The importance of prompt design for LLM performance has been extensively documented since the introduction of GPT-3 [1], which demonstrated that few-shot prompting could rival fine-tuning on many benchmarks. Subsequent work formalized the study of prompt sensitivity, showing that even semantically equivalent reformulations of prompts could produce dramatically different results [7], [22]. This observation motivated a transition from manual prompt crafting toward systematic optimization approaches.

Early automated methods include prefix tuning [23] and prompt tuning [24], which optimize continuous prompt embeddings via gradient descent. While effective, these soft-prompting methods require access to model parameters and produce non-interpretable prompts, limiting their applicability to black-box API settings and their transferability across models. In contrast, discrete prompt optimization methods operate directly on natural language tokens. APE [18] uses an LLM to generate candidate instructions given input-output demonstrations and selects the highest-performing one. OPRO [19] maintains an optimization trajectory of prompt-score pairs and instructs the LLM to propose improved prompts based on this history. ProTeGi [25] employs gradient-inspired textual feedback to iteratively refine prompts. PromptBreeder [26] introduces a self-referential evolutionary approach where mutation operators themselves evolve alongside the prompts they generate.

Despite these advances, existing prompt optimization methods overwhelmingly target English-language tasks and assume the availability of substantial development sets for evaluation. Our work extends this line of research to the low-resource setting, where both assumptions are violated.

The disparity in NLP resources across languages has been a longstanding concern. Joshi et al. [12] categorized the world's languages into six classes based on available resources and community engagement, revealing that the vast majority of languages, including many with millions of speakers, fall into the lowest resource tiers. This taxonomy has served as a foundational reference for subsequent low-resource NLP research.

Traditional approaches to low-resource NLP have focused on data augmentation [27], cross-lingual transfer from high-resource languages [28], [29], and multilingual pretraining [15], [30]. More recently, the emergence of multilingual LLMs has provided new avenues for low-resource language processing. Studies have examined the multilingual capabilities of models such as GPT-4 [31], BLOOM [32], and LLaMA-2 [3], finding that while these models exhibit some cross-lingual transfer, their performance on low-resource languages remains significantly below their English-language capabilities [13], [14], [33].

Specialized models targeting specific low-resource languages have also been developed, including AceGPT for Arabic [34], SEA-LION for Southeast Asian languages [35], and AfroLM for African languages [36]. While these models improve baseline performance, they require substantial resources for training and do not address the fundamental challenge of prompt optimization in data-scarce settings.

EAs have a long history of application to NLP tasks, predating the current LLM era. Genetic Algorithms (GAs) [37] have been applied to feature selection for text classification [38], grammar optimization [39], and information extraction rule learning [40]. The seminal work of Goldberg [37] established the theoretical foundations for GAs, including schema theory and the building block hypothesis, which inform our hierarchical genome design.

Simulated Annealing (SA) [41] has been employed for combinatorial optimization in NLP contexts, including word sense disambiguation [42], text summarization [43], and machine translation decoding [44]. The combination of GA and SA, hybridizing the population-based exploration of GAs with the controlled acceptance of inferior solutions characteristic of SA, has shown particular promise for avoiding local optima in rugged fitness landscapes [45], [46].

Most recently, EvoPrompt [20] demonstrated the effectiveness of connecting EAs with LLMs for prompt optimization, achieving state-of-the-art results across 31 datasets. However, EvoPrompt operates exclusively on instruction-level prompts in English and does not address the hierarchical structure of modern prompts (which include examples, reasoning scaffolds, and format specifications) or the unique challenges of low-resource settings. Our work substantially extends this paradigm.

Cross-lingual transfer learning aims to leverage resources from high-resource languages to improve performance on low-resource targets. Foundational work includes multilingual word embeddings [47], cross-lingual model adaptation [48], and zero-shot cross-lingual classification [15]. In the context of LLMs, studies have explored cross-lingual prompting strategies, where prompts in one language are used to elicit responses in another [49–51].

Particularly relevant to our work, Shi et al. [51] demonstrated that CoT reasoning can be elicited in multilingual settings by providing English-language CoT prompts even when the input is in another language, suggesting that prompt structure transfers across languages more readily than prompt content. Winata et al. [52] examined cross-lingual in-context learning, finding that the language of demonstrations can be decoupled from the language of the query with varying degrees of success depending on language similarity and model architecture. Our cross-lingual fitness transfer mechanism builds upon these findings by systematically exploiting prompt-structure transferability to reduce the evaluation burden on low-resource languages.

Let M denote a pretrained language model and T = {(x , y )} n denote a small labeled dataset in a low-resource

i i i=1

target language Lt, where n is typically in the range of 50–200 examples. A prompt P is a structured natural language input that, when concatenated with a test instance x, elicits a response ŷ = M(P ⊕ x) from the model. The objective is to find the optimal prompt P* that maximizes task performance:

P* = arg maxP ∈ (x,y) ∼ T [ f(M(P ⊕ x), y)], (1)

where denotes the space of all valid prompts, f(0,0) is a task-specific evaluation metric (e.g., accuracy, F1 score, ROUGE), and ⊕ denotes the concatenation operation. The key challenge in the low-resource setting is that is combinatorially large, n is small (making fitness estimates noisy), and the model M has limited competence in Lt.

We decompose the prompt P into four structural components:

P = (Pinst, Pdemo, Pcot, Pfmt), (2)

where Pinst is the instruction template (task description), Pdemo is the in-context demonstration configuration (selection and ordering of K examples from a seed pool S), Pcot is the CoT scaffold (reasoning template with variable decomposition depth), and Pfmt is the output format specification. This decomposition enables independent optimization of each component while preserving the interdependencies through the fitness evaluation.

We represent each prompt candidate as a four-level hierarchical genome G = (g1, g2, g3, g4), where each level corresponds to one of the four structural components defined in Eq. (2).

Level 1 (instruction template gene (g1)). The instruction template gene encodes the task description using a combinatorial template grammar. We define a context-free grammar Ginst = (V, Σ, R, S) where V is the set of non-terminal symbols representing structural slots (e.g., ⟨task_verb⟩, ⟨domain_context⟩, ⟨constraint⟩), Σ is the set of terminal symbols (language-specific lexical items), R is the set of production rules, and S is the start symbol. The gene g1 is a vector of production rule indices:

g1 = (r1, r2, ..., rm), where rj ∈ R(vj). (3)

This encoding ensures that all generated instructions are syntactically valid while enabling combinatorial exploration of phrasing alternatives. For example, the non-terminal ⟨task_verb⟩ for sentiment analysis may expand to (classify, determine, identify, analyze, assess), producing semantically similar but structurally distinct instruction variants.

Level 2 (demonstration configuration gene (g2)). The demonstration gene encodes which K examples are selected from a seed pool S = {s1, ..., sN} and in what order they are presented. We represent this as a permutation-selection pair:

g2 = (σ, K), where σ ∈ Perm(S), K ∈ {0, 1, ..., Kmax}. (4)

The first K elements of the permutation σ are selected as in-context demonstrations, preserving their order. This encoding naturally captures both selection and ordering decisions in a single genetic representation. We set Kmax = 8 in our experiments, consistent with the context length constraints of the base models.

Level 3 (CoT scaffold gene (g3)). The CoT scaffold gene encodes a reasoning template with variable decomposition depth. We represent reasoning chains as Directed Acyclic Graphs (DAGs) where nodes correspond to reasoning steps and edges encode dependencies:

g = (D, {t } D, E), where D ∈ {0, 1, ..., D }. (5)

3 d d=1 max

Here, D is the decomposition depth (number of explicit reasoning steps), td is the template for step d (drawn from a predefined set of reasoning templates such as "First, identify the key entities," "Then, analyze the relationship between," "Consider the context of"), and E ⊆ {1, ..., D} × {1, ..., D} is the set of dependency edges. Setting D = 0 corresponds to direct (non-CoT) prompting. We set Dmax = 5 based on preliminary

The core of EvoPrompt-LR is a hybrid algorithm that combines the population-based search of GAs [37] with the controlled stochastic acceptance of SA [41]. The GA component maintains a diverse population of prompt candidates and applies genetic operators (selection, crossover, mutation) to explore the search space, while the SA component governs the acceptance of offspring into the population, allowing temporary acceptance of inferior solutions to escape local optima.

The fitness of a prompt genome G is defined as

F(G) = α · ftask(G) + β · frobust(G) + γ · flength(G), (7)

where ftask is the primary task metric (accuracy or F1), frobust measures performance stability across multiple evaluation runs (inverse of standard deviation), flength penalizes excessively long prompts to remain within context length budgets, and α, β, γ are weighting hyperparameters with α = 0.8, β = 0.15, γ = 0.05 in our

A central challenge in optimizing prompts for low-resource languages is the cost of fitness evaluation: Each candidate prompt must be applied to multiple test instances in the target language, consuming both computational resources and scarce labeled data. We address this through a cross-lingual fitness transfer mechanism that exploits the observation that prompt structure transfers across languages more robustly than prompt content [51], [52].

Let Lp denote a high-resource proxy language (English or Chinese in our experiments) with an abundant evaluation set Tp. For a candidate genome G, we define the proxy fitness as

F̃ (G) = F(Translate(G, Lt → Lp); Tp, M), (10)

where Translate(G, Lt → Lp) translates the textual components of the genome (instruction template, CoT steps, format specifications) from the target language to the proxy language while preserving the structural components (example selection indices, ordering, depth parameters). Translation is performed using NLLB-200 [54] for all language pairs.

The cross-lingual fitness transfer operates in two phases, as detailed in Algorithm 2.

Algorithm 2. Cross-lingual fitness transfer.

Input:

Offspring set

, proxy language

, target language

model

, top-

ratio

= 0.2 Output:

Offspring with fitness scores // Phase 1: Proxy evaluation (all candidates)

for each do

proxy ← Translate(

) ← Evaluate(

proxy

end for // Phase 2: Target evaluation (top-k candidates only)

top ← Top-⌈

|⌉ by

for each

top do

) ← Evaluate(

// Direct low-resource evaluation

end for // Assign estimated fitness to remaining candidates for each

top do

) ← λ ·

// Discounted proxy fitness, λ = 0.85

end for

return

with fitness scores

By evaluating only the top 20% of candidates on the actual low-resource language, we reduce the number of expensive low-resource evaluations by approximately 80% per generation. The discount factor λ = 0.85 accounts for the imperfect correlation between proxy and target fitness, ensuring that candidates not directly evaluated on the target language are not overvalued.

To calibrate the proxy-target correlation, we perform a one-time calibration step at initialization where a subset of 30 randomly generated prompts are evaluated on both Lp and Lt. We compute the Spearman rank correlation ρ between the two fitness rankings and adjust the discount factor accordingly: λ = 0.7 + 0.2 · ρ. In our experiments, ρ ranges from 0.62 (Khmer–English) to 0.81 (Arabic–English), reflecting the varying degrees of structural similarity between proxy and target languages.

A distinctive feature of EvoPrompt-LR is the incorporation of mutation operators tailored to the typological characteristics of the target language. Standard mutation operators that randomly substitute words or phrases are linguistically naïve and frequently produce ungrammatical or semantically incoherent prompts, particularly in morphologically rich languages. We design three families of mutation operators corresponding to major typological classes:

Agglutinative language operators (Vietnamese, Uzbek): In agglutinative languages, words are formed by concatenating morphemes, each carrying a single meaning or grammatical function. Mutation operators for these languages operate at the morpheme level rather than the word level. Specifically, our morpheme-aware substitution operator identifies the morphological decomposition of instruction keywords using the UniMorph database [55] and substitutes individual morphemes while preserving the overall morphological structure. For example, in Uzbek, the verb stem yoz- (write) may be mutated to belgi- (mark) while preserving the suffix chain -ish-ing-iz (reciprocal-present-2PL.formal). For Vietnamese, which is analytic but uses classifier systems and serial verb constructions, our operator respects classifier-noun agreement and serial verb ordering constraints.

Fusional language operators (Arabic): Arabic exhibits root-and-pattern morphology where consonantal roots are interleaved with vocalic patterns to produce word forms. Our root-preserving mutation operator maintains the triconsonantal root while varying the morphological pattern. For instance, the root k-t-b (related to writing) may be realized as kataba (he wrote), kitāb (book), kātib (writer), or maktūb (written), allowing semantically related variations that remain morphologically well-formed. Additionally, our operator handles Arabic dialectal variation by maintaining a mapping between Modern Standard Arabic (MSA) forms and their dialectal equivalents for Gulf, Egyptian, and Maghrebi Arabic.

Isolating language operators (Khmer): As a predominantly isolating language, Khmer has minimal morphological inflection but exhibits complex syntactic constraints, including specific word ordering, particle use, and register-dependent vocabulary. Our operator for Khmer focuses on syntactic slot permutation (reordering of adverbial modifiers, topic-comment restructuring) and register-appropriate lexical substitution, drawing on a curated register lexicon of formal/informal variant pairs.

For Swahili and Yoruba, which exhibit elements of agglutination and tonal marking respectively, we implement hybrid operators that combine morpheme-level substitution with tone-preserving constraints (for Yoruba) and noun class agreement preservation (for Swahili).

To prevent premature convergence to a narrow region of the prompt space, a common failure mode in evolutionary optimization, we incorporate a semantic diversity preservation mechanism inspired by the crowding distance used in NSGA-II [21].

We compute the semantic embedding of each prompt genome G by encoding its rendered prompt text using a multilingual sentence encoder (LaBSE) [56]. Let e(G) ∈ ℝd denote the embedding. For each individual in the population, the crowding distance is computed as

CD(G ) = Σ ‖e(G ) − e(G )‖ , (11)

i j ∈ NNk(i) i j 2

where NNk(i) denotes the k-nearest neighbors of individual i in the embedding space (we use k = 3). During population trimming (when the combined parent-offspring population exceeds Npop), individuals with lower fitness are eliminated, but ties are broken in favor of higher crowding distance, preserving diverse prompt structures.

Additionally, we implement a diversity injection mechanism: Every δ generations (we use δ = 10), we replace the bottom 10% of the population (by fitness) with freshly generated random genomes. This mechanism ensures continuous exploration of unexplored regions of the prompt space and has been shown to improve long-run convergence in similar evolutionary optimization contexts [57].

We evaluate EvoPrompt-LR on six languages selected to represent diverse typological families, geographic regions, and resource availability levels, following the taxonomy of Joshi et al. [12]. Table 1 summarizes the language characteristics and datasets used.

For the seed pool used in prompt optimization, we sample 100 labeled examples per task per language from the respective training sets. The evaluation is performed on held-out test sets of 500 examples (or the full test set if smaller). For languages where existing benchmarks are unavailable (Uzbek NER, Uzbek QA, Khmer sentiment, Khmer NER, Khmer QA), we construct custom datasets through expert annotation, following established annotation guidelines [65], [66]. Custom datasets undergo two rounds of annotation by native speakers with inter-annotator agreement (Cohen's κ) exceeding 0.80 for all tasks.

We evaluate EvoPrompt-LR with three base LLMs chosen to represent different model families and linguistic specializations:

I. LLaMA-3-8B [4]: A general-purpose open-source model with 8 billion parameters, pretrained on over 15 trillion tokens of multilingual data. While primarily English-centric, LLaMA-3 includes non-trivial coverage of Arabic, Vietnamese, and Swahili.

II. Mistral-7B [5]: A 7-billion-parameter model employing grouped-query attention and sliding window attention mechanisms. Known for strong performance-to-parameter efficiency, it serves as a representative of efficient LLM architectures.

III. AceGPT-7B [34]: A 7-billion-parameter model specifically fine-tuned for Arabic using continued pretraining on Arabic corpora and reinforcement learning with culturally aligned feedback. Included to assess the interaction between language-specialized models and prompt optimization.

All models are used in their instruction-tuned variants. Inference is performed with greedy decoding (temperature = 0) for classification tasks and nucleus sampling (top-p = 0.9, temperature = 0.3) for generation tasks. The maximum output length is set to 512 tokens for QA and summarization tasks and 64 tokens for sentiment and NER tasks.

We compare EvoPrompt-LR against six baseline methods spanning manual and automated prompt optimization approaches:

I. Zero-shot: Direct task instruction without demonstrations, using the best manually crafted prompt identified through preliminary experiments.

II. Few-shot (Standard): Fixed set of K = 5 randomly selected demonstrations prepended to the task instruction, using the same instruction as zero-shot.

III. Manual prompting: Task-specific prompts hand-crafted by native speakers of each target language with NLP expertise, following best practices from the prompt engineering literature [6], [69].

IV. Automatic prompt engineer [18]: APE, which generates candidate instructions via an LLM and selects the best-performing one. We use the iterative refinement variant.

V. Optimization by prompting [19]: OPRO, which uses an LLM to iteratively propose improved instructions based on an optimization trajectory. We use LLaMA-3-8B as the optimizer LLM.

VI. Evolutionary Prompt Optimization [20]: EvoPrompt uses GA and differential evolution operators applied to instruction-level prompts, as described in the original implementation.

Additionally, we compare against SFT using LoRA [70] with the full available training data for each dataset (typically 5,000–50,000 labeled examples) to establish an upper bound reference.

We adopt task-specific evaluation metrics consistent with the conventions of each task:

I. Sentiment analysis: Macro-averaged F1 score.

II. NER: Entity-level F1 score, following the CoNLL evaluation scheme.

III. QA: Exact Match (EM) and token-level F1 score.

IV. Text summarization: ROUGE-L F1 score [71] and BERTScore F1 [72].

All experiments are repeated with three random seeds, and we report the mean ± standard deviation. Statistical significance is assessed using the Wilcoxon signed-rank test at the p < 0.05 level.

EvoPrompt-LR is implemented in Python using the Hugging Face Transformers library [73]. Key hyperparameters for the evolutionary optimization are: Population size Npop = 50, maximum generations Gmax = 100, tournament size k = 3, crossover probability pcross = 0.5, mutation rates μ1 = 0.3, μ2 = 0.2, μ3 = 0.15, μ4 = 0.1, cooling rate η = 0.95, proxy evaluation ratio r = 0.2, diversity injection interval δ = 10. The template grammar for each language is constructed by two native-speaking NLP researchers and contains between 120 (Khmer) and 450 (Arabic) production rules.

Experiments are conducted on a cluster of 8 NVIDIA A100 80GB GPUs. A single optimization run (100 generations × 50 population size) requires approximately 12–18 hours of compute time depending on the language and task, compared to approximately 60–90 hours without the cross-lingual fitness transfer mechanism.

Table 2 presents the sentiment analysis results across all six languages using LLaMA-3-8B as the base model. EvoPrompt-LR consistently achieves the highest F1 scores, with improvements ranging from 8.3% (Vietnamese) to 22.1% (Yoruba) over the best-performing baseline.

Notably, EvoPrompt-LR surpasses SFT with LoRA on Swahili (+1.9%), Yoruba (+2.3%), Uzbek (+1.4%), and Khmer (+4.6%), despite using only 100 labeled examples compared to the 5,000–50,000 examples used for SFT. On Arabic and Vietnamese, where more pretraining data exists, the gap between EvoPrompt-LR and SFT narrows to 2.4% and 3.2%, respectively.

Table 3 shows the NER results. EvoPrompt-LR achieves particularly strong gains on NER, a task where prompt structure (entity type specifications, output format, and demonstration quality) plays a critical role. The improvement over baseline EvoPrompt ranges from 8.8% (Vietnamese) to 10.4% (Khmer). On four of six languages (Swahili, Yoruba, Uzbek, Khmer), EvoPrompt-LR exceeds the SFT upper bound, suggesting that well-optimized prompts can compensate for the model's limited linguistic knowledge more effectively than fine-tuning on noisy low-resource data.

Table 4 reports the QA results, where EvoPrompt-LR achieves an average F1 improvement of 10.1 percentage points over baseline EvoPrompt. The gains are especially pronounced on Yoruba (+10.2% F1) and Khmer (+10.6% F1), confirming the hypothesis that our framework provides the greatest benefit on languages where LLMs have the least pretraining coverage. Notably, EvoPrompt-LR outperforms SFT on four of six languages for QA.

Table 5 presents the summarization results. This task shows the clearest advantage of EvoPrompt-LR's hierarchical genome, as effective summarization prompts require careful coordination between instruction specificity, demonstration selection, CoT scaffolding (for extractive reasoning), and output format constraints. EvoPrompt-LR improves ROUGE-L by 5.8–6.4 points over baseline EvoPrompt across languages and surpasses SFT on all languages except Arabic and Vietnamese.

Table 6 compares EvoPrompt-LR performance across the three base models, averaged over all four tasks.

Several observations emerge from Table 6. First, EvoPrompt-LR provides consistent improvements across all three models, with relative gains of 19–25% over the unoptimized baseline. Second, the language-specialized AceGPT-7B achieves the best Arabic performance when combined with EvoPrompt-LR (68.7%), demonstrating that prompt optimization and model specialization are complementary. Third, LLaMA-3-8B benefits most from EvoPrompt-LR on average, likely due to its broader multilingual pretraining enabling better exploitation of cross-lingual fitness transfer. Fourth, the relative improvement from EvoPrompt-LR is inversely correlated with baseline model performance (Pearson r = −0.73), confirming that prompt optimization is most valuable when model capability is limited.

We conduct comprehensive ablation studies to quantify the contribution of each component of EvoPrompt-LR. Table 7 reports the results, averaged across all six languages and four tasks, using LLaMA-3-8B.

The ablation results reveal that the hierarchical genome encoding is the single most important component, with its removal causing an 8.2-point drop. Within the genome, the demonstration configuration (Level 2) contributes most (−5.8 when removed), followed by the CoT scaffold (Level 3, −4.1), confirming that what examples to show and how to structure reasoning are more impactful than instruction phrasing or output format for low-resource languages. Language-specific mutation operators contribute 4.6 points, validating the importance of linguistically informed search operations. The SA component adds 2.9 points over pure GA, consistent with its role in escaping local optima. Cross-lingual fitness transfer contributes a modest 1.5 points to performance but provides an approximately 5× reduction in computational cost (Section 5.4).

Table 8 reports the computational costs and convergence characteristics of EvoPrompt-LR compared to baselines.

EvoPrompt-LR requires 14.6 GPU-hours per language-task pair, which is higher than simpler baselines but provides a substantially better performance-to-cost ratio. The cross-lingual fitness transfer reduces low-resource evaluations from ~5,000 to ~1,000 (an 80% reduction) while actually improving performance by 1.5 points on average, as the proxy evaluation provides a less noisy fitness signal by leveraging larger evaluation sets in the proxy language.

Convergence analysis reveals that EvoPrompt-LR typically reaches 90% of its final performance within 40 generations and 95% within 60 generations. The SA component is most active during generations 20–50, when the temperature is sufficiently low to be selective but still allows occasional acceptance of inferior solutions. The diversity injection mechanism (every 10 generations) produces observable fitness improvements in approximately 30% of injections, indicating that it successfully introduces beneficial genetic material.

The effectiveness of EvoPrompt-LR can be attributed to several interacting factors. First, the hierarchical genome encoding allows the evolutionary search to explore a structured prompt space rather than the flat, astronomically large space of all possible token sequences. By decomposing prompts into four functionally distinct levels, EvoPrompt-LR enables building-block assembly [37]: good instruction templates can combine with good demonstration configurations and good CoT scaffolds, creating synergistic prompt structures that would be extremely unlikely to arise through random or gradient-free search on a flat representation.

Second, the language-specific mutation operators ensure that the evolutionary search remains within the space of linguistically valid prompts. Our experiments with ablated (language-agnostic) mutation operators revealed that approximately 35% of mutated prompts in Arabic and 28% in Uzbek were grammatically ill-formed, leading to both wasted evaluations and misleading fitness signals. By constraining mutations to respect morphological and syntactic rules, we simultaneously improve search efficiency and fitness estimation quality. Third, the cross-lingual fitness transfer exploits a non-obvious insight: The structural effectiveness of a prompt (e.g., the number and ordering of demonstrations, the depth of CoT scaffolding, the granularity of output format specification) is more language-agnostic than the lexical content. This property allows the proxy evaluation to serve as an effective filter even when the proxy and target languages are typologically distant, as confirmed by the positive Spearman correlations (0.62–0.81) observed in our calibration experiments.

Despite its overall effectiveness, EvoPrompt-LR exhibits limitations in several scenarios. First, on Arabic and Vietnamese, the two highest-resource languages in our evaluation, the gap between EvoPrompt-LR and SFT remains significant for classification tasks (2–4%), suggesting that when sufficient training data exists, parameter-level adaptation captures linguistic knowledge that cannot be fully expressed through prompts alone.

Second, for tasks requiring extensive world knowledge about the low-resource language's culture or geography (e.g., QA questions about local entities in Khmer or Yoruba), no prompt optimization can compensate for the fundamental absence of this knowledge in the model's parameters. In these cases, EvoPrompt-LR often produces prompts that effectively instruct the model to state "I don't know" rather than hallucinate, which improves precision but reduces recall.

Third, the template grammar approach for instruction generation (Level 1) requires manual construction by native-speaking NLP researchers, which itself is a resource-intensive process. While the grammars are reusable across tasks and models, the initial construction cost (estimated at 20–40 person-hours per language) may be prohibitive for languages with very few NLP researchers.

Our results have several important implications for the broader low-resource NLP community. First, the finding that optimized prompts can match or exceed SFT on genuinely low-resource languages (Yoruba, Khmer, Uzbek, Swahili) challenges the prevailing assumption that LLM adaptation necessarily requires labeled data at scale. This finding is particularly significant for the estimated 4,000+ languages classified in Joshi et al.'s [12] lowest resource tiers, where even collecting 100 labeled examples may be feasible but assembling thousands is not.

Second, the effectiveness of cross-lingual fitness transfer suggests a promising research direction in which prompt optimization infrastructure developed for high-resource languages can be systematically repurposed for low-resource targets, reducing the marginal cost of extending NLP capabilities to new languages.

Third, the strong performance of language-specific mutation operators validates the importance of incorporating typological knowledge into NLP systems, even in the era of language-agnostic pretrained models. This finding aligns with recent arguments for typologically informed NLP (Ponti et al., 2019; Bjerva et al., 2019) and suggests that the benefits of typological awareness extend beyond model architecture to the optimization process itself.

Several limitations of this study should be acknowledged. First, our experiments are limited to models in the 7–8 billion parameter range; the effectiveness of EvoPrompt-LR on larger models (e.g., 70B parameters) or smaller models (e.g., 1–3B parameters) remains to be investigated. Second, the custom datasets for Uzbek and Khmer, while carefully annotated, are relatively small and may not fully represent the linguistic diversity of these languages. Third, our cross-lingual fitness transfer currently uses only English and Chinese as proxy languages; incorporating additional proxy languages (e.g., Russian for Uzbek, Thai for Khmer) could further improve transfer quality. Fourth, the computational cost of EvoPrompt-LR, while reduced by the cross-lingual transfer mechanism, remains substantially higher than simple prompting approaches, which may limit its accessibility for resource-constrained research groups.