Publications

Below is a complete list of my publications.

References

Benchmarking graph neural networks in solving hard constraint satisfaction problems

Skenderi, G., Buffoni, L., D’Amico, F., Machado, D., Marino, R., Negri, M., Ricci-Tersenghi, F., Lucibello, C., & Angelini, M. C.

2026. arXiv preprint arXiv:2602.18419

TL;DR: We constructed a publicly available principled dataset to benchmark graph neural networks performance on hard optimization problems and compare them with some algorithms. We show that most advanced network existing today have worst performance with respect to simple algorithms.

Pseudo-likelihood produces associative memories able to generalize, even for asymmetric couplings

D’Amico, F., Bocchi, D., Del Bono, L. M., Rossi, S., & Negri, M.

2026. Physica A: Statistical Mechanics and its Applications, 131497. doi: 10.1016/j.physa.2026.131497

TL;DR: We studied pseudo-likelihood/cross-entropy training in a two-bodies model. We show why it behaves as an associative memory, and that at training time it interpolates between classical learning rules. We show that such model is capable of generalization and that it works even if couplings are not symmetric.

Algorithmic thresholds in combinatorial optimization depend on the time scaling

Angelini, M., Avila-González, M., D’Amico, F., Machado, D., Mulet, R., & Ricci-Tersenghi, F.

2025. Phys. Rev. X, DOI 10.1103/dw9m-95vv

TL;DR: We studied the scaling behaviour with training time of algorithmic thresholds of simulated Annealing algorithm in k-SAT and Q-col problems.

Implicit bias produces neural scaling laws in learning curves, from perceptrons to deep networks

D’Amico, F., Bocchi, D., & Negri, M.

2025. arXiv preprint arXiv:2505.13230

TL;DR: We show with a simple Perceptron theory how the norm of a model can be used as a useful parameter to predict its scaling laws. We show with extensive simulations that also deep-networks in real datasets classification show the same behavior.

Statistical mechanics of vector hopfield network near and above saturation

Nicoletti, F., D’Amico, F., & Negri, M.

2025. J. Phys. A: Math. Theor. 58 505005, DOI 10.1088/1751-8121/ae2bd0

TL;DR: We studied the Vector Hopfield network with Hebb couplings in arbitrary number of dimensions, computing the equilibrium phase diagram and discovering an out-of-equilibrium denoising behaviour absent in standard hopfield model.

Self-attention as an attractor network: Transient memories without backpropagation

D’Amico, F., & Negri, M.

2024. 2024 IEEE workshop on complexity in engineering (COMPENG), 1–6

TL;DR: We show that a simplified self-attention model trained directly via pseudo-likelihood minimization behaves as an associative memory at the very first step of dynamics.