Recent advances in text-to-SQL models, which translate natural language questions (NLQs) into executable SQL queries, have made interacting with relational databases more accessible, even for those with limited technical ability. This task has seen significant improvement with the release of multiple English datasets and benchmarks such as WikiSQL, SPIDER, and BIRD, each covering different domains and levels of complexity. Non-English high-resource languages, such as Chinese, Russian, and Arabic, have also benefited from these advances, either through the translation of existing datasets or the creation of new ones. Dialect2SQL, a newly released text-to-SQL dataset, is dedicated to the Moroccan dialect (Darija), which is known for its complexity and distinctiveness compared to other Arabic dialects and Modern Standard Arabic. In this article, we conduct a comprehensive study on text-to-SQL for Darija by conducting several experiments mainly on the Dialect2SQL dataset using different approaches and configurations with two code-based large language models, StarCoder2 and Qwen-2.5-Coder. The experiments reveal the performance gap between models fine-tuned on English data and those fine-tuned on Darija. Additionally, the results illustrate the positive impact of incorporating multi-language datasets during training. In particular, the gap decreases from 10.1% to 6.7% in BLEU, and from 12.5% to 5.7% in TSED.

Recent page-image retrievers such as ColPali have improved retrieval over visually rich documents, yet little is known about how they behave in cross-lingual, low-resource settings. We introduce LëtzCross, a benchmark for cross-lingual page-level retrieval over Luxembourgish PDF documents, with document pages indexed as images and queries provided in English, French, German, and Luxembourgish. The benchmark combines text-focused QA pairs with visually grounded QA pairs, covering both textual and visual retrieval needs in PDF-based RAG. We use LëtzCross to compare OCR-based text-only retrievers with ColPali-style page-image retrievers and find that the latter perform better across query languages in this system-level comparison. We also examine single-language and multilingual fine-tuning. Fine-tuning transfers across query languages, with French yielding the highest mean performance on Luxembourgish queries among the single-language settings. In the multilingual setting, including Luxembourgish gives the strongest results and substantially improves retrieval for Luxembourgish queries.

PDF files are primarily intended for human reading rather than automated processing. In addition, the heterogeneous content of PDFs, such as text, tables, and images, poses significant challenges for parsing and information extraction. To address these difficulties, both practitioners and researchers are increasingly developing new methods, including the promising Retrieval-Augmented Generation (RAG) systems to automated PDF processing. However, there is no comprehensive study investigating how different components and design choices affect the performance of a RAG system for understanding PDFs. In this paper, we propose such a study (1) by focusing on Question Answering, a specific language understanding task, and (2) by leveraging two benchmarks from the financial domain, including TableQuest, our newly generated, publicly available benchmark. We systematically examine multiple PDF parsers and chunking strategies (with varied overlap), along with their potential synergies in preserving document structure and ensuring answer correctness. Overall, our results offer practical guidelines for building robust RAG pipelines for PDF understanding.

As software systems increasingly rely on natural language interfaces, ensuring the reliability of these systems is crucial. One critical component is the ability to accurately translate natural language queries into corresponding SQL queries, a field known as Text-to-SQL. However, the scarcity of high-quality, large-scale, and domain-specific Text-to-SQL datasets hinders the development of reliable and robust models. To tackle these challenges, we propose SelectCraft, a novel automatic generation approach designed to create realistic Text-to-SQL datasets tailored to specific domains. Our method leverages existing databases and their structures to generate complex text-SQL pairs that mirror real-world usage scenarios. As a proof of concept, we have successfully generated a substantial financial Text-to-SQL dataset, denominated as BanQies, encompassing over 1 million samples utilizing our proposed approach. Moreover, we introduce BanQL, a new large language model (LLM) based on StarCoder2, a state-of-the-art code-based LLM, and fine-tuned on our newly created dataset. We evaluate BanQL performance against several state-of-the-art models, demonstrating significant enhancements in accuracy and generalizability, highlighting the advantages of incorporating domain-specific data in Text-to-SQL tasks. We firmly believe that our contributions have the potential to improve the overall reliability of Text-to-SQL software systems.

Low-resource languages (LRLs) lack sufficient linguistic resources and are underrepresented in benchmark datasets, resulting in persistently lower translation quality than high-resource languages, especially in privacy-sensitive and resource-limited contexts. Firstly, this study systematically evaluates state-of-the-art smaller Large Language Models in 200 languages using the FLORES-200 benchmark, highlighting persistent deficiencies and disparities in the translation of LRLs. To mitigate these limitations, we investigate knowledge distillation from large pre-trained teacher models to Small Language Models (SLMs) through supervised fine-tuning. The results show substantial improvements; for example, the translation performance of English to Luxembourgish (EN to LB), measured by the LLM-as-a-Judge score, increases from 0.36 to 0.89 in the validation set for Llama-3.2-3B. We further investigate various fine-tuning configurations and tasks to clarify the trade-offs between data scale and training efficiency, verify that the model retains its general capabilities without significant catastrophic forgetting after training, and explore the distillation benefits to other LRLs on SLMs (Khasi, Assamese, and Ukrainian). In general, this work exposes the limitations and fairness issues of current SLMs in LRL translation and systematically explores the potential of using the distillation of knowledge from large to small models, offering practical, empirically grounded recommendations to improve LRL translation systems

Arabic automatic speech recognition (ASR) faces unique challenges due to diglossia, extensive regional dialect variation, and limited speech resources. Existing speech datasets often focus on single dialects or large-scale broadcast/web data, leading to trade-offs between linguistic diversity and annotation quality. We present BULBUL, a multi-dialect Arabic ASR dataset collected from 275 speakers in 11 Arab countries. BULBUL includes structured dialect and sub-dialect coverage, as well as recordings of classical Arabic and modern standard Arabic spoken by participants in their native dialectal accents to support accent-aware modeling. The quality of the recordings was ensured through a two-level human verification process. We further benchmark a range of recent ASR systems, establishing strong baselines for modern dialectal and accented Arabic ASR.

LLM-generated outputs in software engineering rarely exist in isolation. They must plug into toolchains, APIs, and data pipelines that impose strict, often organization-specific structural contracts. A semantically correct output that violates the expected format is, from the consuming system's perspective, indistinguishable from a wrong answer, making structural fidelity an operational prerequisite for deploying LLMs in practice. Yet current models routinely produce syntactically invalid or structurally non-compliant outputs. Unlike encoders, autoregressive decoders generate text token-by-token with a local rather than global focus, amplifying structural fragility whenever the target format deviates from familiar training distributions. We present a systematic evaluation of structural reliability across four representative SE tasks, categorizing failures into syntax, structural, and semantic errors. We benchmark ways of mitigation targeting the decoder: grammar-constrained decoding, regex-based validation, and a strict template-driven control (Template Token Match Generation, TTMG) to isolate the sources of these failures. TTMG nearly eliminates syntax errors, yet substantial structural and semantic errors persist, demonstrating that the core bottleneck lies beyond syntax formatting. A detailed case study further illustrates how residual errors cascade in downstream workflows. Our findings show that current structure-enforcing tools are necessary but insufficient, and highlight the need for approaches that jointly ensure structural fidelity and semantic correctness in LLM-driven workflows.

Recently, code-oriented large language models (LLMs) have demonstrated strong capabilities in translating natural language into executable code. Text-to-SQL is a significant application of this ability, enabling non-technical users to interact with relational databases using natural language. However, state-of-the-art models continue to struggle with highly complex logic, particularly deeply nested statements involving multiple joins and conditions, as well as with real-world database schemas that are noisy or poorly structured. In this paper, we investigate whether curriculum learning can improve the performance of code-based LLMs on Text-to-SQL tasks. Employing benchmarks including Spider and BIRD, we fine-tune models under different curriculum strategies. Our experiments show that naive curriculum, simply ordering training samples by complexity in a single epoch, fails to surpass standard fine-tuning due to catastrophic forgetting. To overcome this, we propose a Modular Adapter Composition (MAC) strategy. By sequentially training tier-specific adapters on incremental complexity levels (Easy to Extra-Hard), we create a scaffolded learning environment that improves performance on complex queries. Our approach not only produces measurable performance gains on the Spider and BIRD benchmarks but also provides a flexible, "Lego-like" architecture, allowing models to be composed and deployed based on specific schema difficulty requirements. These findings demonstrate that structured, modular learning is a superior alternative to monolithic fine-tuning for mastering the syntax and logic of complex code generation.

API-driven chatbot systems are increasingly integral to software engineering applications, yet their effectiveness hinges on accurately generating and executing API calls. This is particularly challenging in scenarios requiring multi-step interactions with complex parameterization and nested API dependencies. Addressing these challenges, this work contributes to the evaluation and assessment of AI-based software development through three key advancements: (1) the introduction of a novel dataset specifically designed for benchmarking API function selection, parameter generation, and nested API execution; (2) an empirical evaluation of state-of-the-art language models, analyzing their performance across varying task complexities in API function generation and parameter accuracy; and (3) a hybrid approach to API routing, combining general-purpose large language models for API selection with fine-tuned models and prompt engineering for parameter generation. These innovations significantly improve API execution in chatbot systems, offering practical methodologies for enhancing software design, testing, and operational workflows in real-world software engineering contexts.

Code change representation plays a pivotal role in automating numerous software engineering tasks, such as classifying code change correctness or generating natural language summaries of code changes. Recent studies have leveraged deep learning to derive effective code change representation, primarily focusing on capturing changes in token sequences or Abstract Syntax Trees (ASTs). However, these current state-of-the-art representations do not explicitly calculate the intention semantic induced by the change on the AST, nor do they effectively explore the surrounding contextual information of the modified lines. To address this, we propose a new code change representation methodology, Patcherizer, which we refer to as our tool. This innovative approach explores the intention features of the context and structure, combining the context around the code change along with two novel representations. These new representations capture the sequence intention inside the code changes in the code change and the graph intention inside the structural changes of AST graphs before and after the code change. This comprehensive representation allows us to better capture the intentions underlying a code change. Patcherizer builds on graph convolutional neural networks for the structural input representation of the intention graph and on transformers for the intention sequence representation. We assess the generalizability of Patcherizer ’s learned embeddings on three tasks: (1) Generating code change description in NL, (2) Predicting code change correctness in program repair, and (3) Code change intention detection. Experimental results show that the learned code change representation is effective for all three tasks and achieves superior performance to the state-of-the-art (SOTA) approaches. For instance, on the popular task of code change description generation (a.k.a. commit message generation), Patcherizer achieves an average improvement of 19.39

This paper explores the significant advancements in generating Structured Query Language (SQL) from natural language, primarily driven by Large Language Models (LLMs). These advancements have led to the development of sophisticated text-to-SQL integrated applications, enabling easier database (DB) querying for users unfamiliar with SQL syntax using natural language queries. However, reliance on LLMs exposes these applications to potential attacks through the introduction of malicious prompts or by compromising models with malicious data during the training phase. Such attacks pose severe risks, including unauthorized data access or even complete DB destruction upon success. To address these concerns, we introduce a novel large-scale dataset comprising malicious and safe prompts along with their corresponding SQL queries, enabling model fine-tuning on malicious query detection tasks. Moreover, we propose the implementation of two transformer-based classification solutions to aid in the detection of malicious attacks. Finally, we present a secure agent-based text-to-SQL architecture that incorporates these solutions to enhance overall system security, resulting in a 70% security enhancement overall compared to solely relying on a conventional text-to-SQL model.

The generation of highly fluent text by Large Language Models (LLMs) poses a significant challenge to information integrity and academic research. In this paper, we introduce the Multi-Domain Detection of AI-Generated Text (M-DAIGT) shared task, which focuses on detecting AI-generated text across multiple domains, particularly in news articles and academic writing. M-DAIGT comprises two binary classification subtasks: News Article Detection (NAD) (Subtask 1) and Academic Writing Detection (AWD) (Subtask 2). To support this task, we developed and released a new large-scale benchmark dataset of 30,000 samples, balanced between human-written and AI-generated texts. The AI-generated content was produced using a variety of modern LLMs (e.g., GPT-4, Claude) and diverse prompting strategies. A total of 46 unique teams registered for the shared task, of which four teams submitted final results. All four teams participated in both Subtask 1 and Subtask 2. We describe the methods employed by these participating teams and briefly discuss future directions for M-DAIGT.

The task of converting natural language questions (NLQs) into executable SQL queries, known as text-to-SQL, has gained significant interest in recent years, as it enables non-technical users to interact with relational databases. Many benchmarks, such as SPIDER and WikiSQL, have contributed to the development of new models and the evaluation of their performance. In addition, other datasets, like SEDE and BIRD, have introduced more challenges and complexities to better map real-world scenarios. However, these datasets primarily focus on high-resource languages such as English and Chinese. In this work, we introduce Dialect2SQL, the first large-scale, cross-domain text-to-SQL dataset in an Arabic dialect. It consists of 9,428 NLQ-SQL pairs across 69 databases in various domains. Along with SQL-related challenges such as long schemas, dirty values, and complex queries, our dataset also incorporates the complexities of the Moroccan dialect, which is known for its diverse source languages, numerous borrowed words, and unique expressions. This demonstrates that our dataset will be a valuable contribution to both the text-to-SQL community and the development of resources for low-resource languages.

The hospitality industry in the Arab world increasingly relies on customer feedback to shape services, driving the need for advanced Arabic sentiment analysis tools. To address this challenge, the Sentiment Analysis on Arabic Dialects in the Hospitality Domain shared task focuses on Sentiment Detection in Arabic Dialects. This task leverages a multi-dialect, manually curated dataset derived from hotel reviews originally written in Modern Standard Arabic (MSA) and translated into Saudi and Moroccan (Darija) dialects. The dataset consists of 538 sentiment-balanced reviews spanning positive, neutral, and negative categories. Translations were validated by native speakers to ensure dialectal accuracy and sentiment preservation. This resource supports the development of dialect-aware NLP systems for real-world applications in customer experience analysis. More than 40 teams have registered for the shared task, with 12 submitting systems during the evaluation phase. The top-performing system achieved an F1 score of 0.81, demonstrating the feasibility and ongoing challenges of sentiment analysis across Arabic dialects.

Large language models (LLMs) have shown remarkable progress in reasoning abilities and general natural language processing (NLP) tasks, yet their performance on Arabic data, characterized by rich morphology, diverse dialects, and complex script, remains underexplored. This paper presents a comprehensive benchmarking study of multiple reasoning-focused LLMs, with a special emphasis on the newly introduced DeepSeek models, across a suite of fifteen Arabic NLP tasks. We experiment with various strategies, including zero-shot, few-shot, and fine-tuning. This allows us to systematically evaluate performance on datasets covering a range of applications to examine their capacity for linguistic reasoning under different levels of complexity. Our experiments reveal several key findings. First, carefully selecting just three in-context examples delivers an average uplift of over 13 F1 points on classification tasks—boosting sentiment analysis from 35.3% to 87.5% and paraphrase detection from 56.1% to 87.0%. Second, reasoning-focused DeepSeek architectures outperform a strong GPT o4-mini baseline by an average of 12 F1 points on complex inference tasks in the zero-shot setting. Third, LoRA-based fine-tuning yields up to an additional 8 points in F1 and BLEU compared to equivalent increases in model scale. The code is available at https://anonymous.4open.science/r/AraReasoner41299

Software development teams generally welcome any effort to expose bugs in their code base. In this work, we build on the hypothesis that mobile apps from the same category (e.g., two web browser apps) may be affected by similar bugs in their evolution process. It is therefore possible to transfer the experience of one historical app to quickly find bugs in its new counterparts. This has been referred to as collaborative bug finding in the literature. Our novelty is that we guide the bug finding process by considering that existing bugs have been hinted within app reviews. Concretely, we design the BugRMSys approach to recommend bug reports for a target app by matching historical bug reports from apps in the same category with user app reviews of the target app. We experimentally show that this approach enables us to quickly expose and report dozens of bugs for targeted apps such as Brave (web browser app). BugRMSys ’s implementation relies on DistilBERT to produce natural language text embeddings. Our pipeline considers similarities between bug reports and app reviews to identify relevant bugs. We then focus on the app review as well as potential reproduction steps in the historical bug report (from a same-category app) to reproduce the bugs. Overall, after applying BugRMSys to six popular apps, we were able to identify, reproduce and report 20 new bugs: among these, 9 reports have been already triaged, 6 were confirmed, and 4 have been fixed by official development teams.

In recent years, patch representation learning has emerged as a necessary research direction for exploiting the capabilities of machine learning in software generation. These representations have driven significant performance enhancements across a variety of tasks involving code changes. While the progress is undeniable, a common limitation among existing models is their specialization: they predominantly excel in either predictive tasks, such as security patch classification, or in generative tasks such as patch description generation. This dichotomy is further exacerbated by a prevalent dependency on potentially noisy data sources. Specifically, many models utilize patches integrated with Abstract Syntax Trees (AST) that, unfortunately, may contain parsing inaccuracies, thus acting as a suboptimal source of supervision. In response to these challenges, we introduce PATCHSYNTH, a novel pre-training framework for patches and natural language text. PATCHSYNTH deploys a triple-loss training strategy for (1) patch-description contrastive learning, which enables to separate patches and descriptions in the embedding space, (2) patch-description matching, which ensures that each patch is associated to its description in the embedding space, and (3) patch-description generation, which ensures that the patch embedding is effective for generation. These losses are implemented for joint learning to achieve good performance in both predictive and generative tasks involving patches. Empirical evaluations focusing on patch description generation, demonstrate that PATCHSYNTH sets new state of the art performance, consistently outperforming the state-of-the-art in metrics like BLEU, ROUGE-L, METEOR, and Recall.