Structured Neural Intelligence Lab
Efficient, explainable, and structured neural networks in science and technology, from high-energy physics to telecommunications, medicine, and earth observation.
A research group at Sapienza University of Rome, led by Simone Scardapane.
Two main-conference papers were accepted at ICLR 2026: activation sparsity and MASS. We also have a workshop paper on hypergraph neural network liftings.
"Attention Sinks in Diffusion Language Models" was accepted to Findings of ACL 2026.
Research
Networks that adapt their own structure and computation to each input, so they run faster and cost less. Current work includes, e.g., conditional computation, compute-bound training, and KV cache compression.
Ways to understand what a model has learned and to change its behaviour after training, including model merging and editing.
Training models that keep learning from new data without forgetting the old by working on the geometry of the latent representations.
Building modular and efficient agents (modular skills, controllable behaviour, memory) and studying how they coordinate to reach a goal (we are not there yet...).
Learning on data that has structure: graphs, simplicial and cell complexes. We co-developed TopoBench, and contributed to the field in a joint position paper.
We work with collaborators in high-energy physics, telecommunications, medicine, and earth observation.
People
Where to find us
You can (generally) find us in two departments in central Rome.
Join us
We are always looking for new PhD students, thesis students, and collaborators. Interested? Send us an email!