Postdoctoral Fellow @TechnionLive — Geometric Deep Learning in some of its various forms — PhD @imperialcollege — Previously @twitter, @fabula_ai and @polimi
Joined June 2019
- Tweets531
- Following560
- Followers2.5K
- Likes1.6K
Pinned Tweet
🧵"Neural Message Passing on Attention Graphs for Hallucination Detection" at #ICLR2026 !
🕸️We apply GNNs on the structured data LLMs produce as they generate text (e.g. attentions) to predict their errors.
📄 arxiv.org/abs/2509.24770
🤝 @GuyBarSh (co-1st) @YftahZ @HaggaiMaron
Fabrizio Frasca retweeted
GLOW is back tomorrow at 5 PM CEST!🌟
Great talks by Michael Ito on new approaches on graph canonicalization and by
Ioannis Kalogeropoulos on metanetworks as regulatory operators!
Join us on Zoom:
sites.google.com/view/graph-…
Hope to see you there!
#GraphLearningOnWednesday
Take a look at our new blogpost! ☺️
New blogpost out 📃
"Detecting LLM Misbehaviors from the Inside Out with Deep Learning on Structured Data" (ffabffrasca.substack.com/p/d…) [1/8]
Fabrizio Frasca retweeted
Check out our new ICLR 2026 paper - we explore hallucination detection through graph learning. Take a look!
🧵"Neural Message Passing on Attention Graphs for Hallucination Detection" at #ICLR2026 !
🕸️We apply GNNs on the structured data LLMs produce as they generate text (e.g. attentions) to predict their errors.
📄 arxiv.org/abs/2509.24770
🤝 @GuyBarSh (co-1st) @YftahZ @HaggaiMaron
✅ CHARM consistently outperforms leading approaches across benchmarks, spanning token- and response-level detection, and multiple hallucination types.
This paper is part of a broader direction on structure-aware learning for the detection of LLM misbehaviours:
- ACT-ViT (NeurIPS 2025): arxiv.org/abs/2510.00296
- LOS-Net (AAAI 2026): arxiv.org/abs/2503.14043
A blogpost is coming soon ☺️
See you in Rio 🇧🇷
Fabrizio Frasca retweeted
📄 Beyond Token Probes: Hallucination Detection via Activation Tensors with ACT-ViT
w/ @GuyBarSh , @ffabffrasca, Yaniv Galron, @YftahZ
arxiv.org/abs/2510.00296
Fabrizio Frasca retweeted
I respect that @iclr_conf had to respond to the OR leak, but I disagree with resetting scores. Many students worked hard on rebuttals and improved their papers in good faith. I hope the organizers reconsider and revert the reset. If you agree, feel free to retweet.
🌟 GLOW is back tomorrow, Oct 29th, 5pm CEST on Zoom!
- Celia Rubio-Madrigal will talk about feature-similarity-based rewiring (openreview.net/pdf?id=g6v09V…)
- Lorenz Kummer will show how to attack GNNs with bit-flips (dl.acm.org/doi/pdf/10.1145/3…)
Plenty of discussions as usual ☺️
GLOW is "so back" after the summer break!
Tomorrow: an interaction-heavy, plenary session on oversmoothing/squashing, homo/heterophily and their interplay.
Tune in! ☺️
(spoiler: prominent researchers on the topics will also join Adrián and Federico in the discussions 😇)
🌟Next GLOW session: Oversmoothing, “Oversquashing”, Heterophily, Long-Range, and more: Demystifying Common Beliefs in Graph ML
👤 with Adrián & Federico.
🔥 Discussion-heavy continuation of our last packed sessions.
📅 Sep 3rd, 5pm CEST on Zoom!
sites.google.com/view/graph-…
“Balancing Efficiency and Expressiveness: Subgraph GNNs with Walk-Based Centrality” is at #ICML2025!
Drop by our poster on Tuesday 11:00–13:30, East Exhib. Hall A-B # E-3005.
Joint work w/ @JoshSouthern13 @ytn_ym @GuyBarSh @mmbronstein @HaggaiMaron
arxiv.org/abs/2501.03113
🧵
5/
We seamlessly scaled HyMN to graphs with thousands of nodes, far beyond what full-bag Subgraph GNNs can handle.
HyMN is competitive with — and can outperform — sparse Graph Transformers at a fraction of their practical runtime.
6/
@JoshSouthern13 and I be at #ICML2025, poster session Tuesday — stop by and chat if you're around!
... I would also be happy to meet up and chat about graphs, (graphs and) LLMs, and how to detect their hallucinations 😳
Feel free to reach out!