Abstract Geometric Semantic Genetic Programming ( $$\mathsf {GSGP}$$ ) is a powerful variant of Genetic Programming (GP) that defines genetic operators inducing unimodal fitness landscapes. In recent years, a new mutation operator, Geometric Semantic Mutation with Local Search (GSM-LS), has been proposed to include a local search step in the mutation process. The core idea of GSM-LS is to incorporate a linear regression step during mutation, thereby accelerating convergence toward high-quality solutions. While GSM-LS helps the convergence of the evolutionary search, it is prone to overfitting. Thus, it was suggested to apply GSM-LS only for a limited number of generations and then revert to standard geometric semantic mutation. A more recently defined variant of $$\mathsf {GSGP}$$ (called $$\mathsf {GSGP}$$ -reg) also includes a local search step, but shares similar strengths and weaknesses with GSM-LS. Here, we investigate several strategies to mitigate overfitting in GSM-LS and $$\mathsf {GSGP}$$ -reg, ranging from simple regularized regression techniques to adaptive methods that estimate overfitting risk at each mutation. The latter approaches partition the training set into two subsets: one used to perform the mutation, and the other to evaluate the risk of overfitting based on the mutation’s impact on held-out data. Experimental evaluations across seven real-world regression benchmarks show that, while plain GSGP underperforms on all datasets, methods incorporating local search often achieve significantly better test performance. For example, on the Airfoil dataset, the GSM-LS variant achieves a median RMSE below 10 compared to 30 with standard GSGP. On the LD50 and Bioavailability datasets, the proposed gen and ridge-regularized variants effectively mitigate overfitting, reducing test RMSE by up to 40% relative to baseline GSGP. We conclude that local search, when used with regularization strategies, enhances GSGP’s performance and generalization capability across a diverse range of tasks.

Understanding source code is a topic of great interest in the software engineering community, since it can help programmers in various tasks such as software maintenance and reuse. Recent advances in large language models (LLMs) have demonstrated remarkable program comprehension capabilities, while transformer-based topic modeling techniques offer effective ways to extract semantic information from text. This paper proposes and explores a novel approach that combines these strengths to automatically identify meaningful topics in a corpus of Python programs. Our method consists in applying topic modeling on the descriptions obtained by asking an LLM to summarize the code. To assess the internal consistency of the extracted topics, we compare them against topics inferred from function names alone, and those derived from existing docstrings. Experimental results suggest that leveraging LLM-generated summaries provides interpretable and semantically rich representation of code structure. The promising results suggest that our approach can be fruitfully applied in various software engineering tasks such as automatic documentation and tagging, code search, software reorganization and knowledge discovery in large repositories.

Large language models (LLMs) are increasingly gaining relevance in every-day life, due to their apparent ability in solving tasks that demand intricate linguistic comprehension. Recent studies state that one of the key points that impact their outcome is the quality of the prompt used to interact with them. This work proposes a grammar-based evolutionary approach for exploring the prompt space of LLMs, driven by a fitness function that aims at optimizing the performance on a given task. We tested our technique by steering two state-of-the-art models through evolved prompts, and by comparing the performance they obtain on 8 benchmark tasks with that obtained when using other baseline prompts on the same tasks, showing that in most cases our prompts yield better results. Further, we defined a constrained mutation operator that limits the changes to specific grammar non-terminals, allowing to study and highlight the elements in the prompt that mostly affect the output of the LLM. Finally, a thorough discussion points out some issues that limit the relevance of the emerging prompt engineering discipline, given the existence of many effective prompt structures and the possible diversity that can be observed in the LLM output given the same input to the model.

Research in source code summarization, that is the description of the functionality of a program with short sentences expressed in natural language, is a topic of great interest in the software engineering community, since it can help in automatically generating software documentation, and in general can ease the effort of the developers in understanding the code they are working on. In this work, which is conceived as a negative results paper, we study the existing neural models designed for this purpose, pointing out their high sensitivity to the natural elements present in the source code (i.e. comments and identifiers) and the related drop in performance when such elements are ablated or masked. We then propose a novel source code summarization approach based on the aid of an intermediate pseudo-language, through which we are able to fine-tune the BRIO model for natural language on source code summarization, and to achieve results comparable to that obtained by the state-of-the-art source code competitors (e.g. PLBART and CodeBERT). We finally discuss about the limitations of these NLP-based approaches when transferred in the domain of source code processing, and we provide some insights for further research directions.

Neural networks for source code processing have proven to be effective for solving multiple tasks, such as locating bugs or detecting vulnerabilities. In this paper, we propose an evolutionary approach for probing the behaviour of a deep neural source code classifier by generating instances that sample its input space. First, we apply a grammar-based genetic algorithm for evolving Python functions that minimise or maximise the probability of a function to be in a certain class, and we also produce programs that yield an output near to the classification threshold, namely for which the network does not express a clear classification preference. We then use such sets of evolved programs as initial popu-lations for an evolution strategy approach in which we apply, by following different policies, constrained small mutations to the individuals, so to both explore the decision boundary of the network and to identify the features that most contribute to a particular prediction. We furtherly point out how our approach can be effectively used for several tasks in the scope of the interpretable machine learning, such as for producing adversarial examples able to deceive a network, for identifying the most salient features, and further for characterising the abstract concepts learned by a neural model.

Abstract An interesting thread in the research of Boolean functions for cryptography and coding theory is the study of secondary constructions: given a known function with a good cryptographic profile, the aim is to extend it to a (usually larger) function possessing analogous properties. In this work, we continue the investigation of a secondary construction based on cellular automata (CA), focusing on the classes of bent and semi-bent functions. We prove that our construction preserves the algebraic degree of the local rule, and we narrow our attention to the subclass of quadratic functions, performing several experiments based on exhaustive combinatorial search and heuristic optimization through Evolutionary Strategies (ES). Finally, we classify the obtained results up to permutation equivalence, remarking that the number of equivalence classes that our CA-XOR construction can successfully extend grows very quickly with respect to the CA diameter.