←
Itay Evron586 posts
Itay Evron avatar

Itay Evron

@itayevron

AI Research Scientist @nvidia; PhD; Science is everyone's but opinions are my own

◷ Joined July 2018▧ 60 Photos and videos
Itay Evron avatar
Itay Evron@itayevron·16 Nov 2020

Why does L1 regularization induce sparse models? Many illustrate this using the least squares problem with a norm constraint. The least squares level sets are drawn next to the different unit "circles". I prepared a cool animation which I believe makes it even clearer 🙂

asyaf avatar
⟳ Itay Evron retweeted
asyaf@asya_fru·Nov 16

כנראה הזדמנות טובה לספר שגם אליי הגיע אתמול טרנד הפיטורים (לראשונה אחרי 15 שנה בתחום). הצעות למשרות מעניינות של הובלת AI וצוותי AI/DS יתקבלו בברכה בDM.

קודם פיטרו את בוגרי הבוטקאמפ, ושתקתי, כי היה לי תואר. "קורס של שמונה חודשים לא הופך אף אחד למהנדס." אחר כך פיטרו את הג'וניורים, ושתקתי, כי הייתי סיניור. "אין להם עומק. זה תיקון בריא של השוק."

Itay Evron avatar
Itay Evron@itayevron·Nov 9

Reviewed a paper for NeurIPS 2026. Pointed out several serious problems, recommended rejection. Indeed rejected. Now it's on OpenReview for ICLR 2027. Barely changed. Just resubmitted. Bad ratings again. What an utter waste of time. There should be a cost to this.

Itay Evron avatar
Itay Evron@itayevron·Nov 8

Searched ICLR’s OpenReview for interesting papers based on their titles and abstracts. Found about two dozen. To decide which ones to read, I first looked at their review scores. Only one had a review rated 3/4. The rest were all 1s or 2s out of 4. Read the news instead.

Jiaxin Pei avatar
⟳ Itay Evron retweeted
Jiaxin Pei@jiaxin_pei·Nov 6

My biggest struggle with reviews in 2026 is that (AI) reviewers don’t seem to know the difference between limitations and critical issues. An LLM can easily find 100+ limitations in any paper (even the best ones in history), but that doesn’t mean the paper shouldn’t be accepted.

Hen Davidov avatar
⟳ Itay Evron retweeted
Hen Davidov@hendav136·Nov 6

1/ Our paper “Selective Safety Steering via Value-Filtered Decoding” was accepted at #NeurIPS2026 🎉 Most safety interventions change every response, even though most responses were already safe. Value-Filtered Decoding only touches the ones heading somewhere unsafe.

1/ 🧵 Decoding-time safety methods improve LLM safety, but they rewrite *safe* generations too, an "alignment tax" on helpfulness. Value-Filtered Decoding steers only when needed, with a calibrated bound on how often we touch safe outputs. 📄Accepted at AI4Good @ ICML 2026✨

Post attachment
Maziyar PANAHI avatar
⟳ Itay Evron retweeted
Maziyar PANAHI@MaziyarPanahi·Oct 16

WE ARE SO BACK! 🔥 nvidia’s new open model posted the top score on Clinical System One nemotron 3.5 super vl: 647 of 669 decisions, zero severe errors, tied #1 with pplx-decider and gpt-6 luna the top 2 are now both open weights! SO COOL!

Post attachment

Meet NVIDIA Nemotron 3.5 Super VL Our new open model combines reasoning and tool use with text, image, document, and video understanding for long-running AI agents. 120B total parameters with 12B active 🧵

1:20
Itay Evron avatar
Itay Evron@itayevron·Oct 27

None. If it was fun, then it was not a waste of time.

So if it ends up P=NP, what percentage of Complexity Theory in Computer Science was kind of a waste of time?

Edward Milsom avatar
⟳ Itay Evron retweeted
Edward Milsom@edward_milsom·Oct 23

Are you interested in a PhD in deep learning scaling and optimisation? I'm seeking students to join me at University of Bath, in order to expand the literature on automatic / measurement-based approaches to hyperparameter tuning. See below for motivation. findaphd.com/phds/project/hy…

Our paper "Function-Space Learning Rates" is on arXiv! We give an efficient way to estimate the magnitude of changes to NN outputs caused by a particular weight update. We analyse optimiser dynamics in function space, and enable hyperparameter transfer with our scheme FLeRM! 🧵👇

Post attachment
Gon Buzaglo avatar
⟳ Itay Evron retweeted
Gon Buzaglo@gon_buzaglo·Oct 12

In nonsmooth nonconvex stochastic optimization, existing results suggested that constraints make the problem harder. In a new paper with Guy Kornowski, we show the oracle complexity is in fact unchanged.

Post attachment
Elad Hoffer avatar
⟳ Itay Evron retweeted
Elad Hoffer@eladhoffer·Oct 8

RAG and long context are usually treated as rival approaches, but both pick the evidence that matters, just at different scales. What if one LLM did both? UNREAL: one frozen LLM retrieves from a 3B-token corpus and filters its own 128K prompt. arxiv.org/abs/2610.08463

Post attachment
Nicolò Cesa-Bianchi avatar
⟳ Itay Evron retweeted
Nicolò Cesa-Bianchi@NicoloCB·Oct 7

My dream had always been to prove beautiful theorems as the AI agents unloaded my dishwasher. Well, it went the wrong way... 😂