Pooja Rani

Joel Niklaus, Veton Matoshi, Pooja Rani et al.

Lately, propelled by the phenomenal advances around the transformer architecture, the legal NLP field has enjoyed spectacular growth. To measure progress, well curated and challenging benchmarks are crucial. However, most benchmarks are English only and in legal NLP specifically there is no multilingual benchmark available yet. Additionally, many benchmarks are saturated, with the best models clearly outperforming the best humans and achieving near perfect scores. We survey the legal NLP literature and select 11 datasets covering 24 languages, creating LEXTREME. To provide a fair comparison, we propose two aggregate scores, one based on the datasets and one on the languages. The best baseline (XLM-R large) achieves both a dataset aggregate score a language aggregate score of 61.3. This indicates that LEXTREME is still very challenging and leaves ample room for improvement. To make it easy for researchers and practitioners to use, we release LEXTREME on huggingface together with all the code required to evaluate models and a public Weights and Biases project with all the runs.

Melih Catal, Alex Wolf, Tiago Ferreiro Matos et al.

Large Language Models (LLMs) have become an integral part of software development, especially with the advent of agentic capabilities. Yet, many frontier LLMs are affiliated with specific providers. This raises the question of whether generated code favors the provider's own ecosystem over comparable alternatives, potentially constraining developers' choices and increasing dependence on a single provider. We define this behavior as Vertical Integration Bias (VIB) and introduce \textsc{VIBench}, a benchmark for measuring VIB in direct and agentic code generation across $20$ provider-selectable software-integration scenarios. Evaluating $10$ frontier provider-affiliated models against $3$ non-affiliated controls, we find positive VIB in direct generation, with six of ten affiliated models showing statistically significant effects up to $+18.8$ percentage points (pp). Agentic workflows further amplify VIB, reaching $+39.2$ pp. Moreover, early affiliated-ecosystem choices in agentic workflows can persist into conceptually decoupled downstream files, with persistence as high as $90.3\%$. These findings underscore the need to measure and account for VIB in code generation, especially as agentic capabilities become more prevalent.

Melih Catal, Pooja Rani, Harald C. Gall

Large language models specialized for code (CodeLLMs) have demonstrated remarkable capabilities in generating code snippets, documentation, and test cases. However, despite their promising capabilities, CodeLLMs can inadvertently memorize and reproduce snippets from their training data, which poses risks of privacy breaches and intellectual property violations. These risks restrict the deployment of CodeLLMs in sensitive domains and limit their training datasets to publicly available sources. To mitigate the memorization risk without compromising their task performance, we apply Differential Privacy (DP) to CodeLLMs. To the best of our knowledge, this is the first comprehensive study that systematically evaluates the effectiveness of DP in CodeLLMs. DP adds calibrated noise to the training process to protect individual data points while still allowing the model to learn useful patterns. To this end, we first identify and understand the driving reasons of the memorization behaviour of the CodeLLMs during their fine-tuning. Then, to address this issue, we empirically evaluate the effect of DP on mitigating memorization while preserving code generation capabilities. Our findings show that DP substantially reduces memorization in CodeLLMs across all the tested snippet types. The snippet types most prone to memorization are also the most effectively mitigated by DP. Furthermore, we observe that DP slightly increases perplexity but preserves, and can even enhance, the code generation capabilities of CodeLLMs, which makes it feasible to apply DP in practice without significantly compromising model utility. Finally, we analyze the impact of DP on training efficiency and energy consumption, finding that DP does not significantly affect training time or energy usage, making it a practical choice for privacy-preserving CodeLLMs training.

Luís Cruz, João Paulo Fernandes, Maja H. Kirkeby et al.

The environmental impact of Artificial Intelligence (AI)-enabled systems is increasing rapidly, and software engineering plays a critical role in developing sustainable solutions. The "Greening AI with Software Engineering" CECAM-Lorentz workshop (no. 1358, 2025) funded by the Centre Européen de Calcul Atomique et Moléculaire and the Lorentz Center, provided an interdisciplinary forum for 29 participants, from practitioners to academics, to share knowledge, ideas, practices, and current results dedicated to advancing green software and AI research. The workshop was held February 3-7, 2025, in Lausanne, Switzerland. Through keynotes, flash talks, and collaborative discussions, participants identified and prioritized key challenges for the field. These included energy assessment and standardization, benchmarking practices, sustainability-aware architectures, runtime adaptation, empirical methodologies, and education. This report presents a research agenda emerging from the workshop, outlining open research directions and practical recommendations to guide the development of environmentally sustainable AI-enabled systems rooted in software engineering principles.

Pooja Rani, Jan-Andrea Bard, June Sallou et al.

The rapid technological evolution has accelerated software development for various domains and use cases, contributing to a growing share of global carbon emissions. While recent large language models (LLMs) claim to assist developers in optimizing code for performance and energy efficiency, their efficacy in real-world scenarios remains under exploration. In this work, we explore the effectiveness of LLMs in reducing the environmental footprint of real-world projects, focusing on software written in Matlab-widely used in both academia and industry for scientific and engineering applications. We analyze energy-focused optimization on 400 scripts across 100 top GitHub repositories. We examine potential 2,176 optimizations recommended by leading LLMs, such as GPT-3, GPT-4, Llama, and Mixtral, and a senior Matlab developer, on energy consumption, memory usage, execution time consumption, and code correctness. The developer serves as a real-world baseline for comparing typical human and LLM-generated optimizations. Mapping these optimizations to 13 high-level themes, we found that LLMs propose a broad spectrum of improvements--beyond energy efficiency--including improving code readability and maintainability, memory management, error handling while the developer overlooked some parallel processing, error handling etc. However, our statistical tests reveal that the energy-focused optimizations unexpectedly negatively impacted memory usage, with no clear benefits regarding execution time or energy consumption. Our qualitative analysis of energy-time trade-offs revealed that some themes, such as vectorization preallocation, were among the common themes shaping these trade-offs. With LLMs becoming ubiquitous in modern software development, our study serves as a call to action: prioritizing the evaluation of common coding practices to identify the green ones.

Marco Edoardo Palma, Pooja Rani, Harald C. Gall

Syntax highlighting is a critical feature in modern software development environments, enhancing code readability and developer productivity. However, delivering accurate highlighting in real time remains challenging for online and web-based development tools due to strict time and memory constraints on backend services. These systems must serve highlights rapidly and frequently, even when code is partially valid or invalid. This has led to on-the-fly syntax highlighting, where visual annotations are generated just before content is served, often at high request rates and under incomplete input conditions. To meet these demands efficiently, state-of-the-art models use deep learning to learn the behavior of brute-force syntax highlighting resolvers, tools that are easy to implement but too slow for production. Through the Deep Abstraction process, brute-force strategies are encoded into fast statistical models that achieve both high accuracy and low-latency inference. Despite their success, such models face key challenges: they support only one programming language per model, require large datasets from slow brute-force generators, and involve resource-intensive training. In multi-language environments, this means maintaining multiple independent models, increasing system complexity and operational cost. This work addresses these issues by introducing a unified model capable of highlighting up to six mainstream programming languages, reducing deployment complexity by a factor of six and improving performance on unseen languages. A novel normalization technique significantly enhances model generalization, while few-shot learning experiments show that a small number of oracle samples can replace large datasets, minimizing dependence on brute-force generators. Combined, these innovations enable efficient, scalable, and cost-effective syntax highlighting across diverse programming languages.

Pooja Rani, Suada Abukar, Nataliia Stulova et al.

Assessing code comment quality is known to be a difficult problem. A number of coding style guidelines have been created with the aim to encourage writing of informative, readable, and consistent comments. However, it is not clear from the research to date which specific aspects of comments the guidelines cover (e.g., syntax, content, structure). Furthermore, the extent to which developers follow these guidelines while writing code comments is unknown. We analyze various style guidelines in Java and Python and uncover that the majority of them address more the content aspect of the comments rather than syntax or formatting, but when considering the different types of information developers embed in comments and the concerns they raise on various online platforms about the commenting practices, existing comment conventions are not yet specified clearly enough, nor do they adequately cover important concerns. We also analyze commenting practices of developers in diverse projects to see the extent to which they follow the guidelines. Our results highlight the mismatch between developer commenting practices and style guidelines, and provide several focal points for the design and improvement of comment quality checking tools.

Pooja Rani, Mathias Birrer, Sebastiano Panichella et al.

Code comments are important for program comprehension, development, and maintenance tasks. Given the varying standards for code comments, and their unstructured or semi-structured nature, developers get easily confused (especially novice developers) about which convention(s) to follow, or what tools to use while writing code documentation. Thus, they post related questions on external online sources to seek better commenting practices. In this paper, we analyze code comment discussions on online sources such as Stack Overflow (SO) and Quora to shed some light on the questions developers ask about commenting practices. We apply Latent Dirichlet Allocation (LDA) to identify emerging topics concerning code comments. Then we manually analyze a statistically significant sample set of posts to derive a taxonomy that provides an overview of the developer questions about commenting practices. Our results highlight that on SO nearly 40% of the questions mention how to write or process comments in documentation tools and environments, and nearly 20% of the questions are about potential limitations and possibilities of documentation tools to add automatically and consistently more information in comments. On the other hand, on Quora, developer questions focus more on background information (35% of the questions) or asking opinions (16% of the questions) about code comments. We found that (i) not all aspects of comments are covered in coding style guidelines, e.g., how to add a specific type of information, (ii) developers need support in learning the syntax and format conventions to add various types of information in comments, and (iii) developers are interested in various automated strategies for comments such as detection of bad comments, or verify comment style automatically, but lack tool support to do that.

Pooja Rani, Sebastiano Panichella, Manuel Leuenberger et al.

Most software maintenance and evolution tasks require developers to understand the source code of their software systems. Software developers usually inspect class comments to gain knowledge about program behavior, regardless of the programming language they are using. Unfortunately, (i) different programming languages present language-specific code commenting notations/guidelines; and (ii) the source code of software projects often lacks comments that adequately describe the class behavior, which complicates program comprehension and evolution activities. To handle these challenges, this paper investigates the different language-specific class commenting practices of three programming languages: Python, Java, and Smalltalk. In particular, we systematically analyze the similarities and differences of the information types found in class comments of projects developed in these languages. We propose an approach that leverages two techniques, namely Natural Language Processing and Text Analysis, to automatically identify various types of information from class comments i.e., the specific types of semantic information found in class comments. To the best of our knowledge, no previous work has provided a comprehensive taxonomy of class comment types for these three programming languages with the help of a common automated approach. Our results confirm that our approach can classify frequent class comment information types with high accuracy for Python, Java, and Smalltalk programming languages. We believe this work can help to monitor and assess the quality and evolution of code comments in different program languages, and thus support maintenance and evolution tasks.

Pooja Rani

Previous studies have shown that high-quality code comments assist developers in program comprehension and maintenance tasks. However, the semi-structured nature of comments, unclear conventions for writing good comments, and the lack of quality assessment tools for all aspects of comments make their evaluation and maintenance a non-trivial problem. To achieve high-quality comments, we need a deeper understanding of code comment characteristics and the practices developers follow. In this thesis, we approach the problem of assessing comment quality from three different perspectives: what developers ask about commenting practices, what they write in comments, and how researchers support them in assessing comment quality. Our preliminary findings show that developers embed various kinds of information in class comments across programming languages. Still, they face problems in locating relevant guidelines to write consistent and informative comments, verifying the adherence of their comments to the guidelines, and evaluating the overall state of comment quality. To help developers and researchers in building comment quality assessment tools, we provide: (i) an empirically validated taxonomy of comment convention-related questions from various community forums, (ii) an empirically validated taxonomy of comment information types from various programming languages, (iii) a language-independent approach to automatically identify the information types, and (iv) a comment quality taxonomy prepared from a systematic literature review.

Pooja Rani, Sebastiano Panichella, Manuel Leuenberger et al.

Previous studies have characterized code comments in various programming languages to support better program comprehension activities and maintenance tasks. However, very few studies have focused on understanding developer practices to write comments. None of them has compared such developer practices to the standard comment guidelines to study the extent to which developers follow the guidelines. This paper reports the first empirical study investigating commenting practices in Pharo Smalltalk. First, we analyze class comment evolution over seven Pharo versions. Then, we investigate the information types embedded in class comments. Finally, we study the adherence of developer commenting practices to the official class comment template over Pharo versions. The results of this study show that there is a rapid increase in class comments in the initial three Pharo versions, while in subsequent versions developers added comments to both new and old classes, thus maintaining a similar code to comment ratio. We furthermore found three times as many information types in class comments as those suggested by the template. However, the information types suggested by the template tend to be present more often than other types of information. Additionally, we find that a substantial proportion of comments follow the writing style of the template in writing these information types, but they are written and formatted in a non-uniform way.This suggests the need to standardize the commenting guidelines for formatting the text, and to provide headers for the different information types to ensure a consistent style and to identify the information easily. Given the importance of high-quality code comments, we draw numerous implications for developers and researchers to improve the support for comment quality assessment tools.