Denis Tarasov

PhD at ETH Zurich. Ex Meta, InstaDeep, JetBrains, Yandex, Tinkoff.

[Denis Tarasov @ML_is_overhyped]

📜Find the paper here: https://t.co/DRDYbtj3Cb

[Denis Tarasov @ML_is_overhyped]

Our paper on the benefits of actual RL for In-Context RL has been accepted to RLC!

Looking forward to attending @RL_Conference and discussing the work there.

[Denis Tarasov @ML_is_overhyped]

LLMs are amazing because they can learn in context — read, adapt, and act.

Can we do the same for reinforcement learning? That’s the promise of In-Context RL (ICRL).

But existing offline ICRL methods don’t even optimize rewards.

Our new paper shows why RL matters
🧵

[Denis Tarasov @ML_is_overhyped]

Happy to share that I’m officially starting my PhD in Reinforcement Learning and Robotics at ETH Zurich under the supervision of Prof. Robert Katzschmann @katzschmann.

It’s been quite a journey to get to this point, and I’m excited for what’s ahead!

[Denis Tarasov @ML_is_overhyped]

Super excited to share that our paper was accepted as an ICLR 2026 Oral 🚀
See you in Rio 🇧🇷

[Denis Tarasov @ML_is_overhyped]

I have a strange tradition to work on LM + RL project once a year😶‍🌫️

[Albina Klepach @_cinemere]

🤩Excited to share our #AAAI2026 Oral on #EmbodiedAI:

How can robots learn from noisy, unlabeled videos—like those on the internet—without action labels?

Our new work, Object-Centric Latent Action Learning, tackles exactly this.

Thread ↓

[Denis Tarasov @ML_is_overhyped]

Don’t want to spam your feeds, but I’m incredibly thankful to the AI community for all the feedback and support on my last tweet. I didn’t expect so many responses -- really happy to be part of such a brilliant community. Thank you so, so much!

[Denis Tarasov @ML_is_overhyped]

@amirhkarimi_ @hllo_wrld @_rockt @robertarail Thank you very much for your suggestion. I'm afraid Canada is not an option for me as well as US (and UK now) due to the visa difficulties:(

[Denis Tarasov @ML_is_overhyped]

I’m asking for help. I was meant to start my PhD with @_rockt and @robertarail at UCL, but my UK background check was refused. My appeal seems unlikely to succeed, so I’m urgently searching for any PhD or research positions in academia or industry. Any help is appreciated.

[Denis Tarasov @ML_is_overhyped]

@sarsanaee @_rockt @robertarail Yes, that's ATAS. They do not provide any reasons but most likely it is due to the combination of my research area and nationality (I know other people with the same issue who didn't get ATAS). I can't reapply for 3 months. Was waiting for their decision 6 months

[Denis Tarasov @ML_is_overhyped]

@SnchzPedro_ @_rockt @robertarail Had an interview for this position last year but it was canceled:(

[Denis Tarasov @ML_is_overhyped]

I guess my scholar describes my background the best: https://t.co/gTb0Rfvh0E

[Denis Tarasov @ML_is_overhyped]

Many thanks to my collaborators @how_uhh @suessmannn and @vladkurenkov!

[Denis Tarasov @ML_is_overhyped]

Meet NinA — Normalizing flows in Action.

We tried a simple idea: replace diffusion decoders in VLA models with normalizing flows.
🧵

[Denis Tarasov @ML_is_overhyped]

Paper: https://t.co/k4aSggZQMi
Code: https://t.co/hHKMOf6AdH

[Denis Tarasov @ML_is_overhyped]

I have a strange tradition to work on LM + RL project once a year😶‍🌫️

[Vladislav Kurenkov @vladkurenkov]

🚀 Introducing cadrille: a new SOTA model for CAD reconstruction from images, point clouds, and text—all in one framework with the use of RLVR.

Multimodal inputs + RLVR = clean, editable 3D models.

🧵👇

[Denis Tarasov @ML_is_overhyped]

Big thanks to my collaborators @how_uhh @suessmannn @vladkurenkov!

[Denis Tarasov @ML_is_overhyped]

LLMs are amazing because they can learn in context — read, adapt, and act.

Can we do the same for reinforcement learning? That’s the promise of In-Context RL (ICRL).

But existing offline ICRL methods don’t even optimize rewards.

Our new paper shows why RL matters
🧵

[Denis Tarasov @ML_is_overhyped]

🧠 We believe offline ICRL should actually do RL.

Our results show that reward-aware training improves many crucial aspects of ICRL.

Check out the paper here: https://t.co/DRDYbtj3Cb