🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Before, it was all about finding the next architecture. Now, it's all about finding the best tokenization. Next, it will be all about finding ways to collect new data and exploring.

We presented this work at RL_Conference this year at the Finding The Frame Workshop and we now put a version on arxiv. The purpose of benchmarking in RL seems convoluted and benchmarking choices are often not thought about much. We argue that we should change this!

🎉Claas and I have another preprint out: MAD-TD🤪 After combatting divergence in high utd RL, we can now stabilize ood action value prediction using model-based synthetic data. This makes resetting networks unnecessary and leads to great performance. Check it out👇

After putting it off for a while, I’m diving into transformer research this winter. Joining John Langford’s team Microsoft Research to work on transformer representations. Don't worry, I'm not leaving RL behind but I need to learn some new things so we can get to point 3.

CoLLAs stacking their line-up early. Met Jakob at ICML and he is amazing! Even more excited because the conference is gonna be here at Penn next year. Come visit us!

Yes, yes I'm on it, ACs. Here is a live, completely real picture of me working on it.

My first week at Microsoft Research is ending today. One of the highlights of this week was how excited everyone got when people shared that this work was public now and how supportive everyone was. Definitely check it out: microsoft.github.io/OmniParser/

Been hearing a lot about #NotebookLM but haven't tried it out myself. To my fellow researchers, do you find it useful and how do you use it? Is it work getting started and if so, what are some good things to try? #Research #Science #AcademicTwitter

What is the perception of #TMLR in the community at this point. How much is it "worth" compared to conference papers in core AI/ML? I've had very different reactions to this IRL, so I want to get a more general feeling. #mltwitter

Eugene got me to try it out. I also made a starter pack for machine learning #theory people but I haven't found too many I know. So if you want in just message me there x.com/EugeneVinitsky…

As everyone in #mltwitter, I'm probably going to be more active on ye goode other app in the sky going forward. Same handle, same shit-posting!

Bluesky adds 700,000 new users in a week trib.al/HlfBOpD

Also on🦋Swaminathan.bsky.social! If you're looking for where to get started, the following starter packs of people to follow should be helpful: go.bsky.app/3WPHcHg, go.bsky.app/21nFz12, go.bsky.app/PcwNoSy, go.bsky.app/DfAoaJ1, go.bsky.app/SipA7it.

Throwing compute at data has been incredibly powerful in other domains but used to be difficult in RL. Not so much anymore! Our MAD-TD paper will be presented as a spotlight 🚨 at #ICLR. As Claas and I are both not active here anymore, come discuss with us on the other platform!

We often use #VAML/ #MuZero losses with deterministic models. But if we want stochastic models to measure uncertainty or to leverage current SOTA models such as #transformers and #diffusion, we need to take care! Naively translating the loss functions leads to mistakes!

Come say hi at our #ICML poster today during Poster Session 1 (W-600)! Joint work with Aaron Roth Michael Kearns Jess Sorrell Marcel Hussing and Eric Eaton! arxiv.org/abs/2502.11828