PhD @KTHuniversity
Stockholm, Sweden
Joined February 2018
- Tweets880
- Following212
- Followers192
- Likes24.4K
Pinned Tweet
اگه به ژانرهای متالکور یا پراگرسیو یا آلترناتیو متال (یا آهنگایی نزدیک به این ژانرا) علاقه دارید، این پلیلیست احتمالا به دردتون میخوره.
open.spotify.com/playlist/27…
Amir Karimi retweeted
We're releasing dense Argus-format annotations for >5,000 hours (290k episodes) of public robot data (ABC-130k, MolmoAct2 YAM, DROID, GM-100, ABC-Sim), labeled with GPT 6.1 Sol.
Per-hand subgoals and actions, plus outcomes, mistakes, recoveries and object-state changes 👇🧵
🤗 huggingface.co/datasets/nkal…
Amir Karimi retweeted
This project started with a literature search on "growing" work in modern vision pipelines, and I was surprised that there was not much, although there have been more recent works in language. Since I started my PhD by working on vision (object-centric?!) I was interested.
Excited to share “Growing to See”: growing SSL vision models provide better representations with up to 25% fewer FLOPs!
Co-led with @KaplFer.
With @Google’s Paradigms of Intelligence team (@kaitlinmaile & @oswaldjoh), @andrea_dittadi, Karl H. Johansson & @stefanAbauer.
1/13
Excited to share “Growing to See”: growing SSL vision models provide better representations with up to 25% fewer FLOPs!
Co-led with @KaplFer.
With @Google’s Paradigms of Intelligence team (@kaitlinmaile & @oswaldjoh), @andrea_dittadi, Karl H. Johansson & @stefanAbauer.
1/13
Planning to train MAE or DINOv3 yourself? We include extensive hyperparameter ablations to inform training choices!
We also report additional evaluations, such as robustness to missing visual evidence under random and semantic masking.
More details in the appendix!
12/13
We’re excited about the possibilities beyond pretraining: cheaper feature extraction and adaptive dense prediction. Growing autoregressive visual or multimodal models could also support self-speculative decoding.
Paper: arxiv.org/abs/2610.16817
13/13
Amir Karimi retweeted
🧵 1/10
One-step generation is advancing fast: from flow maps to drifting models.
But how do we train a one-step sampler without target samples? 🤔
Introducing OSIS: “Amortised Inference through One-Step Implicit Sampling”.
Work co-led with @ktamogashev, as part of an incredible team with Arran Carter, @shchoi55, @stefanAbauer and Esmeralda S. Whitammer! 🙌
📄 arxiv.org/abs/2610.12679
More details about the method 🧵👇
Amir Karimi retweeted
🆕 Very happy to share:
“One for All, All for One: Coordinated Multi-Agent Diffusion Steering via Stochastic Optimal Control”
Work co-led with top-notch collaborators @BarbanoRiccardo and Runchang Li 🤝
We show how to teach pretrained diffusion models to work together: coordinating robot arms, human motion, navigation and image components.
With frozen pretrained weights! ❄️
📄 Paper: arxiv.org/abs/2610.08595
Grateful to an incredible team: George Webber, Alexander Denker, Željko Kereta, @stefanAbauer, Francisco Vargas and Esmeralda S. Whitammer. 🙌
Idea, theory, and some examples below 🧵👇
Amir Karimi retweeted
🗓️ Three months ago today #Iran shut off access to the global internet. While connectivity has now largely returned, metrics indicate that users still face heavy filtering, similar to the interim period between the January protests and the start of the war.
Amir Karimi retweeted
⏳ The digital blackout in #Iran has just entered its twelfth week and is now in its 78th day. The internet shutdown, which has sent a country of 90 million largely offline for an unprecedented duration, continues to erode human rights, the economy and basic liberties at scale.
Amir Karimi retweeted
🗓️ Today marks the 70th day of #Iran's internet blackout, with the incident now surpassing 1656 hours. Digital connectivity is vital in times of crisis, and limiting service harms those most in need - people with disabilities, students, small businesses and the general public.
