Niccolò Biondi

Assistant Professor (RTD-A) @ UniTn -- Representation Learning, Continual Learning, Compatible Learning

[Multimedia & Human Understanding Group (MHUG) @mhug_unitn]

🚀 CVPR, here we come!

The MHUG team is bringing a massive lineup of 17 papers this year, and we couldn't be more excited to share what we've been building! 🧠💻

🧵👇

[Simone Ricci @ricciunifi]

Excited to present our #NeurIPS2025 paper on "λ-Orthogonality Regularization for Compatible Representation Learning" today! 🎉

Come chat with me at Poster #2503 in Exhibit Halls C, D, E
📅 Fri, Dec 5 | 🕚 11am – 2pm PST
#MachineLearning #AI #Compatibility #RepresentationLearning

[Federico Pernici @FedPernici]

λ-Orthogonality: a tiny near-orthogonal adapter with a controllable λ "knob" to balance geometry preservation and adaptation. Enables backward-compatible embeddings (no full backfill) with near new-model performance and <50% gallery updates. @NeurIPSConf #NeurIPS2025

[Federico Pernici @FedPernici]

Encouraging to see compatibility discussed in this NeurIPS workshop. The theoretical understanding of the problem’s structure is still limited, with room for useful progress. @NeurIPSConf #NeurIPS2025
https://t.co/rvR943BM5o

[Niccolò Biondi @NiccoBio]

LeJEPA what else

[Randall Balestriero @randall_balestr]

LeJEPA: a novel pretraining paradigm free of the (many) heuristics we relied on (stop-grad, teacher, ...)
- 60+ arch., up to 2B params
- 10+ datasets
- in-domain training (>DINOv3)
- corr(train loss, test perf)=95%
Paper: https://t.co/NpfB9G1pOP
Code: https://t.co/BsK5wmNEHc

[Niccolò Biondi @NiccoBio]

Thanks! Scores are 5 5 4 4 : )

[Niccolò Biondi @NiccoBio]

🎉 Thrilled to share that our paper has been accepted to #NeurIPS2025 — more details coming soon!

@NeurIPSConf @miccunifi

[Aleksa Gordić (水平问题) @gordic_aleksa]

New in-depth blog post - "Inside vLLM: Anatomy of a High-Throughput LLM Inference System". Probably the most in depth explanation of how LLM inference engines and vLLM in particular work!

Took me a while to get this level of understanding of the codebase and then to write up this one - i quickly realized i understimated the effort. 😅 It could have easily been a book/booklet (lol).

I covered:

* Basics of inference engine flow (input/output request processing, scheduling, paged attention, continuous batching)

* "Advanced" stuff: chunked prefill, prefix caching, guided decoding (grammar-constrained FSM), speculative decoding, disaggregated P/D

* Scaling up: going from smaller LMs that can be hosted on a single GPU all the way to trillion+ params (via TP/PP/SP) -> multi-GPU, multi-node setup

* Serving the model on the web: going from offline deployment to multiple API servers, load balancing, DP coordinator, multiple engines setup :)

* Measuring perf of inference systems (latency (ttft, itl, e2e, tpot), throughput) and GPU perf roofline model

Lots of examples, lots of visuals!

---

I realize i've been silent on social - many of you noticed and thanks for reaching out! :) --> I'm so back! lots of things happened.

Also, in general, I'm a bit sick of superficial content, it really is an equivalent of junk food (h/t @karpathy).

I want to do the best/deepest technical work of my life over the next years and write much more in depth (high quality organic food ;)) so I might not be as frequent around here as i used to be (? we'll see). I'll make it a goal to share a few paper summaries a week or stuff that's relevant / in the zeitgeist.

If you have any topics that happened over the past few weeks/months drop it down in the comments i might focus on some of those in my next posts.

---

Huge thank you to @Hyperstackcloud for giving me an H100 node to run some of the experiments and analysis that i needed to write this up. The team there led by Christopher Starkey is amazing!

Also a big thank you to Nick Hill (who did a very thorough review of the post - basically a code review lol; Nick's a core vLLM contributor and principal SWE at RedHat) and to my friends Kyle Krannen (NVIDIA Dynamo), @marksaroufim (PyTorch), and @ashVaswani (goat) for taking the time during weekend when they didn't have to!

[Lele @EmanueleVivoli]

Calling all AI researchers & comic lovers! 🚀
Can your model pick the right comic panel? 🤖

Join the #ICDAR2025 Comics Understanding Competition & submit your results by 15/04/2025! 📆

Challenge: https://t.co/G5kHwUtkVi
Dataset: https://t.co/7HGYY0MNiU

Submissions are open!

[Marco Bertini @BertiniMarco]

Are you a professional working on video in Milan ? Come to the Milan Video Tech Meetup this Thursday (Nov. 7, 6:00 PM CET) at the NTTData headquarters, Via Ernesto Calindri, 4.
I'll talk about how to use GenAI to improve the performance of video codecs.
https://t.co/SZT3hQ4Ors

[Niccolò Biondi @NiccoBio]

🕝 The Learning Backward-Compatible tutorial kicks off at 2:30 pm in Room 205.
See you there!

#ACMMM2024 @miccunifi

[Niccolò Biondi @NiccoBio]

🚀 Heading to Melbourne for #ACMMM2024!

Join our Learning Backward-Compatible tutorial 🧩 on Oct 28th, where we survey compatibility methods with fresh interpretation in the era of LLMs & VLMs.

💻 Packed with code examples and practical insights!

See you there @ACMMM24 🇦🇺🎉!

[Niccolò Biondi @NiccoBio]

🚀 Heading to Melbourne for #ACMMM2024!

Join our Learning Backward-Compatible tutorial 🧩 on Oct 28th, where we survey compatibility methods with fresh interpretation in the era of LLMs & VLMs.

💻 Packed with code examples and practical insights!

See you there @ACMMM24 🇦🇺🎉!

[MICC- Media Integration and Communication Center @miccunifi]

🔔 MICC News from #ECCV2024!
The 18th European Conference on Computer Vision is live in Milan, Italy, from Sept 28 to Oct 4.

👏 Huge congrats to our MICC researchers showcasing their work in workshops, posters, and tutorials! #computervision #research #miccunifi #unifi #milan

[MICC- Media Integration and Communication Center @miccunifi]

👏Congratulations to MICC researchers presenting a poster "𝗔𝗿𝗲 𝗖𝗟𝗜𝗣 𝗳𝗲𝗮𝘁𝘂𝗿𝗲𝘀 𝗮𝗹𝗹 𝘆𝗼𝘂 𝗻𝗲𝗲𝗱 𝗳𝗼𝗿 𝗨𝗻𝗶𝘃𝗲𝗿𝘀𝗮𝗹 𝗦𝘆𝗻𝘁𝗵𝗲𝘁𝗶𝗰 𝗜𝗺𝗮𝗴𝗲 𝗢𝗿𝗶𝗴𝗶𝗻 𝗔𝘁𝘁𝗿𝗶𝗯𝘂𝘁𝗶𝗼𝗻?"@eccvconf  https://t.co/2rK7YWxfiK
#ECCV2024 #TWYN #QMUL #miccunifi

[Lele @EmanueleVivoli]

💥 But that’s not all! We'll be in Milan for #ECCV2024 with our work, "ComiCap: A VLMs Pipeline for Dense Captioning of Comic Panels" 🚀

This paper presents a VLM-based approach to generate dense captions for comic panels (no 🚂), and a new metric !
https://t.co/XmYYyRFxBt

[Lele @EmanueleVivoli]

🚀 Exciting news! Our paper "CoMix: A Comprehensive Benchmark for Multi-Task Comic Understanding" has been accepted at #NeurIPS2024 (B&D)!

It sets a new standard for multi-task comic analysis, please dive into the details here: https://t.co/EI3DXqQcpu, and: see you in Vancouver!

[Federico Pernici @FedPernici]

Apple🍏shows interest in strategies for evolving🔄compatible🤝language models that minimize inconsistencies with previous versions:
https://t.co/506wcW7pUD

[Lorenzo Agnolucci @AgnolucciL]

📢 Interested in low-level vision? Join our UHD-IQA challenge, held in conjunction with the AIM 2024 workshop at #ECCV2024 @eccvconf 🇮🇹

🚀 The top two contenders will win a PlayStation 5! 🎮

Challenge: https://t.co/kdxGQMowsK

Workshop: https://t.co/5hm99Wdl78

[Niccolò Biondi @NiccoBio]

@CVPR in three words?
Training’s over, guys!

#CVPR2024

[Federico Pernici @FedPernici]

Stop by board #447 in Arch 4E today, June Friday 21st, to see our🌟 highlight🌟poster🖼️:
"Stationary Representations: Optimally Approximating Compatibility and Implications for Improved Model Replacements".
📌Join us from 5:15-6:45 PM PDT! #CVPR2024