News

📍 Rome, Italy

September 2026 - Cambridge Talk

I gave a talk entitled “RAG Beyond NLP: Learning Relevant Context for Cellular Prediction” at GMTD 2026, University of Cambridge.

I presented our work on differentiable retrieval-augmented generation for predicting cellular responses to gene perturbations, discussing how models can learn to retrieve relevant external knowledge through the prediction task itself. The talk explored the potential of RAG beyond natural language processing, with a focus on cellular modeling and its applications to therapeutic discovery.

September 2026 - NeurIPS

Main track paper accepted at NeurIPS

Following the acceptance of PT-RAG: Retrieval-Augmented Generation for Predicting Cellular Responses to Gene Perturbations at an ICLR workshop, we extended our approach to additional model backbones and more challenging tasks, evaluating generalization to both unseen cells and unseen perturbations.

Our experiments showed consistent results across models and tasks. This extended work, “Differentiable Retrieval-Augmented Generation for Predicting Cellular Responses to Gene Perturbation” has now been accepted to the NeurIPS main track as a poster.

I’ll be in Sydney to present the paper in person!

June 2026 - Token Collapse Talk at ISTI-CNR, Pisa

On June 3, I gave a talk entitled 'Graph-NeuralRAG: On the Opportunities and Challenges of GNNs for GraphRAG' at ISTI-CNR, sharing my work with colleagues and collaborators from the HPC Lab and other research groups across the institute.

I discussed the challenges of aligning GNN representations of large knowledge graphs with the token space of LLMs, including a phenomenon we call token collapse, where graph representations become nearly identical across queries.

April 2026 - ICLR 2026

I have been at the Gen² Workshop @ ICLR 2026!🧬 in RIO

Happy to have finally presented my work (2 papers) at a Top-Tier Conference!

PT-RAG: Retrieval Augmented Generation for predicting cellular responses to gene perturbations.
GGE: A Standardized Framework for Evaluating Gene Expression Generative Models.

April 2026 - GraphNeuralRAG Talk at University of Cambridge

On April 9th, I gave a talk at the Computer Lab of the University of Cambridge entitled: Graph Neural RAG: On the Opportunities and Challenges of GNNs for GraphRAG, from Multi-Hop Question Answering to Perturbation Modelling.

I discussed the line of research I worked on in the past months here at the University of Cambridge. In particular, going from past works on GraphRAG and Knowledge Graph Q&A with GNNs, I show that there is a concrete opportunity to leverage GNNs as the de facto standard to empower GraphRAG systems. This derives directly from GNNs being naturally suited to learn on graphs. I show how such a connection is possible, and what the advantages and concrete limitations are that come from this GNN-based approach.

For instance, I also propose how these limitations can be plausibly overcome, and the related Research Opportunities for GNNs, GraphRAG, and Graph Foundation Models communities.

This was my first Talk, and I am honored to have given it at such a prestigious University.

Slides and Video.

March 2026 - 2/2 Accepted Papers at ICLR 2026

Excited to share that I got accepted at the Gen² Workshop @ ICLR 2026!🧬 2/2.

These papers are part of a new line of research I'm pursuing here at Cambridge, sitting at the intersection of generative modeling, retrieval-augmented systems, and single-cell biology.

PT-RAG: Retrieval Augmented Generation for predicting cellular responses to gene perturbations.

We use RAG for a completely different modality, namely Gene Expressions. We introduce the first differentiable RAG framework for predicting cellular responses to gene perturbations. We show that Naive RAG hurts performance while cell-type conditioning, and end-to-end differentiable retrieval leads to statistically significant improvements over classic baselines.

GGE: A Standardized Framework for Evaluating Gene Expression Generative Models.

How do you know if your generative model for single-cell data is actually better? Turns out, you often can't — because "Wasserstein distance" can mean completely different things depending on implementation choices, making cross-paper comparisons essentially meaningless. GGE fixes this with a unified, open-source evaluation framework that makes every choice explicit, so the field can finally compare methods fairly.

September 2025-April 2026 - Visiting PhD student at the University of Cambridge

I am currently a visiting PhD student at Cambridge, actively collaborating with Professor Pietro Liò’s research group. While we are exploring various domains, my primary focus is developing a foundation model for GraphRAG using Graph Neural Networks.

January 2026-March 2026 - Teaching Assistant for the Geometric Deep Learning course

I am working as a TA for the GDL course (L65) at the University of Cambridge, taught by Prof. Pietro Liò and Dr. Petar Veličković. In this role, I formulated project proposals and currently mentoring four students, on topic about efficient GNN design, link prediction, LLMs, and robotics applications.

July 2025 - London Geometry and Machine Learning Summer School

From July 7–11, I had the opportunity to attend LOGML25 Summer School, hosted at the Imperial Colledge of London, a highly selective program focusing on the intersection of mathematics, geometry, topology, and their role in Machine Learning. Coming from an engineering background, I found it particularly enriching to deepen my understanding of mathematical concepts that I can now integrate into my research.

As part of the program, I worked on a project entitled Iterative Reasoning in Graph Neural Networks for Drug Repurposing. This was a valuable chance to apply GNNs to a practical task and, importantly, to step outside my comfort zone by exploring applications in biology and medicine, fields I had not previously worked in. Updates on this project will be shared soon!

The Summer School also offered the opportunity to present my latest preprint, Early-Exit Graph Neural Networks, during the poster session.

June 2025 - International Joint Conference on Neural Networks

From 30 June to 5 July 2025, i had the chance to attend IJCNN25 in Rome, at the Pontificia Università Gregoriana as my first time as an author. Having 3 accepted papers at the conference, i had more than one occasion to present my work in front of an audience. I presented

Link Prediction with Physics-Inspired Graph Neural Networks;
GATSY: Music Artist Similarity with Graph Attention Networks;
IJCNN 2025 Competition: Learning with noisy graph labels;

Beyond presenting, the conference was a great opportunity to discover cutting-edge research and engage in valuable networking with fellow researchers.

June 2025 - Visiting PhD student at the University of Cambridge

Starting 2nd June 2025, i began my period abroad for my Ph.D. I work at the Computer Laboratory of the University of Cambridge under the supervision of Professor Pietro Liò. My research revolves around the intersection of GraphRAG and Graph Neural Networks.

Page updated

Report abuse