Roberto Bifulco

Our novel training method learns #neuralnetwork layers’ width automatically, reducing tuning costs and enabling efficient compression for large‑scale foundational models. Learn how in our #ICLR26 accepted paper: #publications-2182" target="_blank" rel="nofollow noopener">neclab.eu/research-group…. #NECLabs

Our latest work in #machinelearning and #neuralnetworks is highlighted in three papers recently accepted at the International Conference on Learning Representations (@iclr_conf) 2026. Learn more: neclab.eu/research-group…. #NECLabs

Our latest work explores how #LLMs can automate CVSS risk scoring, improving #riskassessment speed and scalability even when labelled data is scarce. Learn how in our #EMNLP2025 paper: #publications-2165" target="_blank" rel="nofollow noopener">neclab.eu/research-group…. #NECLabs

🚨I'm beyond ecstatic to announce that we've raised $6M in seed funding to launch Unikraft Cloud -- the only platform that can start *any* workload in < 10ms and scale to 100K+ instances on a single server with instantaneous scale to zero...think catering to millions of users on a few servers instead of an entire data center. All of with strong, hardware-level isolation of course! 🔥 A huge shut out to our amazing investors @heavybit @vercel Ventures, Mango Capital, @Firestreakvc , @FlyVC and @_firstmomentum for believing that a group of tech geeks could build a fundamentally different and exponentially better cloud platform -- that dream and vision is now reality. 🙏 To my co-founders @nderjung and @s_kuenzer , an immense pleasure to be sharing this amazing adventure with you, and now the sky's the limit! 🚀 Oh, and if you're unhappy with and disillusioned with the way the cloud works, here's a message of hope: there's better out there, come check us out! (Link to Business Wire release article in the first comment below 👇)

Our latest research presents a framework that enhances deep #graphnetworks by optimizing message depth and filtering, effectively modeling long-range interactions while overcoming limitations like oversmoothing and underreaching. Learn how: #publications-2141" target="_blank" rel="nofollow noopener">neclab.eu/research-group…. #NECLabs #ML

Our new physics-informed method for training machine-learned #interatomicpotentials addresses atomistic simulation challenges by improving robustness & accuracy, reducing errors & minimizing data needs. Learn how in our #ICML2025 accepted paper: #publications-2143" target="_blank" rel="nofollow noopener">neclab.eu/research-group…. #NECLabs

In computer science, you need to know the system that you’re building on top to have any real leverage. You need to know what will and won’t work, and why. This also generalizes to most fields, which is precisely why AI will take far fewer jobs than people think.

@jxnlco We did it for you 😆

@RoBBifulco Omg should I be building a bibtex for instructor

Federico Errica is presenting at NAACL25 two metrics we use to assess the quality of LLM prompts: sensitivity and consistency. We started the work after seeing how LLMs changed behavior when changing variable names in our pydantic classes (we were using @jxnlco’s instructor)

We also had fun building a small integration with @langfuse for data collection and then to feed data to a @streamlit dashboard. It was an easy and convenient way to quickly test our work and make it more usable

We apply these metrics in classification tasks. Sensitivity tells us how rephrasing of the prompt affects the LLM output. Consistency tells how an element of a class is affected by rephrasing.

Congratulations to B. Bütün, D. de Andrés Hernández, M. Gucciardo & M. Fiore from @NECLabsEU, @IMDEA_Networks & @uc3m on winning #BestPaperAward at INFOCOM 2025 for DUNE – a framework for deploying #ML models in #network user planes! Learn more: #publications-2063" target="_blank" rel="nofollow noopener">neclab.eu/research-group…. #NECLabs

Our new benchmark – AL4PDE – uses active learning to solve partial differential equations more efficiently with #neuralnetworks. It reduces data needs and lowers errors by up to 71% compared to random sampling. Learn how: #publications-2076" target="_blank" rel="nofollow noopener">neclab.eu/research-group…. #NECLabs #AI #MachineLearning

Pandas is dying a slow, painful death. It's the world's most popular data library, but it's slow, and many libraries have significantly improved over it. The problem with many of these alternatives is that nobody wants to learn a new API. Let's face it: people won't migrate their codebase unless they have to. I heard about FireDucks recently. It's up to 48x faster than Pandas, and you don't have to touch your code. Well, there are two ways. You can change *one* line of code and get everything else run as-is: > import fireducks.pandas as pd You can also run your code *without* changing a single line by using an import hook: $ python -mfireducks.imhook yourfile[.]py FireDucks is a multi-threaded, compiler-accelerated library with a fully compatible pandas API. It's also faster than Polars. Below is a link to some benchmarks that compare Pandas, Polars, and FireDucks. FireDucks wins, hands down.

Federico Errica, Giuseppe Siracusano, Davide Sanvito, Roberto Bifulco: “What Did I Do Wrong? Quantifying LLMs’ Sensitivity and Consistency to Prompt Engineering”, accepted at NAACL 2025: neclab.eu/fileadmin/user…

"What Did I Do Wrong? Quantifying LLMs' Sensitivity and Consistency to Prompt Engineering' will appear in NAACL. Thanks to Federico and all co-authors for the fun work (Link👇) You can guess what is a regular source of frustration in the modern software development process...

@martin_casado I have been missing YAIG, I’d love to join… hopefully you’ll eventually see this 😅

Hey infra folks. We're standing up a new Discord server to discuss CS infra. If you want an invite DM me (reply and I'll follow). thanks!

💯

Congrats to @ViktorZaverkin and the team 🍾 I loved how this paper included a larger collaboration within the group, down to the platform-level bits (and bytes 😉)