Professor of Electrical and Computer Engineering at Tel Aviv University. On Sabbatical at Apple AI Research. Interested in AI, multimodal and signal processing
Joined September 2020
- Tweets459
- Following100
- Followers498
- Likes1.3K
Pinned Tweet
I’m excited to be spending my sabbatical at #Apple, collaborating with Vladlen Koltun and the team on a range of cutting-edge AI projects. Looking forward to this new chapter!
Raja Giryes retweeted
Give a VLM a few demos as text: ~100% accuracy.
As images, with no task hint: 45%.
VLMs don’t learn from what they see in context. We teach them to watch and learn 🧵
Paper: arxiv.org/abs/2610.30913
Dataset: huggingface.co/datasets/Nimr…
Raja Giryes retweeted
To channel my inner Schmidhuber, we found this softmax "bug" ~1.5 years ago and published in this paper arxiv.org/pdf/2603.11682
In reality, it's pretty hard to come up with a training scenario where this matters because gradient w.r.t. highly likely tokens is tiny anyway. Nonetheless I recommend fixing it -- one less thing to worry about! Check out @jon_barron 's work for an elegant fix!
Raja Giryes retweeted
(1/9) 🐍 Mamba Recall Scaling Laws: accepted to #NeurIPS2026! 🎉
How many facts can Mamba actually store and recall?
By reverse-engineering Mamba models, we reveal an interpretable hash-based recall mechanism, and theoretically derive precise scaling laws - predicting recall accuracy directly from model dimensions. Here's the story. 🧵👇
📄 Paper: arxiv.org/abs/2609.07681
💻 Code: github.com/yuvalko1/mamba-re…
Raja Giryes retweeted
(1/7)📢 New paper! Speech AI is getting better at hearing and speaking many languages. But what if the input mixes two languages, or is in one the model never learned? It guesses one language it knows and imposes that language's sounds on everything.
We find why, and fix it 🧵
Raja Giryes retweeted
SepGen extends a pretrained audio-video model to also output one waveform per source (alongside the video and mix) in a single pass. Sources stay individually accessible, so they can be localized and rendered as spatial audio in 4D. Dual generation + separation modes; strongest on speech.
Check out sepgen.github.io/
Raja Giryes retweeted
1/ Rendering a 4D audio-visual scene from a novel viewpoint requires every source as an individual waveform. Joint audio-video generators emit a single audio-mix in which the sources are not accessible.
We present SepGen - multi-stem audio-video separation and generation in a single model 🧵
Raja Giryes retweeted
I’ve spent the last decade working to make AI as powerful for 3D as it is for text and images - from industry, through my PhD, to leading my own research lab as a professor.
I started Thrixel to put that research directly into people's hands: a 3D creation platform for the agentic era that lets anyone build complete, interactive 3D worlds using AI.
It’s especially exciting to build Thrixel alongside people I’ve researched with for years. Together, we’re bringing fundamental research breakthroughs straight into our core engine.
Here’s a peek behind the scenes at Thrixel ↓
Thrixel is building the 3D creation engine that powers the agentic world. Born out of years of core 3D AI research, Thrixel aims to lower the barrier to entry to building in 3D and democratize the 3D creation process. Thrixel creates assets with named, editable parts, that human creators and AI agents can directly control across games, simulations, and a wide range of other applications.
See what we’re building: thrixel.com
Raja Giryes retweeted
We are hiring PhD level interns for efficiency of LLMs at . The topics include but not limited to speculative decoding, contextual sparsity, quantization, compression and on-device models. Zurich: jobs.apple.com/de-de/details… or Munich: jobs.apple.com/de-de/details…
Raja Giryes retweeted
In this era of artificial intelligence, it is becoming urgent to distinguish human art from what machines produce. There is an ontological difference, even before an aesthetic one, between art and what a machine can generate through statistical calculation based on millions of images created by others. Algorithms lack the spark of humanity. For this reason, the Church wishes to renew an alliance with artists and cultural institutions to safeguard our humanity.
Raja Giryes retweeted
(1/7) Humans can acquire many capabilities without (catastrophically) forgetting previous ones. Why can't neural networks do the same?
Local Support Learning (LSL) takes us a step closer towards this goal.
In LLMs of up to 7B parameters, LSL augments gradient-based training to retain prior capabilities while learning new tasks at full capacity. Notably, it does so without access to prior data.
Key idea: casting forgetting as a geometric problem, and optimizing for the worst case.
Paper: arxiv.org/abs/2610.02126
Website and code: assafbk.github.io/lsl
Raja Giryes retweeted
Amazing few days at #ECCV2026 presenting ID-LoRA!
Next stop: Finland! Presenting "Consistent Zero-Shot Video Editing with Multimodal Feature Flow" at #ICIP2026, oral spotlight Wed Sep 16, Duetto 2, Tampere Hall
w/ @eyal_gomel @liorwolf @RGiryes
Shanah Tovah! 🍯
Raja Giryes retweeted
I'll present CLIMP - The first fully mamba based CLIP on #ECCV2026 Friday, Sep 11, 10:30 AM – 12:30 PM, poster session 3, ExHall - poster # 65.
Come and check out how switching CLIP towers to mamba can do magic!
@RGiryes @Eli_Schwartz @ItamarZimerman
Happy to share CLIMP was accepted to ECCV 2026!
@eccvconf #ECCV2026
Check out our work if you want to see how Mamba models meet CLIP!
@ItamarZimerman @Eli_Schwartz @RGiryes
teapawt.net/NimrodShabtay/status/2…
Raja Giryes retweeted
You ask for a talking parrot, but the model makes the pirate talk AND gives him a beak! 🦜😮💨
-> Introducing The Attention Triangle in Audio-Video Models: how semantics leak between text, audio, and video, and how we steer them back. 🧵 [1/5]
Raja Giryes retweeted
This entire game was built using Claude + Thrixel, an agentic 3D creation platform.
Thrixel: Generated high-quality, thematically consistent, editable 3D assets.
Claude: Assembled in Unity + wrote game logic.
1 /goal prompt + Thrixel Skills + <5 follow-ups
Repo coming👇
Raja Giryes retweeted
נכון שהילדים שלכם תמיד רצו לדעת מה מדענים עושים במעבדות שלהם חוץ מלהצטלם עם משקפי מעבדה וחלוק? אז האקדמיה הצעירה עשתה משהו סוף! היא אספה כמה חוקרים שיספרו על המחקרים שלהם לילדים, יש שם את עודד רכבי המלך ואת איגור אוליצקי המעולה, וארז בן יוסף ונעמה, וכולם חוקרים מעולים! ומעניינים>
Is it the one extra day of rest that helped Spain win?
Since 2010, 14 of the last 15 men’s and women’s World Cup and European Championship finals have been won by the team that played its semi-final a day earlier.
Interestingly, the only previous exception was… Spain in 2010.
Raja Giryes retweeted
Work w/ @moranynk @RGiryes @begusgasper
Check out our project page for more info and examples of auto-conlangs: conlangcrafter.github.io
Raja Giryes retweeted
ConlangCrafter is in the news! Check out the recent coverage of our work in Science and IEEE Spectrum (in thread) -- great to see people are interested in how AI can help us create constructed languages!
Raja Giryes retweeted
Happy to share our recent TMLR publication: "Jacobian-Aware Posterior Sampling for Inverse Problems" 🥳
JAPS is based on a clever discovery and achieves SOTA results!
arXiv: arxiv.org/abs/2511.18471
OpenReview: openreview.net/forum?id=m63G…
code: github.com/liavhen/JAPS
