Publications

Conference & Workshop Papers

1. [MGB26a]
   L. Miglior, V. Gervasi, and D. Bacciu. CompFlow: Composing velocity fields for multi-condition generation. 2nd Workshop on Compositional Learning (CompLearn) at ICML 2026, Seoul, South Korea, 2026. pdfopenreviewproject page A training-free flow-matching framework for compositional inference. Conditional velocity residuals are proportional to gradients of noised conditional likelihoods, so their weighted addition implements a tempered Product-of-Experts, generalising classifier-free guidance from one condition to arbitrarily many. On CLEVR, a model trained only with single-object conditioning composes shape, color and position for up to five objects at 99.1–86.5% per-object accuracy, using 30× fewer network evaluations than prior baselines. Applied zero-shot on top of FLUX.1[dev], it improves the Share-CoT score on the T2I-CompBench non-spatial split from 0.7809 to 0.8021.
2. [MTGB26]
   L. Miglior, M. Tolloso, A. Gravina, and D. Bacciu. Can you hear me now? A benchmark for long-range graph propagation. The Fourteenth International Conference on Learning Representations (ICLR), Rio de Janeiro, Brazil, 2026. arXivdoi A benchmark for evaluating how far graph neural networks actually propagate information, rather than inferring it from downstream accuracy.
3. [MSPB25]
   L. Miglior, L. Simone, M. Podda, and D. Bacciu. Towards efficient molecular property optimization with graph energy based models. Proceedings of the 33rd European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN), pp. 289–294, 2025. pdfarXivdoi Energy-based models over molecular graphs for efficient, implicit property optimization.
4. [VMG⁺23]
   E. Vignali, L. Miglior, E. Gasparotti, et al. An application for remote respiratory patterns detection in post-surgery patients. Proceedings of the 28th Congress of the European Society of Biomechanics, 2023. handle

Preprints

1. [MGB26b]
   L. Miglior*, A. Gravina*, and D. Bacciu. Stable transformers for graph generation. arXiv preprint arXiv:2609.39739, 2026. *Equal contribution. pdfarXiv Graph generative models increasingly rely on Graph Transformer denoisers, but deeper is not automatically better: from a dynamical-systems perspective, repeated self-attention becomes increasingly dissipative with depth, contracting node representations and causing vanishing gradients. We replace the attention map with a Cayley-based orthogonal node-mixing operator (a permutation-equivariant, non-dissipative drop-in for existing GT denoisers), plus a damping knob that interpolates towards contractive regimes. On synthetic and molecular discrete flow-matching benchmarks the gap widens with depth: contractive denoisers degrade and eventually collapse, while non-dissipative ones keep benefiting from added layers.

Journal Articles

1. [SMG⁺25]
   L. Simone, L. Miglior, V. Gervasi, L. Moroni, E. Vignali, E. Gasparotti, and S. Celi. Deep learning framework for cardiorespiratory disease detection using smartphone IMU sensors. Computers in Biology and Medicine, 196:110595, 2025. doiarXiv A deep learning framework that classifies cardiorespiratory conditions from smartphone inertial sensor signals. The earlier preprint is titled On the effectiveness of smartphone IMU sensors and Deep Learning in the detection of cardiorespiratory conditions. Supported by Regione Toscana (CUP I55F20000610007).
2. [VGM⁺24]
   E. Vignali, E. Gasparotti, L. Miglior, V. Gervasi, L. Simone, D. Haxhiademi, L. Frediani, G. Borelli, S. Berti, and S. Celi. A new smartphone-based method for remote health monitoring: Assessment of respiratory kinematics. Electronics, 13(6):1132, 2024. pdfhtmldoi An acquisition protocol and Android application for measuring chest and abdomen respiratory kinematics from smartphone IMU sensors, validated on 77 healthy volunteers. Open access (CC BY); Feature Paper in the special issue Wearable and Implantable Sensors in Healthcare.