A code smell is a surface indication that usually corresponds to a deeper problem in the system. Detecting and removing code smells is crucial for sustainable software development. However, manual detection can be daunting and time-consuming. Machine learning (ML) is a promising approach towards the automation of code smell detection. The first ML-based methods were classifiers trained on feature vectors comprising software metrics extracted by off-the-shelf tools. Determining the optimal set of metrics is a complex problem that requires both ML and software engineering expertise. Recently source code embedding models emerged as a viable feature-inferring alternative. However, their potential is yet to be fully explored. To that aim, we compare state-of-the-art source code embedding models (CuBERT and CodeT5) with the models trained on metrics returned by the CK Tool and RepositoryMiner tools. We focus on detecting the Data Class and Feature Envy code smells within a large-scale, manually labeled, publicly available dataset. After extensive experiments (51 test/train splits), we found that source code embedding models have comparable performances with software metrics, a that they indeed can capture important characteristics of the source code. We discuss our findings in detail in the paper.

Annotated medical imaging data remain scarce, and labels are often weak and noisy: in mammography, an examination comprises several high-resolution views, yet the diagnostic outcome is recorded only at the breast level. Such problems are naturally cast as Multiple Instance Learning (MIL), where the model must infer instance-level structure from bag-level labels alone. Although contemporary foundation encoders supply strong general-purpose embeddings, augmenting MIL data in this representation space remains an open problem as established techniques act on one instance at a time and ignore the statistical dependencies binding a bag together. We address this with SetFlow, a generative model that learns the distribution of complete MIL bags directly in a frozen encoder’s embedding space. SetFlow couples flow-matching training with a Set Transformer-inspired backbone, making it invariant to instance ordering while modeling intra-bag relationships. Generation is conditioned jointly on class label and per-instance scale, yielding coherent, semantically faithful bags rather than isolated vectors. Evaluating on two large public mammography datasets and two encoders, we assess distributional fidelity, nearest-neighbor behavior, and downstream augmentation utility. We show that generated bags reproduce real-data statistics and improve classification in certain configuration, with performance gains varying on the amount of synthetic data. An architecture ablation confirms each design choice contributes to performance.