Multilingual Prompt Exploration

• Multilingual Prompt Exploration is the systematic design and evaluation of prompts across languages to improve zero-shot and few-shot transfer.
• It categorizes methods into regimes such as shared prompts, language-specific translations, prompt transformation, and dynamic language routing, enhancing alignment and performance.
• Empirical studies show optimized multilingual prompts enhance classification, NLI, and generation tasks by addressing training-inference mismatches and cultural biases.

Searching arXiv for recent and foundational papers on multilingual prompt exploration. Multilingual Prompt Exploration is the systematic design, adaptation, and evaluation of prompts across multiple languages in order to improve zero-shot transfer, few-shot learning, robustness, factual recall, diversity, or controllability in multilingual and cross-lingual settings. In current research, the term spans several prompt regimes: language-agnostic unified prompts for multilingual pretrained LLMs, translated or language-specific discrete prompts, soft prompts and prompt translators, multilingual prompt bundles used simultaneously, and language-routed prompts selected dynamically at inference time. Across these variants, the central problem is that prompt behavior is rarely invariant across languages: prompt wording, verbalizers, reasoning language, typological distance, and language-resource imbalance can materially affect model performance, consistency, and deployment complexity (Huang et al., 2022, Qiu et al., 2024, Vatsal et al., 16 May 2025).

1. Conceptual scope and problem formulation

Multilingual prompt exploration arises from a mismatch between the multilingual coverage of pretrained models and the language sensitivity of prompting interfaces. In prompt-based tuning for multilingual pretrained LLMs, one often trains on a source language such as English and evaluates directly on other languages without target-language labels. The prompt then becomes part of the transfer mechanism rather than a neutral wrapper (Huang et al., 2022). This issue is particularly acute in low-resource and few-shot regimes, where prompt choice can dominate downstream performance (Zhou et al., 2022, Qiu et al., 2024).

A recurring distinction in the literature is between multilingual prompting and cross-lingual prompting. In the former, prompts are intended to work across several languages, sometimes with one shared prompt, sometimes with language-specific variants. In the latter, prompts are optimized or learned in one language and then transferred to another, either directly, via translation, or through an intermediate transformation (Huang et al., 2022, Qiu et al., 2024). A closely related distinction is between prompt language and reasoning or retrieval language: some methods answer in the query language while internally exploring other languages to access latent knowledge or aligned representations (Diskind et al., 23 Jun 2026).

Another axis concerns prompt representation. Research separates discrete prompts, which use natural-language tokens, from soft prompts, which are trainable continuous embeddings, and from model-based prompts, where a prompt is itself encoded by part of a multilingual model before being fused with the input (Huang et al., 2022). This distinction is not merely taxonomic. Discrete prompts align with masked-language-model pretraining but are language-specific and labor-intensive to design; soft prompts are language-agnostic in principle but may suffer from a pretrain–finetune mismatch because the prompt vectors never appeared during pretraining (Huang et al., 2022). Model-based prompt encoders such as UniPrompt attempt to reconcile these trade-offs by preserving multilingual semantic alignment while avoiding per-language prompt engineering (Huang et al., 2022).

A broader perspective treats multilingual prompt exploration as a search problem over prompt components, languages, and aggregation rules. This view appears in work on multilingual system-prompt optimization, prompt translators, and cross-lingual exploration for parametric knowledge, where prompt performance depends jointly on language selection, routing, answer aggregation, and inference budget (Zhang et al., 2 Dec 2025, Qiu et al., 2024, Diskind et al., 23 Jun 2026). A plausible implication is that prompt design in multilingual settings should be regarded less as translation of a fixed English artifact and more as structured control over model-internal multilingual geometry.

2. Prompt regimes and architectural patterns

The literature has converged on several recurrent architectures for multilingual prompt exploration. The first is the shared prompt regime, where one prompt is used across languages. UniPrompt is the clearest example: it uses a unified, language-agnostic, model-based prompt built from a Template Tower and a Context Tower, both initialized from the lower layers of XLM-R, followed by a shared Fusion Tower initialized from upper layers (Huang et al., 2022). The prompt is encoded separately from the input for the first $p$ layers, concatenated with the input representation, and fused only in the upper layers, with the empirical choice $p=9$ for XLM-R-base (Huang et al., 2022). This design operationalizes the common hypothesis that lower multilingual layers encode more language-specific signals while upper layers are more semantically aligned.

The second regime is translated or language-specific prompting, where prompts are rendered separately for each target language. This includes translated hard prompts and translated verbalizers, but also minimal-translation approaches in which only labels or class descriptions are translated (Zhou et al., 2022, Chen et al., 2022). Work on relation classification is representative: prompts are kept language-agnostic at the template level by using the sentence and entity spans directly, while only the relation verbalizers are translated as needed (Chen et al., 2022). Legal judgement prompting follows a similar pattern: the structural template remains fixed across English, German, French, and Italian, while the legal question itself is adapted to the court and language (Trautmann et al., 2022).

The third regime is prompt transformation, in which a source-language prompt is mapped into a multilingual or target-appropriate prompt rather than copied directly. Multilingual Prompt Translator exemplifies this approach. It learns a source prompt $\boldsymbol{p}^S$ on English NLI data and passes it through a two-layer MLP translator to produce $\boldsymbol{p}^T$, trained jointly with a symmetric KLD alignment objective on a small auxiliary parallel corpus derived from PAWS-X (Qiu et al., 2024). The resulting target prompt is not language-specific per target language but a multilingual prompt intended to retain task knowledge while altering language knowledge (Qiu et al., 2024).

The fourth regime is language-specific dynamic prompting, where prompts are learned independently per language and selected at inference time. PolyPrompt follows this strategy for multilingual MMLU: it learns five continuous trigger embeddings $T^\lambda \in \mathbb{R}^{k \times d}$ for each language, detects the input language with langid, and prepends the corresponding trigger tokens to the input of a frozen multilingual LLaMA model (Roll, 27 Feb 2025). This does not attempt to force one universal prompt; instead it treats language-conditioned prompt selection as an efficient form of PEFT.

The fifth regime is parallel multilingual prompting, in which several language realizations of the same prompt are presented together. In text-to-image generation, PMT2I concatenates English with translations into several high-resource languages and conditions a multimodal generator on the resulting multilingual bundle (Mu et al., 13 Jan 2025). In diversity-oriented LLM prompting, multilingual prompting creates several culturally cued prompt variants in different languages, generates separate responses, and aggregates them by concatenation (Wang et al., 21 May 2025). This pattern differs from language routing: instead of choosing one language, it exploits redundancy and complementarity across several languages simultaneously.

The following table summarizes the major regimes.

Regime	Representative method	Core mechanism
Shared unified prompt	UniPrompt (Huang et al., 2022)	One language-agnostic prompt encoded and fused in multilingual semantic space
Translated/language-specific prompt	Universal Prompting, multilingual RC (Zhou et al., 2022, Chen et al., 2022)	Prompt components translated or minimized per language
Prompt transformation	MPT (Qiu et al., 2024)	Source prompt mapped to multilingual target prompt via learned translator
Dynamic language-routed prompt	PolyPrompt (Roll, 27 Feb 2025)	One learned soft-trigger set per language, selected at inference
Parallel multilingual prompt bundle	PMT2I, multilingual diversity prompting (Mu et al., 13 Jan 2025, Wang et al., 21 May 2025)	Multiple language versions used together and aggregated

These regimes are not mutually exclusive. This suggests that multilingual prompt exploration is best viewed as a design space rather than a single method family.

3. Training–inference alignment and prompt semantics

A central theme across the literature is alignment between pretraining, prompt representation, and inference conditions. The strongest early articulation appears in work on Universal Prompting and Dual Prompt Augmentation. That work observes that cross-lingual prompting often fails because the model is trained with English templates and English verbalizers but tested with translated templates and translated label words, creating a source–target discrepancy in both the prompt scaffold and the prediction targets (Zhou et al., 2022). Universal Prompting removes template words entirely and keeps English verbalizers at both training and inference, using a language-agnostic structure like A . B ? <mask> . (Zhou et al., 2022). On XNLI, this yields substantial gains over translated templates and verbalizers, such as 43.18 vs. 38.81 average accuracy at 16 shots and 48.26 vs. 41.42 at 32 shots (Zhou et al., 2022).

UniPrompt addresses the same alignment problem from an architectural rather than purely lexical perspective. Instead of concatenating prompt tokens with the input at the embedding layer, it encodes the prompt and the context separately through duplicated lower layers and fuses them in shared upper layers (Huang et al., 2022). Because the Template Tower is initialized from a multilingual PLM, the prompt representation is anchored in the same multilingual space as the input. The paper also shows that multilingual XLM-R initialization for the Template Tower outperforms monolingual RoBERTa initialization, with averages of approximately 40.97 for XLM-R initialization versus 37.38 for RoBERTa-base and 39.08 for random initialization (Huang et al., 2022). This is direct evidence that multilingual prompt exploration is sensitive to the representation manifold in which prompts live.

Label semantics constitute a second alignment problem. UniPrompt introduces soft label words and initializes them not randomly but as the average of [mask] representations for training examples in a class, computed by the original pretrained model $\theta^0$ (Huang et al., 2022). In ablations, this proposed initialization improves average transfer performance to approximately 40.97, compared with 37.19 for discrete label words, 38.85 for random soft labels, and 39.82 for standard LM-head initialization (Huang et al., 2022). Dual Prompt Augmentation reaches a similar conclusion from a different angle: label words should be aligned across languages during training via a multilingual verbalizer rather than introduced only at inference (Zhou et al., 2022).

A third alignment issue concerns reasoning language. Cross-lingual exploration for parametric knowledge shows that simply increasing monolingual inference in the query language is less effective than routing the same query through other languages, particularly when those languages are culturally or geographically closer to the fact’s origin (Diskind et al., 23 Jun 2026). The paper formalizes multilingual prompt exploration along four dimensions—language selection, exploration routing, answer aggregation, and inference budget—and shows that cross-lingual exploration dominates native-language scaling on the compute–accuracy Pareto frontier (Diskind et al., 23 Jun 2026). A plausible implication is that prompt alignment in multilingual systems includes not just prompt form and label space, but also the latent retrieval path induced by the choice of intermediate language.

4. Empirical performance across tasks

Empirical results vary by task, but several broad patterns recur. On multilingual sentiment and text classification, shared or minimally language-dependent prompts perform well under few-shot transfer. UniPrompt on MARC, trained only on English and tested zero-shot on German, Spanish, French, Japanese, and Chinese, achieves the best average performance for almost all $k$-shot settings. At $k=16$, its average is 40.97 compared with 37.87 for vanilla fine-tuning, 37.30 for English Prompt, 35.24 for Soft Prompt, and 32.43 for Translation Prompt (Huang et al., 2022). At $k=32$, UniPrompt reaches 46.61 versus 42.59 for vanilla fine-tuning and 40.11 for English Prompt (Huang et al., 2022). These improvements are strongest in moderate few-shot regimes, where prompt transferability matters most.

On cross-lingual NLI, prompt transformation and language-agnostic prompting both outperform conventional baselines. Dual Prompt Augmentation reaches 46.54 average accuracy on XNLI with only 16 English training examples per class, compared with 34.99 for fine-tuning and 43.18 for Universal Prompting (Zhou et al., 2022). MPT improves further in few-shot settings: on XNLI 4-shot, average accuracy is 43.4 for MPT, compared with 38.3 for soft prompting, 35.9 for PCT, and 34.1 for full fine-tuning (Qiu et al., 2024). At 256-shot, MPT reaches 59.7, compared with 56.0 for PCT and 51.9 for SP (Qiu et al., 2024). The gains are especially large in typologically distant and low-resource target languages such as Chinese, Swahili, and Urdu, where MPT reports relative gains of about 18.4% for Chinese over SP at $k=32$, and +14.8% and +13.8% for Swahili and Urdu respectively (Qiu et al., 2024).

On multilingual relation classification, efficient prompt construction from relation triples yields strong gains over fine-tuning and task-agnostic null prompts. On SMiLER, in fully supervised conditions, in-language prompting reaches a macro average of 85.0, compared with 84.1 for code-switch prompting, 82.7 for soft prompt, 68.2 for XLM-R_EM fine-tuning, and 66.2 for null prompts (Chen et al., 2022). In few-shot settings, in-language prompting reaches 40.6, 46.3, and 65.5 macro average at 8, 16, and 32 shots, while code-switch prompting is best at 16 shots with 48.9 and close at 32 shots with 60.8 (Chen et al., 2022). These results show that prompt engineering with minimal translation can scale beyond small-label classification tasks.

On dialogue generation, prompt learning is used not primarily to improve task performance directly but to preserve multilinguality during fine-tuning. Prompt Learning to Mitigate Catastrophic Forgetting in Cross-lingual Transfer for Open-domain Dialogue Generation documents that few-shot cross-lingual transfer and multitask learning both cause catastrophic forgetting in all six target languages unless fine-tuning is made more pretraining-compatible through fixed-prompt LM tuning with handcrafted prompts (Liu et al., 2023). This suggests that prompt exploration can serve a regularization function in multilingual generative systems, not only an instruction or control function.

On legal judgement prediction, zero-shot legal prompt engineering beats random and majority baselines across English, German, French, and Italian, despite no legal-domain fine-tuning. On ECHR, GPT-J-6B reaches macro-F1 0.528 and GPT-NeoX-20B 0.521, compared with 0.459 for the majority baseline, though still below the supervised state of the art at 0.820 (Trautmann et al., 2022). On the Swiss Federal Supreme Court corpus, mGPT-xl reaches macro-F1 0.493 for German, 0.502 for French, and 0.484 for Italian, compared with baselines around 0.447–0.449 and supervised results of 0.685, 0.702, and 0.598 respectively (Trautmann et al., 2022). This indicates that multilingual prompting can transfer even into high-stakes domain-specific settings, albeit with clear limitations.

On recommendation, multilingual prompts expose a different failure mode: prompt language itself can reduce performance when the recommender has been tuned only on English prompt templates. In OpenP5-style recommendation, Spanish and Turkish prompts generally reduce performance relative to English, particularly for Turkish and for sequential recommendation under sequential indexing (Ozsoy, 2024). Multilingual retraining balances this disparity but slightly reduces English performance (Ozsoy, 2024). A plausible implication is that prompt exploration in multilingual settings can reveal whether a system is genuinely multilingual or merely English-conditioned with superficial cross-lingual transfer.

5. Multilingual prompts for diversity, multimodality, and knowledge access

Multilingual prompt exploration is not limited to standard classification or generation metrics. Several recent works use multilingual prompts to alter what the model knows, notices, or produces.

One line of work targets knowledge access and factual recall. Cross-Lingual Exploration for Parametric Knowledge argues that parametric knowledge is not equally accessible across languages and that multilingual prompt exploration can unlock latent factual knowledge more efficiently than additional monolingual inference (Diskind et al., 23 Jun 2026). The method evaluates single-path and multi-path strategies on 17 languages over ECLeKTic and CLIKE. On CLIKE, native monolingual prompting yields about 54.8% macro accuracy for a Gemini 2.5 Flash example, while origin-aware single-path language selection reaches about 70.0%, fixed-set multi-path majority vote 72.4%, and repeated autonomous selection with majority vote 74.4% (Diskind et al., 23 Jun 2026). On ECLeKTic, native baseline is about 38.4% macro, while origin-aware selection reaches 53.1% and repeated autonomous selection with minority-aware aggregation about 59.3% (Diskind et al., 23 Jun 2026). The paper further shows that consistency gains exceed what can be explained by accuracy gains alone (Diskind et al., 23 Jun 2026).

A second line addresses generation diversity and cultural plurality. Multilingual Prompting for Improving LLM Generation Diversity constructs culturally cued prompt variants in several languages, generates responses separately, and concatenates them. Across GPT-4o, GPT-4o-mini, LLaMA 70B, and LLaMA 8B, multilingual prompting outperforms high-temperature sampling, personas prompting, and step-by-step recall on reason entropy, agreement entropy, and demographic entropy (Wang et al., 21 May 2025). The paper reports that multilingual prompting yields 1.8x–2.38x higher reason entropy, 1.65x–2.86x higher agreement entropy, and 1.1x–1.2x higher demographic entropy than the best non-multilingual diversity baselines (Wang et al., 21 May 2025). It also finds that aligning prompt language with cultural cues reduces hallucination about culturally specific entities, with validity rates for Chinese names rising from 73.3% to 87.6% for the basic setting and from 81.0% to 92.3% for the enhanced setting (Wang et al., 21 May 2025).

A third line uses multilingual prompting in multimodal generation. PMT2I augments text-to-image prompts by concatenating the English prompt with translations into several languages, tagged by language labels, and feeds the whole bundle into large multimodal models (Mu et al., 13 Jan 2025). On MS-COCO, DrawBench, and T2I-CompBench, PMT2I improves CLIP-T, CLIP-I, DINO, and ImageReward compared with English prompts, especially when combined with CLIP-based reranking (Mu et al., 13 Jan 2025). For example, on Emu2-Gen over DrawBench, EN Prompt achieves 38.5% GPT-4o correctness while PMT2I-6 reaches 44.3% and PMT2I-6 + rerank 45.8% (Mu et al., 13 Jan 2025). On Lumina-Next over T2I-CompBench, PMT2I-6 + rerank yields B-VQA improvements such as 55.2% on Color and 50.9% on Texture (Mu et al., 13 Jan 2025). The paper’s interpretation is that multilingual bundles provide semantically redundant yet complementary conditioning signals.

A fourth line appears in speech generation, where prompt exploration becomes prompt retrieval rather than instruction design. M$p=9$0PDB is a multimodal, multi-label, multilingual prompt database for speech generation with 18 languages, over 400k hours of data, and unified label taxonomies across modalities (Zhu et al., 13 Aug 2025). Although this work is not about token prompts in the NLP sense, it extends the notion of multilingual prompt exploration to speech prompt selection, cross-lingual proxying for unseen languages, and latency-aware prompt routing (Zhu et al., 13 Aug 2025). This suggests that the conceptual frame of multilingual prompt exploration is broader than text-only prompting and increasingly intersects with retrieval and selection problems.

6. System-prompt steerability and prompt optimization

A more recent strand treats multilingual prompt exploration explicitly as optimization over prompt components. Cross-Lingual Prompt Steerability studies system prompts that must work reliably across five languages, three LLMs, and three benchmarks, and proposes a four-dimensional evaluation framework: mean accuracy, accuracy variance across languages, cross-lingual consistency, and output-length variance (Zhang et al., 2 Dec 2025). It then constructs a prompt component library of 10,000 items across categories such as role, style, emotion, scenario, behavioral instructions, CoT, safety, and cross-language instructions (Zhang et al., 2 Dec 2025).

The empirical findings are striking. Same-language system prompts slightly outperform English prompts on average, with $p=9$1 0.5097 versus 0.5033 and lower variance and length variance, though the differences are modest (Zhang et al., 2 Dec 2025). More importantly, regression over prompt components shows that Chain-of-Thought, emotion, and scenario components correlate positively with multilingual robustness, while cross-language components, behavioral instructions, role, and style correlate negatively with accuracy and consistency (Zhang et al., 2 Dec 2025). Cross-language system directives such as “translate the question into multiple languages” are particularly harmful, with coefficients around $p=9$2 for mean accuracy and $p=9$3 for consistency (Zhang et al., 2 Dec 2025).

The same paper adapts the Sprig optimizer to the multilingual setting using a reward model trained on prompt–metric pairs. Optimized prompts improve all metrics by roughly 5–10% over random prompts. For Qwen2.5-7B-Instruct, $p=9$4 rises from 0.521 to 0.621, consistency from 0.254 to 0.311, and length variance falls dramatically from 344,154 to 90,419 (Zhang et al., 2 Dec 2025). The reasoning-trace analysis over more than 10 million reasoning units shows that better prompts induce more structured reasoning and reduce unnecessary language switching (Zhang et al., 2 Dec 2025).

This optimization perspective complements task-specific prompt methods. It shifts attention from handcrafted prompt content to prompt component effects that generalize across tasks and models. A plausible implication is that multilingual prompt exploration may benefit from separating task-language alignment from reasoning-language intervention: the former tends to help when it keeps prompt form close to user language, while explicit cross-language manipulation inside the system prompt can be destabilizing unless carefully structured, as in dedicated cross-lingual exploration methods (Zhang et al., 2 Dec 2025, Diskind et al., 23 Jun 2026).

7. Recurrent controversies, misconceptions, and limitations

Several misconceptions recur in this area. One is the belief that translating an English prompt is sufficient. Multiple papers show that this is often false. Translation Prompt baselines are fragile on classification tasks because translated templates and verbalizers create training–inference mismatch and can encode unnatural syntax in the target language (Huang et al., 2022, Zhou et al., 2022). In NLI, translated hard prompts underperform unified or transformed prompt strategies (Qiu et al., 2024). In recommendation, directly translated prompts reduce performance unless the model is retrained on multilingual prompt variants (Ozsoy, 2024). In function calling, naïvely translating JSON tool schemas corrupts function names and argument structures; the successful solution is to translate user-facing text while preserving function schemas in a canonical language (Chen et al., 2024).

A second misconception is that one multilingual prompt should always outperform language-specific prompts. The evidence is mixed. UniPrompt and Universal Prompting show that one shared prompt can work very well, especially when prompt representations are language-agnostic and aligned to multilingual PLM space (Huang et al., 2022, Zhou et al., 2022). However, PolyPrompt demonstrates strong gains from per-language trigger tokens, and MPT shows that a learned transformation of a source prompt can outperform a shared reused soft prompt (Roll, 27 Feb 2025, Qiu et al., 2024). This suggests that universality is not free; it depends on architecture, task, and whether the prompt is supposed to encode task knowledge, language knowledge, or both.

A third misconception is that multilingual prompting only matters for low-resource languages. While gains are often largest there, several studies show improvements even for well-resourced languages. PolyPrompt improves English, Spanish, French, German, and Italian markedly on Global MMLU, with English rising from 37.1 to 43.4 for Llama 3.2 1B Base and from 43.9 to 49.5 for the Instruct variant after one epoch (Roll, 27 Feb 2025). PMT2I improves text-to-image generation even though English is already strong (Mu et al., 13 Jan 2025). Multilingual diversity prompting finds that high-resource languages can still unlock underused cultural knowledge when used in parallel (Wang et al., 21 May 2025).

A fourth issue is evaluation scope. Many methods are validated on a narrow set of tasks, often NLI, sentiment, or QA. The survey literature notes that multilingual prompt engineering spans around 250 languages across 36 papers and 39 techniques, but sophisticated methods are concentrated in a few language families and task types (Vatsal et al., 16 May 2025). Low-resource languages are often covered only by translation-based baselines or in machine translation benchmarks (Vatsal et al., 16 May 2025). Likewise, prompt optimization studies often use only five languages and a handful of benchmarks (Zhang et al., 2 Dec 2025). This suggests that claims about “multilingual prompts” should be interpreted as conditioned on the task, model family, and language coverage actually studied.

Finally, there is the limitation of catastrophic forgetting and interference. Dialogue generation work shows that mPLMs can lose multilingual generation ability during fine-tuning unless prompting is made compatible with pretraining (Liu et al., 2023). Recommendation work shows that multilingual retraining can balance languages only at some cost to English performance (Ozsoy, 2024). These results caution against treating multilingual prompt performance as a pure inference phenomenon; training dynamics can erase the multilingual substrate on which prompting depends.

8. Broader significance and future directions

Multilingual prompt exploration has become a unifying framework for studying how language choice alters access to model capabilities. Early work emphasized practical burdens: per-language prompt engineering is expensive, translated prompts are brittle, and prompt–label mismatch impairs cross-lingual transfer (Huang et al., 2022, Zhou et al., 2022). More recent work has reframed the topic in terms of model-internal multilingual geometry, prompt routing, and component-level steerability (Zhang et al., 2 Dec 2025, Diskind et al., 23 Jun 2026). Across these perspectives, several stable conclusions have emerged.

First, prompt form is part of the multilingual transfer mechanism, not a neutral wrapper. Whether through shared model-based encoders, multilingual verbalizers, learned translators, or language-conditioned trigger tokens, effective methods reduce mismatch between prompt representation and multilingual pretraining space (Huang et al., 2022, Qiu et al., 2024, Roll, 27 Feb 2025).

Second, language can be used as an inference-time search variable. Parallel multilingual prompting, autonomous language selection, and cross-lingual exploration for factual recall all show that models often encode knowledge or behaviors that are accessible only through particular linguistic paths (Mu et al., 13 Jan 2025, Wang et al., 21 May 2025, Diskind et al., 23 Jun 2026). This suggests a conceptual shift from asking “what is the best prompt?” to asking “which language or combination of languages best exposes the required capability?”

Third, shared multilingual prompts and language-specific prompts solve different problems. Shared prompts reduce operational complexity and human cost; language-specific prompts or transformed prompts can better capture language-specific geometry or cultural priors (Huang et al., 2022, Qiu et al., 2024, Roll, 27 Feb 2025). Future systems will likely combine both, using a shared backbone prompt with lightweight language-conditioned modules or routing.

Fourth, prompt evaluation in multilingual settings needs multidimensional metrics. Accuracy alone misses language variance, cross-lingual consistency, hallucination about culturally specific entities, and reasoning-language stability (Zhang et al., 2 Dec 2025, Wang et al., 21 May 2025, Diskind et al., 23 Jun 2026). A plausible implication is that multilingual prompt exploration should be benchmarked as a robustness problem rather than only as a task-performance problem.

Several future directions are already visible in the literature: extension to generation and multimodal tasks beyond classification (Huang et al., 2022, Mu et al., 13 Jan 2025); per-language or typology-aware prompt translators (Qiu et al., 2024); shared prompt spaces augmented with adapters or PEFT mechanisms (Roll, 27 Feb 2025, Huang et al., 2022); larger-scale prompt optimization across more languages and tasks (Zhang et al., 2 Dec 2025); and cross-lingual exploration methods that integrate retrieval, tool use, or multimodal evidence (Diskind et al., 23 Jun 2026, Chen et al., 2024, Zhu et al., 13 Aug 2025). Taken together, these developments indicate that multilingual prompt exploration is evolving from a narrow question of prompt translation into a general methodology for controlling, auditing, and amplifying multilingual model behavior.