Davide D'Ascenzo

Davide D'Ascenzo

Co-founder, Untoken

I study learning as a phenomenon. I build architectures and algorithms to get closer to how intelligence and understanding work.

I co-founded Untoken, where we build small models that think in concepts and understand before they act. Before Untoken I did a PhD in Artificial Intelligence at Politecnico di Torino and the University of Milan. During the PhD I worked in Tomaso Poggio's group at MIT and at Télécom Paris.

I think a model generalizes when it captures the structure of its problem, and that scale alone does not get it there. Most of the work below starts from this view.

Research

scDataset

ICML 2026

A PyTorch data loader for datasets too large for memory. It reads blocks of contiguous samples and fetches several minibatches at once, so training gets shuffled data without one random read per sample.

Hierarchical cross-entropy loss

Nature Computational Science 6, 243–249

When labels form a hierarchy, standard cross-entropy treats a prediction that is more specific than the label as simply wrong. This loss scores each class together with its subclasses, so any prediction inside the label's subtree counts.

Compositional sparsity

ICML 2025, Position Paper Track

Deep networks learn functions of many variables without the exponential cost the curse of dimensionality predicts. We argue the reason is that the functions worth learning are built from pieces that each depend on a few inputs.

Monotonicity of centrality measures

Social Network Analysis and Mining 14, 183

When you add an edge to an undirected network, one of its endpoints can lose rank. We prove that the two can never both lose it, for closeness, betweenness and distance-decay centralities.

Provenance of LLM-generated code

Journal of Systems and Software 245, 113147

Code models sometimes reproduce snippets from their training data without saying where they come from. We find the source by shortlisting candidates with vector search and re-ranking them with code fingerprints.

Papers

Talks