Royi Rassin

PhD candidate @biunlp researching multimodality. Intern @GoogleAI

[Royi Rassin @RoyiRassin]

How diverse are the outputs of text-to-image models and how can we measure that? In our new work, we propose a measure based on LLMs and Visual-QA (VQA), and show NONE of the 12 models we experiment with are diverse. 🧵
1/11

[Guy Dar @guy_dar1]

A potential way to do Aurora with a closed-form construction that avoids the expensive iterative step.

Link in comments

[Emmy Liu @_emliu]

Midtraining is a new part of many training pipelines, but when does it help and can it backfire? 🤔

In our new preprint, we use controlled experiments to pin this down. TL;DR; midtraining helps the most when it “bridges” pretraining and posttraining, and mitigates forgetting after posttraining. Timing is also very important.
🧵

[Life at Google @lifeatgoogle]

Aviv, a research scientist with Google Earth AI, walked us through how his work provides first responders, crisis planners, and others with the information they need to save lives. Interested in similar roles? Explore our open AI/ML jobs ➡️ https://t.co/8iZOsFzuTd

[Royi Rassin @RoyiRassin]

Congrats to the BRIA team! Fibo looks amazing, I can’t wait to try it

[Ron Mokady @ CVPR @MokadyRon]

Generating an image from 1,000 words.

Very excited to release Fibo 😃, the first ever open-source model trained exclusively on long, structured captions.

Fibo sets a new standard for controllability and disentanglement in image generation

[1/6] 🧵

[Shauli Ravfogel @ravfogel]

New NeurIPS paper! 🐣Why do LMs represent concepts linearly? We focus on LMs's tendency to linearly separate true and false assertions, and provide a complete analysis of the truth circuit in a toy model. A joint work with @Giladude, @tallinzen, Joan Bruna and @albertobietti.

[Ilya Sutskever @ilyasut]

truly the greatest day ever🎗️

[Guy Dar @guy_dar1]

NEW PAPER
Scaling up vec2vec!
vec2vec holds an amazing promise for unsupervised alignment of embedding models without matching pairs of data.

But it is very costly!
What can we do?!?!

In this paper, I investigate a simple method to learn this alignment in just 10 minutes!!

[Merriam-Webster @MerriamWebster]

We are thrilled to announce that our NEW Large Language Model will be released on 11.18.25.

[Royi Rassin @RoyiRassin]

@lovodkin93 Congrats 🥳

[Guy Dar @guy_dar1]

A project based on the celebrated elegant vec2vec ("Harnessing the Universal Geometry")!

one can get competitive results *at a fraction* of the compute, and samples, *using linear methods*!

I use a cocktail of combinatorial matching algorithms, lin algbera, and some tricks >>

[Ron Mokady @ CVPR @MokadyRon]

Open-sourcing a model is not enough—it has to be accessible to be useful
That’s why I’m excited our first model is natively supported in @diffuserslib (my favorite repo 🤗)
(Link below)

A small step, but important prep for what’s next 😃

[Royi Rassin @RoyiRassin]

@yanaiela Lucky us 🥳

[Aviya Maimon @AviyaMaimon]

🚨 New paper alert! 🚨
We propose an IQ Test for LLMs — a new way to evaluate models that goes beyond benchmarks and uncovers their core skills.

Think: 🧠🤖 psychometrics for LLMs.
👇
(1/6)

[Ron Mokady @ CVPR @MokadyRon]

Here is my short and practical thread on how to teach your Text-to-Image model to generate readable text:

[1/n]

[Royi Rassin @RoyiRassin]

@yoavgo @SenougaharA Then you can hide content from it specifically

[Royi Rassin @RoyiRassin]

@guy_dar1 @yoavgo The jokes on you

[(((ل()(ل() 'yoav))))👾 @yoavgo]

IRGC, no matter what you do, please do not attack the compute cluster at Bar-Ilan university. It is priceless and impossible to replace. We will be devastated if it will be destroyed. (it also has tons of super sensitive and irreplaceable military stuffs!!!!)

[Royi Rassin @RoyiRassin]

@yoavgo לא כל הספרים שלו אימה. זה אפוקליפטי ונחשב אולי העבודה הטובה ביותר שלו

[Luca Ambrogioni @LucaAmb]

1/2) It's finally out on Arxiv: Feedback guidance of generative diffusion models!

We derived an adaptive guidance methods from first principles that regulate the amount of guidance based on its current state.

Complex prompts are highly guided while simplem ones are almost free