning, and report synthesis over a multilingual question set with verifiable answers. Existing multilingual financial QA benchmarks either lack the nine-language breadth needed for our design (FinanceBench is English-only; MultiFinBen covers five languages) or do not provide structured answers suitable for automated evaluation at scale. We draw a stratified random sample of 400 items (category × complexity type) that underpins both evaluation tracks.

FinanceBench Pilot (Task 3: Domain Validation)

To test whether the cross-lingual patterns observed on Mintaka’s general-knowledge content hold on financial material, we conduct a domain-validation pilot using FinanceBench Islam et al. [2023], an open-source QA benchmark of 150 questions over SEC 10-K filings from 32 publicly traded companies. We draw a stratified sample of 50 items (~17 each of metrics-generated, domain-relevant, and novel-generated questions) and evaluate them in four languages (English, Hindi, Spanish, German) across the same three model families and all nine model-effort configurations. Evidence passages from the source 10-K filings are translated into each target language using the same GPT-4.1 translator [OpenAI nAI, 2025a] (with Claude Sonnet 4.5 [1] as a fallback) employed for the Mintaka evidence corpus, and questions and gold answers are translated identically. The pilot yields $50 \times4 \times9 = 1{,}800$ API calls—a deliberately moderate footprint designed to confirm or refute the Mintaka patterns on domain-specific financial content, including the effect of reasoning effort, rather than to serve as a comprehensive financial benchmark.

Method

All three tasks are evaluated on the same three model families—GPT-5.1 [OpenAI, 2025b], Claude Opus 4.6 [2], and Gemini 3.5 Flash [Google DeepMind, 2026]—enabling cross-task comparison of multilingual effects. Task 1 tests a single reasoning-effort level per family across five prompt languages; Task 2 varies reasoning effort across three levels per family (nine model configurations in total) and nine languages; Task 3 uses the same nine model–effort configurations across four languages (§ 3.3).

Task 1: Cheque Amount Extraction

Each cheque image is sent to the model with a structured prompt requesting JSON output containing two fields: Courtesy Amount (extracted from the amount box as a float) and Legal Amount (converted from words to a numeric float), along with confidence scores for each. The system prompt identifies the model as a cheque information extraction system. The model is instructed to extract each amount independently and not copy one to the other.

To test whether prompt language affects extraction accuracy, we translate the extraction prompt into four additional languages—Korean, Japanese, Hindi, and Spanish—using GPT-4.1 at temperature 0, while keeping the cheque images unchanged (all cheques are in English). GPT-4.1 was chosen because it is not among the evaluated models (GPT-5.1, Claude Opus 4.6, Gemini 3.5 Flash), avoiding circularity between the translator and the systems under test. This isolates the effect of prompt language from document language. Each of the 124 evaluation cheques is processed with all five prompt languages under the same model.

We evaluate three model families through the same orchestration path described in § 4.2 and report three accuracy measures: courtesy accuracy (predicted courtesy amount equals ground truth), legal accuracy (predicted legal amount equals ground truth), and joint accuracy (both correct simultaneously). All comparisons use exact numeric match after normalization.

Task 2: Cross-Lingual QA Design

Our experiment is a fully crossed design over three axes: item (400 Mintaka questions Sen et al. [2022], stratified by category $\times$ complexity type), language (9: en, ar, de, es, fr, hi, it, ja, pt), and model configuration (9, spanning three families—GPT-5.1, Claude Opus 4.6, and Gemini 3.5 Flash—each at three reasoning-effort levels). To control for retrieval variance—repeat runs of the same query returned different evidence in pilot tests—we construct a pinned evidence corpus: one English Google Custom Search per item, captured once and committed to version control. An offline extraction step then produces attributable source passages (1,066 passages across 400 items, mean 2.7 per item), which every subsequent evaluation run reads verbatim.

Task 2 uses two evaluation tracks. For reasoning evaluation, under shared evidence, each question–language pair receives a language-matched translation of the pinned passages. Because the underlying evidence and its translations are pinned across model configurations, cross-lingual differences cannot arise from retrieval variation; they may nevertheless reflect both model reasoning and translation characteristics. Questions use Mintaka’s human translations; evidence is translated by GPT-4.1 (temperature 0) with a URL-preserving validation step and Claude Sonnet 4.5 as a fallback—neither translator is among the evaluated models. For retrieval evaluation, we also employ native deep research with OpenAI and Gemini to execute their own end-to-end retrieval-to-report pipelines, so comparing the two tracks isolates retrieval quality as a factor.

Grading is two-stage: script-aware rule-based matching first (boolean lexicon, numeric comparison, normalised entity containment), then a GPT-5.2 [OpenAI, 2025c] LLM judge for the remaining items.

Semantic-Similarity Diagnostic

As an additional semantic diagnostic, we embed each model answer and its ground truth using OpenAI’s text-embedding-ada-002 model [OpenAI, 2022] and MPNet [Song et al., 2020] to consider encoder-specific artefacts. For each prediction we record the cosine similarity

where $G_i$ is the set of accepted aliases for item $i$. This score is a continuous semantic proximity measure, not a binary correctness label; the analysis is restricted to entity answers.

We pair each target-language response with the English response for the same item and model configuration. The item-level language–configuration interaction is

where $\Delta S$ is the low-to-high (or standard-to-adaptive) change in cosine similarity. A positive $I_{f,i}$ means the configuration change narrows the cross-lingual gap. We average over items and compute 95% percentile intervals from 5,000 bootstrap samples stratified by category × complexity type.

Citation and Source-Selection Metrics

To assess whether models change how they use evidence across languages (RQ3), we extract the numbered references each model cites and compare citation behaviour between each target-language variant and its English counterpart for the same item and model configuration. Six metrics are computed:

• Cited rate: fraction of answers containing at least one citation.

• Citation coverage: share of available references cited, adjusting for the variable number of references per item.

• Blanket citation rate: fraction of answers citing every available reference (i.e. no selective sourcing).

• Proxy precision / recall: overlap between cited references and proxy alias-coattributed references—those co-attributed with extracted passages containing a known gold-answer alias. Boolean items are excluded because their gold strings match ubiquitously.

• Citation-set Jaccard similarity: $\lvert R_{\mathrm{en}} \cap R_{\ell} \rvert/ \lvert R_{\mathrm{en}} \cup R_{\ell} \rvert$ between the valid reference sets selected in English and the target language for the same item.

Binary metrics are tested with exact paired McNemar tests; numerical metrics use paired bootstrap confidence intervals and exact sign tests. All $p$-values are adjusted within each metric using the Benjamini–Hochberg procedure. Because each non-English condition pairs a translated question with translated evidence, observed differences are localised-condition effects relative to English, not pure question-language effects.

Results

Task 1: Cheque Amount Extraction

Table 1 reports extraction accuracy on the 124-item evaluation set. English-prompt joint accuracy ranges from 49.2% (GPT-5.1) to 83.9% (Claude Opus 4.6) to 96.0% (Gemini 3.5 Flash). When the prompt is translated into four additional languages while keeping the cheque images in English, the multilingual effect is model-dependent. Non-English prompts boost GPT-5.1: Spanish reaches 54.8% joint accuracy (+5.6pp over English), and all four languages match or exceed the English baseline. Conversely, non-English prompts degrade Claude Opus 4.6: Korean drops to 76.6% ($-7.3$pp). Notably, the joint drop ($-7.3$pp) exceeds the sum of the individual field drops (courtesy $-4.0$pp, legal $-2.4$pp), indicating that errors which overlapped in English become spread across different items under Korean prompting—a de-correlation effect that amplifies the joint penalty. Gemini 3.5 Flash is near-ceiling across all languages (95.2–97.6% joint), with a range of only 2.4pp.

Model	Prompt	Courtesy	Legal	Joint
GPT-5.1	English	63.7%	83.1%	49.2%
GPT-5.1	Spanish	70.2%	83.1%	54.8%
GPT-5.1	Japanese	65.3%	84.7%	54.0%
GPT-5.1	Hindi	67.7%	83.9%	52.4%
GPT-5.1	Korean	63.7%	84.7%	50.8%
Claude Opus 4.6	English	85.5%	97.6%	83.9%
Claude Opus 4.6	Korean	81.5%	95.2%	76.6%
Claude Opus 4.6	Spanish	83.9%	97.6%	81.5%
Claude Opus 4.6	Hindi	84.7%	97.6%	82.3%
Claude Opus 4.6	Japanese	83.9%	96.8%	80.6%
Gemini 3.5 Flash	English	96.0%	100%	96.0%
Gemini 3.5 Flash	Spanish	97.6%	100%	97.6%
Gemini 3.5 Flash	Japanese	96.8%	100%	96.8%
Gemini 3.5 Flash	Hindi	96.8%	100%	96.8%
Gemini 3.5 Flash	Korean	95.2%	100%	95.2%

Two findings emerge. First, model family dominates prompt language as a factor in extraction accuracy: Gemini 3.5 Flash’s weakest language (Korean, 95.2%) exceeds Claude Opus 4.6’s best (English, 83.9%) by 11.3pp, which in turn exceeds GPT-5.1’s strongest (Spanish, 54.8%) by over 29 percentage points. Second, the multilingual prompting effect is model-dependent: it provides a substantial boost for GPT-5.1 (up to +5.6pp), is negligible for Gemini 3.5 Flash (≤2.4pp range), and actually hurts Claude Opus 4.6 (up to $-7.3$pp for Korean), suggesting that the direction and magnitude of the multilingual effect cannot be assumed uniform across model families.

Cross-Lingual QA Results across Tasks 2 and 3

Accuracy is consistent cross-lingually, but language rankings are domain-dependent. On Mintaka, Claude Opus 4.6 leads at 79.5%, followed by GPT-5.1 at 77.5% and Gemini 3.5 Flash at 77.4%; English is the strongest language for every configuration (Table 2), but the cross-lingual spread is modest—at most 3.3pp for GPT-5.1 low (78.8% vs. 75.5% Japanese) and 3.0pp for Claude standard (80.2% vs. 77.2% Japanese). On FinanceBench, overall accuracy is 79.9% with the same model-family ranking (Claude 81.8%, GPT-5.1 80.3%, Gemini 77.5%)

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
	adaptive	81.0	79.5	80.8	80.5	79.2	79.5	79.0	78.2	79.5	84.0	84.0	82.0	78.0
Gemini 3.5 Flash	low	78.8	76.5	77.8	77.2	77.9	77.5	77.0	76.2	78.2	72.0	82.0	82.0	76.0
	medium	78.5	76.0	77.8	77.5	77.7	77.2	76.8	77.2	77.2	80.0	82.0	76.0	78.0
	high	78.2	76.0	77.8	77.2	77.9	77.2	77.8	77.5	77.5	74.0	78.0	78.0	72.0

and per-language means spanning 5.8pp (77.1–82.9%). However, the language ranking reverses; Hindi is the best non-English language (82.9%), outperforming even English (78.9%), while German is the weakest (77.1%). A plausible explanation is that FinanceBench questions are shorter and more formulaic than Mintaka’s diverse trivia, reducing the surface area for translation errors that disproportionately affect morphologically rich languages. The finding suggests that the “Hindi penalty” on general-knowledge QA is domain-dependent rather than an intrinsic language limitation.

Reasoning-effort gains are domain-dependent: marginal on general knowledge, substantial on financial QA. On Mintaka, higher effort produces at most +0.8pp for GPT-5.1 and Claude, and effectively zero for Gemini. On FinanceBench, the effect is larger but not uniform across families: GPT-5.1 5.1 jumps +7.0pp (76.0% to 84.0% in English), Claude gains only +1.5pp, and Gemini declines −2.5pp. The GPT-dominated lift is consistent with numerical-reasoning demands of SEC 10-K questions benefiting from additional compute in a way that trivia questions do not.

Model selection dominates language as a cost driver, and the verbosity-as-incorrectness signal is domain-dependent. On Mintaka, reasoning effort modulates output length only for Claude (∼11% reduction from standard to adaptive); GPT-5.1 and Gemini 3.5 Flash are virtually invariant (Figure 2). Cross-language variation is likewise family-specific: GPT and Gemini are compact (∼10–12% spread), whereas Claude’s cross-language range exceeds 20%, with Hindi as the most expensive language. Incorrect answers are consistently longer than correct ones—most markedly for Claude (1.95×)—suggesting verbosity as a lightweight quality signal (Figure 3a). On FinanceBench, however, this pattern reverses: incorrect answers are shorter overall (0.88×). Only Claude preserves the Mintaka direction (1.14×), while GPT-5.1 (0.78×) and Gemini (0.85×) show shorter incorrect outputs. The reversal is question-type-dependent: domain-relevant questions still show longer incorrect answers (1.20×), but metrics-generated questions strongly reverse (0.62×), indicating that verbosity as a quality signal does not generalise across domains.

Confidence scores are poorly calibrated across all families on both tasks. On Mintaka, mean confidence on incorrect answers exceeds 0.91 for every family, with family-mean correct–incorrect gaps of only 0.015–0.022 (Figure 3b). Disaggregating by effort level reveals wider per-configuration gaps: GPT-5.1’s gap narrows from 0.028 (low effort) to 0.007 (high), and Claude Opus 4.6’s from 0.017 to 0.010, while Gemini is invariant. Uncertain predictions—those the rule engine could not settle—receive noticeably lower confidence from GPT-5.1 and Claude; Gemini’s uncertain confidence is indistinguishable from correct. On FinanceBench the same pattern holds: mean confidence on incorrect answers is 0.904 overall, with a correct–incorrect gap of only 0.028.

Gemini is the most overconfident (0.958 on incorrect), while GPT-5.1 shows the largest discriminative gap (0.048). Unlike Mintaka, higher reasoning effort widens the gap for GPT-5.1 (from 0.027 at low to 0.066 at high), suggesting that extended reasoning improves calibration on domain-specific content. Across both tasks, 93% of FinanceBench confidence scores exceed 0.8, confirming that self-reported confidence is not a reliable filter for incorrect answers regardless of domain.

Semantic similarity alone does not predict cross-lingual accuracy. On Mintaka, all eight target languages show lower answer-level cosine similarity than English, with a clear three-tier structure (Table 3): Hindi is a pronounced outlier, Arabic and Japanese form a middle tier, and the five European languages cluster near English. This ranking is robust across both Ada and MPNet encoders on Mintaka (Pearson $r = 0.987$), but the two encoders diverge on FinanceBench—German’s Ada gap is −0.004 while its MPNet gap is −0.045—further undermining similarity as a reliable signal. On FinanceBench, answer-level gaps are an order of magnitude smaller (Ada $\Delta$ from −0.007 to −0.001), and at the language level, similarity is anti-correlated with accuracy—Hindi has the lowest answer similarity yet the highest accuracy. Neither encoder’s similarity score correlates with item-level correctness (Pearson $r < 0.01$, $p > 0.75$), indicating that semantic similarity alone is not a usable proxy for cross-lingual accuracy. The language × configuration interaction is negligible for GPT-5.1 and Claude Opus 4.6 (both intervals include zero) and only marginally positive for Gemini 3.5 Flash, indicating that provider-exposed reasoning-depth changes do not differentially benefit non-English languages over English in entity-answer semantic similarity.

Language	$n$	Ada $\Delta$	Ada 95% interval	MPNet $\Delta$	MPNet 95% interval
Task 2: Mintaka (prediction vs ground truth)
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	$-0.0095$	[$-0.0128$, $-0.0066$]	$-0.0108$	[$-0.0199$, $-0.0021$]
Spanish	245	$-0.0085$	[$-0.0117$, $-0.0056$]	$-0.0117$	[$-0.0212$, $-0.0029$]
French	244	$-0.0083$	[$-0.0112$, $-0.0056$]	$-0.0038$	[$-0.0108$, $0.0026$]
Italian	239	$-0.0071$	[$-0.0108$, $-0.0040$]	$-0.0059$	[$-0.0144$, $0.0019$]
German	244	$-0.0047$	[$-0.0063$, $-0.0031$]	$-0.0029$	[$-0.0083$, $0.0021$]
Task 3: FinanceBench (prediction vs ground truth)
Hindi	450	$-0.0072$	[$-0.0104$, $-0.0040$]	$0.0039$	[$-0.0106$, $0.0184$]
Spanish	450	$-0.0013$	[$-0.0046$, $0.0019$]	$-0.0005$	[$-0.0147$, $0.0135$]
German	450	$-0.0036$	[$-0.0073$, $-0.0001$]	$-0.0445$	[$-0.0594$, $-0.0305$]

We also observe that citation behaviour is surprisingly quite stable across languages, with no consistent hierarchy. Citation presence is near-universal (97.6–99.0%), and citation-set Jaccard similarity with English averages 77–79% across all target languages (Table 5 in Appendix § 8), indicating that approximately one-fifth of cited references differ.

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
it	98.0	61.2	43.2	79.3	65.9	77.5
ja	98.1	62.3	45.2	79.0	66.4	77.2
pt	98.9	63.8	46.1	78.8	68.6	79.3

The few statistically significant comparisons (after Benjamini–Hochberg correction) show mixed directions across language, model family, and reasoning level—no consistent language hierarchy emerges. Proxy precision is stable (78–79%), while recall tends slightly lower for several languages—most notably Italian (−3.6pp), Japanese (−3.1pp), and Hindi (−2.9pp), with German close behind (−2.3pp)—weakly suggesting that some localised conditions cite a narrower set, though no individual comparison survives correction. The amount of source-selection change depends more on the model–effort configuration than on language.

Cross-task reusability: which Mintaka patterns transfer to financial QA? Table 4 tests whether seven quantitative patterns observed on Mintaka transfer to FinanceBench. Three transfer (model-family ranking, numerical-answer robustness, and poor confidence calibration), two transfer partially: reasoning-depth lift is directionally consistent but 4.0× larger on financial content, while the cross-lingual gap shrinks and becomes inconsistent across families (Gemini reverses sign). Two fail: the language-difficulty ranking reverses, and the verbosity-as-incorrectness signal reverses on financial content. In both domains, self-reported confidence exceeds 0.90 on incorrect answers, making it an unreliable quality filter.

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→hi)	+0.5 pp	+2.0 pp	Partial
Model fam. rank	Cl>GPT>Gem	Cl>GPT>Gem	Yes{sup}*`
Numer. robust.	+1.0 pp	+2.0 pp	Yes
Verbose=incorrect	All longer (1.95×)	Mixed (0.88×)	No
Poor calibration	gap 0.015–0.022	gap 0.028	Yes

Discussion

Our results converge on several cross-cutting findings that inform multilingual LLM deployment, particularly for financial institutions considering cross-lingual pipeline reuse.

Cross-lingual reusability. Across all three tasks, the evidence supports qualified optimism. Accuracy drops at most 3.3pp on Task 2, Gemini is near-ceiling across prompt languages on Task 1 (≥95.2%), and FinanceBench per-language means span only 5.8pp. However, reusability is model-dependent—Claude degrades with non-English prompts on Task 1—and the patterns that fail or transfer only partially (Table 4, rows 1–3) mean that per-language validation on the target domain is essential before deployment.

Limitations. Several limitations temper these conclusions. First, the pinned evidence corpus was captured in English and translated using GPT-4.1; observed cross-lingual gaps therefore conflate reasoning degradation with translation-quality variation, and native-language retrieval remains untested. Second, proxy precision and recall rely on alias-coattributed references, not human relevance judgments. Third, the semantic-similarity diagnostic uses a single embedding model and no calibrated threshold; it is not a substitute for factual accuracy. Fourth, the reasoning-depth axis is not commensurable across families (GPT uses a graded ladder, Claude an on/off switch, Gemini a narrow range), so cross-family depth comparisons should be read with caution. Finally, the 400-item Mintaka sample—while stratified—may not capture all difficulty profiles present in full-scale deployments. The 50-item FinanceBench pilot addresses the domain gap and yields a substantive finding (Hindi’s ranking reverses between general-knowledge and financial QA), but the pilot remains small and covers only four of the nine languages; larger-scale replication is needed to confirm the pattern.

Conclusion

We presented a systematic multilingual evaluation of three major LLM families across three tasks—financial cheque extraction, cross-lingual question answering on a general-knowledge benchmark, and a domain-validation pilot on financial QA over SEC 10-K filings—spanning up to nine languages and nine model–effort configurations.

Addressing our three reusability dimensions: accuracy reusability is broadly confirmed—non-English prompts can match or exceed English on extraction and QA, though the language-difficulty ranking is domain-dependent (RQ1). Effort reusability holds within domains: the direction of reasoning-effort effects transfers across languages, but their magnitude—and, for some families, their direction—does not transfer reliably across domains (RQ2). Behavioural reusability is stable: citation-set overlap with English averages $\sim78\%$ Jaccard, with no single language consistently worse (RQ3). The FinanceBench domain-validation pilot confirms that model-family dominance and small cross-lingual gaps extend to financial content (79.9% overall, 5.8pp best-to-worst spread), while revealing that the language-difficulty ranking is domain-dependent: Hindi, the weakest language on general-knowledge QA, is the strongest on financial QA. Additionally, output token cost is family-dependent: Claude’s cross-language range exceeds 20% (Hindi is the most expensive language), while GPT is nearly language-neutral, making model selection the dominant cost driver for multilingual deployments.

More broadly, the modest cross-lingual accuracy and citation-behaviour gaps suggest that English-developed pipelines can often be reused across languages through translation, but per-language validation on the target domain remains necessary—semantic-similarity screening does not reliably predict accuracy across domains.

Future work should extend this analysis in three directions: (i) native-language retrieval to disentangle search quality from reasoning quality, (ii) LLM-judged citation support precision and completeness to move beyond proxy reference labels, and (iii) larger-scale financial evaluations across more languages to test whether the Hindi reversal and small cross-lingual gaps observed in the 50-item pilot hold at scale.

Supplementary Tables

Supplementary Figures