The restriction to English-language publications was adopted because the review team could reliably assess methodological quality only in English, and multilingual text summarization research reported in international venues is predominantly published in English. This restriction is acknowledged as a potential limitation, particularly because relevant research on low-resource languages may be published in other languages.

Multilingual sequence-to-sequence (seq2seq) models constituted the most frequently reported architecture class, documented in 19 of the 47 reviewed studies. All 19 studies reported using Transformer-based encoder–decoder frameworks as the primary model backbone. Of these, 12 studies employed mBART or its multilingual extension mBART-50 as the base architecture [17]. Five studies (26%) reported using mT5 in encoder–decoder configurations. The remaining 2 studies reported custom multilingual Transformer encoder–decoder architectures trained from scratch on domain-specific multilingual corpora. With respect to input–output language configurations, 14 of the 19 studies (74%) reported monolingual summarization across multiple languages, with each language evaluated independently. The remaining 5 studies reported cross-lingual summarization settings in which the input document language differed from the output summary language. The number of languages addressed per study within this architecture class ranged from 2 to 45. Seventeen of the 19 studies conducted evaluations exclusively on high-resource language pairs, a common practice claimed in [18]. Two studies explicitly reported degradation in cross-lingual generalization when the architecture was applied to low-resource target languages outside the pre-training distribution.

Hybrid architectures were reported in 7 of the 47 reviewed studies (15%). Studies in this class described systems that combined a neural sequence-to-sequence backbone with one or more supplementary structural or retrieval-based modules. Three distinct hybrid configurations were documented across the reviewed literature. Retrieval-augmented summarization architectures were reported in 3 of the 7 studies. These studies reported incorporating a document retrieval component, proposed in [19], that supplied candidate passages to the generation module as auxiliary context, with retrieved passage counts ranging from top-3 to top-10 per input document. Graph-enhanced Transformer configurations were reported in an additional 3 studies (43%). All three studies reported appending graph neural network (GNN) layers or dependency-parse-based graph encoders to the Transformer encoder to capture inter-sentential and cross-lingual structural relationships before decoding [20]. Two of the three graph-based studies applied the architecture to long-document summarization tasks with input lengths exceeding 1,024 tokens. In one study, 14% of the hybrid class reported a knowledge-infused summarization architecture that integrates a multilingual knowledge graph into the encoder via entity-linking and relation-embedding layers [21]. All hybrid architecture studies reported increased computational overhead relative to standard seq2seq baselines as a documented characteristic of their configurations.

Adapter-based and parameter-efficient model architectures were reported in 10 of the 47 reviewed studies (21%). Studies in this class described configurations in which large pretrained multilingual models were adapted for summarization tasks by updating only a restricted subset of parameters, while the base model weights were held frozen [22]. Three parameter-efficient adaptation strategies were identified: adapter module insertion, Low-Rank Adaptation (LoRA), and prefix-tuning. Adapter module insertion was the most frequently documented approach, reported in 5 of the 10 studies (50%). All five studies reported inserting bottleneck feed-forward adapter layers at each Transformer block. The proportion of trainable parameters relative to total model parameters was explicitly reported in 4 of these studies, with values ranging from 0.5% to 3.2%. LoRA-based adaptation was documented in 4 of the 10 studies. These studies described decomposing weight update matrices of Transformer attention layers into low-rank matrix products during fine-tuning. Reported LoRA rank hyperparameters across these studies ranged from r = 4 to r = 16. Prefix-tuning was reported in 1 study, which described prepending a sequence of trainable continuous prompt vectors to the input of each Transformer layer to condition generation on a target-language summarization objective. Across all 10 studies in this class, the reported target-language counts ranged from 3 to 30 languages. Eight of the 10 studies applied these configurations in low-resource language settings, where full fine-tuning was reported to be computationally infeasible.

The reviewed studies show that recent LLM-based summarization approaches increasingly rely on large pretrained generative models combined with different adaptation and prompting strategies. Of the 11 LLM-based studies, five investigated instruction-tuned models in zero-shot or few-shot settings, four used supervised fine-tuning of models such as mBART-large and mT5-XXL, and two examined both zero-shot prompting and supervised approaches. This indicates a shift from conventional multilingual encoder–decoder architectures toward instruction-based and generative approaches that can perform summarization with limited task-specific training. The use of zero-shot and few-shot prompting is particularly relevant to multilingual summarization because it provides a mechanism to extend summarization capabilities to languages with limited annotated training data. However, the reviewed studies also indicate that the effectiveness of these approaches remains dependent on the language coverage and capabilities of the underlying pretrained models.

Evaluation practices in LLM-based studies remain a significant limitation. Eight of the 11 studies reported concerns about hallucinations or the reliability of generated summaries, while six identified limitations associated with the use of ROUGE for evaluating summaries in non-English and morphologically complex languages [23]. Only three studies incorporated human evaluation, indicating that automatic metrics continue to dominate the assessment of LLM-generated multilingual summaries. These findings suggest that although instruction tuning, zero-shot/few-shot prompting, and supervised fine-tuning have expanded the applicability of LLMs to multilingual summarization, current evaluation practices do not yet fully capture factuality, fluency, relevance, and language-specific quality. More robust evaluation frameworks that combine automated metrics with human assessment are therefore needed to reliably assess the performance of LLM-based multilingual summarization systems.

Multilingual pre-training was the most frequently reported technique category, documented in 18 of the 47 reviewed studies (38%). Studies in this category reported employing large-scale pretrained multilingual models as backbone architectures for summarization, leveraging representations acquired during pretraining on multilingual corpora. Thirteen of the 18 studies in this category reported using mBART or mBART-50 as the primary pretrained model. Four studies reported the application of mT5, a multilingual variant of the T5 model pretrained with a text-to-text denoising objective across 101 languages. One study [24] reported using XLM-R as the encoder component in a summarization pipeline. With respect to pre-training objectives, 11 of the 18 studies (61%) reported sequence-to-sequence denoising as the primary pre-training strategy. Four studies reported masked language modeling (MLM) as the backbone pretraining objective, which was subsequently adapted for generation [25]. Three studies reported causal language modeling objectives within instruction-tuned or generative model configurations. Studies in this category further reported that pre-training corpus composition varied across the reviewed models, with some explicitly noting limited coverage of low-resource languages in the pre-training data as a characteristic of the employed backbones. The reported target-language counts per study ranged from 2 to 45 languages in this category.

Cross-lingual transfer learning was reported in 14 of the 47 reviewed studies (30%). Studies in this category described the transfer of summarization capabilities from one or more source languages, predominantly English, to one or more target languages with limited or no target-language training data. Three distinct transfer configurations were observed across the reviewed literature: zero-shot transfer, few-shot transfer, and supervised cross-lingual fine-tuning. Zero-shot cross-lingual transfer was reported in 6 of the 14 studies in this category. These studies reported fine-tuning models exclusively on English-language summarization data and subsequently evaluating the models on non-English target languages without any target-language supervision. Few-shot transfer configurations were reported in 4 studies, in which a small number of target-language training examples, ranging from 50 to 500 instances across studies, were incorporated during adaptation. Supervised cross-lingual fine-tuning, as explored in [26], in which models pretrained on multilingual corpora were subsequently fine-tuned on parallel or target-language-specific summarization datasets, was reported in the remaining 4 studies. Across the 14 studies, 10 (71%) reported English as the sole transfer source language. Two studies reported multi-source transfer configurations involving more than one high-resource language. The reported target language sets ranged from a single low-resource language to evaluations spanning up to 12 typologically diverse languages.

Shared representation learning techniques were reported in 11 of the 47 reviewed studies (23%). Studies in this category described approaches in which a unified representational space was constructed to process inputs and produce outputs across multiple languages, without maintaining language-specific model components. Two principal shared representation strategies covered in [27] were identified: shared subword tokenization and joint multilingual embedding spaces. Shared subword tokenization, primarily via Byte-Pair Encoding (BPE) or SentencePiece unigram language model vocabularies, was reported in 9 of the 11 studies as the mechanism for achieving multilingual vocabulary coverage. These studies reported vocabulary sizes ranging from 32,000 tokens to 250,000 tokens, with larger vocabulary configurations reported in studies covering broader language sets. Joint multilingual embedding space learning, in which language-agnostic sentence or token representations are trained using contrastive or alignment objectives, was proposed in [28] and reported in 4 studies, with two of these studies reporting shared tokenization alongside joint embedding training as co-occurring techniques. Studies in this category reported the use of shared representations in both encoder-only and encoder–decoder configurations. Six of the 11 studies (55%) reported shared encoder representations with language-specific or shared decoder modules. The remaining 5 studies reported fully shared encoder–decoder architectures in which no language-specific parameters were maintained, or in which language identity tokens were prepended to guide output generation.

Parameter-efficient adaptation methods were the second most frequently reported technique category, documented in 16 of the 47 reviewed studies (34%). Studies in this category reported techniques that enable the adaptation of large pretrained multilingual models for summarization tasks or specific languages by modifying only a restricted subset of model parameters, rather than performing full fine-tuning [29]. Across all 16 studies in this category, the reported proportions of trainable parameters were consistently below 5% of the total number of model parameters, with exact values explicitly stated in 9 of them. Four distinct parameter-efficient adaptation approaches were identified across the reviewed studies: adapter modules, Low-Rank Adaptation (LoRA), prefix-tuning, and prompt tuning. Adapter module insertion was the most frequently reported parameter-efficient technique, documented in 7 of the 16 studies. These studies reported inserting bottleneck feed-forward adapter layers at each Transformer block, with trainable parameter counts ranging from 0.5% to 3.2% of the total model parameters. LoRA-based adaptation was reported in 5 studies, which described decomposing the weight update matrices of attention layers into low-rank products during fine-tuning. Reported rank values across LoRA-based studies ranged from r = 4 to r = 16. Prefix-tuning was reported in 3 studies, which described prepending a sequence of trainable continuous vectors to the input of each Transformer layer to condition the model on target-language or task-specific summarization behavior. One study reported using prompt tuning as a soft-prompt adaptation strategy applied exclusively to the input embedding layer.

Data-centric and augmentation techniques were reported in 12 of the 47 reviewed studies (26%). Studies in this category described techniques applied at the data pipeline level to expand the availability of multilingual training data for summarization. Four augmentation strategies were identified across the reviewed literature: back-translation, parallel corpus construction, cross-lingual data augmentation via machine translation, and synthetic summary generation. Back-translation was the most frequently reported data augmentation technique, documented in 5 of the 12 studies. These studies described translating monolingual target-language documents into a high-resource source language for model training, or generating target-language summaries by translating source-language summaries produced by a pretrained model, as documented in [30]. Parallel corpus construction involving the alignment of document–summary pairs across language pairs was reported in 4 studies. Two of these studies reported constructing new multilingual summarization datasets as a primary contribution alongside the proposed technique.

Cross-lingual data augmentation via neural machine translation (NMT) systems was reported in 3 studies, which described translating existing monolingual summarization training corpora into one or more target languages to create synthetic multilingual training sets. Synthetic summary generation, in which a pretrained summarization model is used to produce pseudo-reference summaries for unannotated multilingual documents, according to [31], was reported in 2 studies. One study reported the combined use of back-translation and synthetic summary generation within a single data preparation pipeline. In all 12 studies in this category, the reported augmented dataset sizes ranged from approximately 5,000 synthetic instances to over 200,000 augmented document–summary pairs.

Low-resource language coverage was the most broadly reported limitation category, documented in 31 of the 47 reviewed studies (66%). Language-wise analysis shows that the reviewed literature remains strongly concentrated on high-resource languages, particularly English, Chinese, French, German, and Spanish. Among the 31 studies reporting limitations related to low-resource language coverage, 24 were associated with the dominance of high-resource pre-training or evaluation languages, while 9 specifically reported an English-centric development or evaluation bias. This concentration limits the extent to which multilingual summarization models are evaluated across linguistically diverse settings. In particular, 12 studies reported limited representation or absence of non-Latin-script languages, including Arabic, Hindi, Amharic, and languages from Southeast Asia. These findings indicate that multilingual coverage remains uneven, with substantial attention given to languages for which larger training and evaluation resources are already available.

The analysis also identified language-specific challenges beyond simple differences in the number of languages represented. Six studies reported vocabulary fragmentation or tokenization-related difficulties affecting languages such as Finnish, Turkish, Swahili, and Yoruba, while four studies associated the under-representation of low-resource languages in pre-training data with lower automatic evaluation scores [32]. However, the latter relationship was reported as an association rather than a causal effect. Overall, the findings suggest that the low-resource language gap involves multiple factors, including limited training resources, inadequate representation of non-Latin scripts, and difficulties in representing morphologically diverse languages. Consequently, conclusions about multilingual summarization performance remain considerably stronger for high-resource languages than for low-resource languages.

Dataset and benchmark limitations were reported in 27 of the 47 reviewed studies (57%). The dataset-wise analysis indicates that the evidence base is also concentrated around a relatively limited range of datasets and domains. Of the 27 studies reporting dataset and benchmark limitations, 21 identified a predominance of news-oriented data, with datasets such as XL-Sum, MLSum, WikiLingua, and CNN/DailyMail appearing among the datasets used in the reviewed studies [33]. Furthermore, 15 of these 21 studies relied primarily on a single dataset, limiting opportunities to assess whether model performance remains consistent across different domains and data characteristics. This concentration is particularly important for multilingual summarization because performance observed in a single domain may not adequately reflect models' ability to summarize content across different languages, topics, and document types.

Language coverage and benchmark standardization were also identified as important dataset-level limitations. Six studies reported skewed language distributions, while eight studies used or discussed datasets that covered fewer than 10 languages. In addition, 13 studies reported the absence of a standardized multilingual benchmark, making direct comparison between models and approaches more difficult [34]. Five studies also identified issues related to noise or inconsistent data quality, while three developed new datasets because existing resources were considered insufficient for their research objectives. Taken together, these findings show that dataset limitations extend beyond the availability of multilingual corpora to include uneven language representation, dependence on individual datasets, differences in data quality, and the lack of standardized benchmarks for evaluating multilingual summarization systems.

Evaluation limitations constituted the most frequently documented limitation category in the reviewed corpus, reported in 34 of the 47 studies (72%). Over-reliance on ROUGE-based automatic evaluation metrics was the dominant pattern, observed in 30 of the 34 studies in this category. Of these, 18 studies (60% of the ROUGE-reliant subset) reported ROUGE-1, ROUGE-2, and ROUGE-L as the sole evaluation metrics, with no supplementary metric or human assessment included. Fourteen studies explicitly acknowledged that ROUGE scores exhibit reduced sensitivity to lexical and morphological variation across languages, a limitation reported in [35], whereby n-gram overlap-based scoring may inadequately capture summary quality in morphologically rich or agglutinative target languages.

Human evaluation was reported in only 11 of the 47 reviewed studies (23%). Of these, 7 studies (64% of the human-evaluation subset) employed human assessment as a supplementary measure alongside automatic metrics, while 4 studies reported human evaluation as the primary assessment protocol. Evaluation criteria employed in human assessment varied across studies: fluency and coherence were assessed in all 11 studies, factual consistency in 6 studies, and relevance in 4 studies. Inconsistent evaluation protocols across the reviewed corpus were reported as a limitation in 19 studies. These studies documented variability in evaluated language sets, summary length constraints, tokenization procedures for metric computation, and test split definitions across the literature, with 8 studies explicitly stating that these inconsistencies prohibited direct numerical comparison of reported scores with prior work [36]. Five studies additionally reported that BERTScore and METEOR were employed as supplementary automatic metrics to partially address ROUGE limitations, though cross-study comparability remained constrained by inconsistent metric configurations.

Generalization and scalability challenges were reported in 29 of the 47 reviewed studies (62%). Within this category, the curse of multilinguality was the most consistently documented recurring phenomenon, reported explicitly in 22 of the 29 studies (76%). The reviewed literature also identified the curse of multilinguality as a scalability challenge, whereby increasing the number of supported languages can reduce per-language performance under fixed model capacity. These studies reported that as the number of languages supported by a single model increased, per-language task performance declined, a pattern attributed to fixed-capacity model parameters being distributed across an expanding set of languages [37]. Fifteen of the 22 studies documenting the curse of multilinguality reported this degradation in the context of models supporting more than 15 languages simultaneously. Seven studies (32% of this subset) reported that the degradation was most pronounced for low-resource languages with limited representation in the pre-training corpus, consistent with the study in [2] about low-resource coverage limitations.

Cross-domain generalization limitations were reported in 17 of the 29 studies (59%) in this category. Fourteen of these studies (82% of the cross-domain subset) reported that models trained and evaluated exclusively on news-domain corpora exhibited performance degradation when applied to out-of-domain text types, including scientific abstracts, legal documents, and conversational transcripts [38]. Cross-lingual generalization to unseen language families was reported as a limitation in 13 studies (45% of this category), with 8 of these studies reporting that zero-shot transfer to typologically distant languages yielded substantially lower automatic metric scores than transfer within the same language family. Computational constraints were explicitly reported as a scalability limitation in 18 of the 29 studies. Of these, 11 studies (61% of the computation-limited subset) reported that full fine-tuning of large multilingual models, particularly those exceeding 600M parameters, was computationally infeasible under the reported experimental resource constraints, with GPU memory limitations cited as the primary constraint in 7 studies. Four studies also reported that multilingual scaling in extending a trained summarization model to additional languages post-training introduced catastrophic forgetting of previously learned language representations, with [39] identifying this as a distinct scalability challenge.