Alibaba's Qwen3-Omni — the end-to-end omni-modal AI unifying text, image, audio & video in one model 🏆 SOTA on 22/36 audio & AV benchmarks 🌍 119L text / 19L speech in / 10L speech out ⚡ 211ms latency | 🎧 30-min audio understanding 🎨 Fully customizable via system prompts

Large Vision Models Can Solve Mental Rotation Problems Sebastian Ray Mason, Anders Gjølbye, Phillip Chavarria Højbjerg, Lenka Tětková, Lars Kai Hansen tl;dr: DINOv3,CLIP know 3D geometry, but only in middle layers. MAE bad, ImageNet bad, cls token bad. arxiv.org/abs/2509.15271

🎯 How can we build faster, smarter video foundation models? At the TDW in Thessaloniki, Matthieu Cord (VALEO) presented VaVIM–VaVAM: boosting video AI with fewer, continuous tokens for faster training & better performance. ▶️ Watch the full talk: youtu.be/ykfwRa-ZQig?...

This is a great example of local actions that both reduce carbon pollution (climate mitigation by enabling carbon free active transportation) and also increase the resilience to climate impacts (climate adaptation by designing systems to handle increased amounts of rainfall due to climate change).

Ever wondered how Energy-Based Models (EBMs) work and how they differ from normal neural networks? ☕️ We go over EBMs and then dive into the Energy-Based Transformers paper to make LLMs that refine guesses, self-verify, and could adapt compute to problem difficulty.

I agree! Current growth is clearly non-sustainable.

Finally the dates for @cvprconference.bsky.social are public! cvpr.thecvf.com/Conferences/... And there's an additional week for supp videos 🥰

Ι'm 100% with Noa on this. But in all fairness, trust in peer review has been damaged long way before.

#NeurIPS2025: I am very happy that we have two choices this year, Mexico 🇲🇽 and the US 🇺🇸. This is GREAT 👏. The cherry on top would have been if Copenhagen 🇩🇰 would have also been a real alternative, instead of an "additional venue" only which still requires attendance in one of the primary ones.

As an SAC for @neuripsconf.bsky.social, I don't agree with PCs approach to reject papers based on ranking. I ranked my accepted papers as requested and explicitly stated that I support the acceptance of all papers. I was not given an explanation of why papers at the end of the rank were rejected.

This week's #PaperILike is "Hi Robot: Open-Ended Instruction Following with Hierarchical Vision-Language-Action Models" (Shi et al., ICML 2025). I'm often asked: how might we combine ideas from hierarchical planning and VLAs? This is a good start! PDF: arxiv.org/abs/2502.19417

𝗧𝗼𝘄𝗮𝗿𝗱𝘀 𝗨𝗻𝗱𝗲𝗿𝘀𝘁𝗮𝗻𝗱𝗶𝗻𝗴 𝗩𝗶𝘀𝘂𝗮𝗹 𝗚𝗿𝗼𝘂𝗻𝗱𝗶𝗻𝗴 𝗶𝗻 𝗩𝗶𝘀𝘂𝗮𝗹 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗠𝗼𝗱𝗲𝗹𝘀 Georgios Pantazopoulos, Eda B. Özyiğit arxiv.org/abs/2509.10345 Trending on www.scholar-inbox.com

In the words of Jitendra Malik: I will not bow the imperialistic stance of NLP. In my own opinion: anyone who thinks this is true for humans is delusional. Although I can understand that when you become a millionnaire thanks to just burning compute on simple ideas, you get addicted to that stuff.

Come for the computer vision, stay for the control theory lecture 😬 "There is no need to change this. This was correct before, it is correct today. Please keep it as it is!"

Unitree, Chinese Robotics company, released the open-weights of UnifoLM-WMA-0 It is Unitree‘s first open-source world-model–action architecture spanning multiple types of robotic embodiments, designed specifically for general-purpose robot learning.

"GenAI is just a tool, scientists are the ones who add value." Great point from @sarameghanbeery.bsky.social, reminding us that human expertise remains central to the research process. #MITGenAISymposium #MGAIC #AIForBiodiversity

First day was AMAZING! Great talks, deep discussions and very meaningful interactions! Best event I've been for computer vision. You get 10x of what you would get in a conference. Big congratulations to the organisers!

If you're interested in human pose estimation and mesh recovery from LiDAR data, we have this massive survey: arxiv.org/abs/2509.12197 Salma and Nermin put a tremendous amount of work in it, there's everything: the tasks, all the methods organized, datasets, numbers, challenges and opportunities.

Interesting graphs from csconferences.org Trying to predict the inflection point of the sigmoid.

I wish all buildings in cities were like this one. Prague, Krymská

New work by Nathan Leroux and coauthors: analog in-memory computing with gain cells cuts attention energy by up to 10,000×, bringing faster, greener large language models closer to reality. www.nature.com/articles/s43...

Outstanding opportunities in AI & Life Sciences, with strong @ellis.eu synergies

Math Inc.'s Gauss by Christian Szegedy They have built on the formalization work of Alex Kontorovich and Terence Tao, but extending their formalization by 22,000 lines of LLM-generated Lean code. www.math.inc/gauss

www.youtube.com/watch?v=XA-U... @maosbot.bsky.social

dino v2 so your model takes an image as input, and outputs a vector the objective is that if you crop and rotate the image at random, it outputs the same vector that's it this turns out to extract tons of information about the image, like "where is the dog" arxiv.org/abs/2304.07193

Want to evaluate open-world object counting in a fine-grained setting? We’re excited to release the LookAlikes dataset, which is a test-set-only benchmark where images contain objects from multiple visually or semantically similar categories! #ComputerVision #DeepLearning #AI #Counting #OpenDataset

Ever wanted to easily edit individual sounds in a complex audio scene—like removing a cough or making a doorbell louder? New paper - Recomposer(arxiv.org/abs/2509.05256) from our Sound Understanding team introduces a powerful way to do just that. Huge congrats to all the authors 🎉 #AudioEditing

📣 Re-launch of a joint ELLIS Reading Group “Mathematics & Efficiency of Deep Learning”, affiliated with the ELLIS Unit Graz and co-organized by ELLIS Members Linara Adylova (🇩🇪 @ruhr-uni-bochum.de) and Olga Saukh (🇦🇹 @tugraz.bsky.social). Learn more here: sites.google.com/view/efficie...

Mini marsupial bounces back from the brink! Once listed as Endangered, the crest-tailed mulgara expanded its range by more than 48,000 km² between 2015 and 2021, lifting its status to Least Concern. Biocontrol contributed to the heartwarming bounce-back. https://f.mtr.cool/afllseseeb

Would you like some good news? A new vaccine has just been approved for protecting koalas from chlamydia, which is one of their leading causes of death.

The arXiv paper has also been updated to the ICCV camera-ready version: arxiv.org/abs/2503.10781. We get SOTA in 5 video grounding datasets: iGround (ours), VidSTG, ActivityNet-Entities, GroundingYouTube, YouCook-Interactions!

YES

𝗩𝗶𝘀𝘂𝗮𝗹 𝗥𝗲𝗽𝗿𝗲𝘀𝗲𝗻𝘁𝗮𝘁𝗶𝗼𝗻 𝗔𝗹𝗶𝗴𝗻𝗺𝗲𝗻𝘁 𝗳𝗼𝗿 𝗠𝘂𝗹𝘁𝗶𝗺𝗼𝗱𝗮𝗹 𝗟𝗮𝗿𝗴𝗲 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗠𝗼𝗱𝗲𝗹𝘀 Heeji Yoon, Jaewoo Jung, Junwan Kim ... Seungryong Kim arxiv.org/abs/2509.07979 Trending on www.scholar-inbox.com

Are you interested in #AIforConservation #AIforBiodiversity #AIforWildlife or #AIforNature?? Are you located in the Boston Area? If so, come join us!! The AI for Conservation Slack community is doing our first local-area Boston meetup, partnering with iNaturalist and TEDx Boston!

first test in plugging the $100 Amazing Hands on a (cute) humanoid instead of grippers - pretty happy about it Hand: github.com/pollen-robot... Humanoid: www.pollen-robotics.com/reachy/

We have released the code, checkpoints and datasets for GROVE! Grab them at github.com/ekazakos/grove. We also provide grove-transformers — an inference-only interface for GROVE, implemented with 🤗 Transformers. You can load GROVE and try it on any video data in just a few lines of code!

Congrats to ELLIS Member @mtoneva.bsky.social (Faculty @ MPI-SWS 🇩🇪) on winning an @erc.europa.eu Starting Grant! Her project BrainAlign will receive ~€1.5M to advance brain-aligned language models & deepen understanding of language + neuroscience. 👏 #WomenInELLIS #ERCStG

We have released the code, checkpoints and datasets for GROVE! Grab them at github.com/ekazakos/grove. We also provide grove-transformers — an inference-only interface for GROVE, implemented with 🤗 Transformers. You can load GROVE and try it on any video data in just a few lines of code!

Cool work!

Sony AI is hiring game integration engineers! We do awesome RL applications to modern video games. If that excites you, check out the posting! We have positions for senior and staff level developers. Senior dev link (staff level link next in 🧵): sonyglobal.wd1.myworkdayjobs.com/en-US/SonyGl...

Modal, a GPU on-demand service provider, has entered the notebook space with Modal Notebooks, competing with Google Colab Pro. modal.com/blog/notebooks

Baidu just released a reasoning model 🔥 ERNIE-4.5-21B-A3B-Thinking huggingface.co/baidu/ERNIE-... ✨ Small MoE - Apache 2.0 ✨ 128K context length for deep reasoning ✨ Efficient tool usage capabilities

Opensourcing a tool to calibrate camera extrinsics painlessly in a minute, no checkerboards! It's based on EasyHEC, using diff. rendering to optimize extrinsics given object meshes+poses. Crazy that even a piece of paper works too. Code: github.com/StoneT2000/s... (paper example in next post)

check out this PhD opportunity for female students from Africa at @upf.edu : www.upf.edu/web/phd-engi... i have hired my amazing student @NnekaOkolo4 via this program, which was a great success. hope i can find someone this time as well! let me know if you have any questions

This is insightful! PS: I wish I had the chance to chat with neuroscientists more.

If you've been using a 📌 emoji to bookmark posts, @rafael.my and @samuel.bsky.team cooked up a special tool you can use to migrate them to your Saved Posts. Try it here: pin2saved.vercel.app

v1.108 is rolling out today 🚚 Now live, at long last: Bookmarks, aka Saved Posts. For all those posts you'll definitely plan to come back to! Update the app and give it a try. The button is right down there 👇

Yo, bookmarks is here!!! 🥳🥳🥳

Crash test your foundational models for object recognition at its finest granularity. Here are the updated results on our instance-level image retrieval benchmark (ILIAS-CVPR'25). DINOv3 and Perception Encoder (PE) are included, with DINOv3 being the new SoA! Oh, but no, look at this...

Is basic image understanding solved in today’s SOTA VLMs? Not quite. We present VisualOverload, a VQA benchmark testing simple vision skills (like counting & OCR) in dense scenes. Even the best model (o3) only scores 19.8% on our hardest split.