Ivan Rubachev

ML Researcher @YandexResearch | Tabular ML

[Ivan Rubachev @puhsuuu]

New report with a benchmark of optimizers for modern tabular MLPs.

Muon is the new best optimizer in our domain!

https://t.co/EwCoPBXl0I

[Yura Gorishniy @YuraFiveTwo]

A small yet practical update for tabular deep learning people:

Muon is a strong alternative to AdamW for training modern tabular MLPs, including TabM. Give it a try!

Overall, our technical report covers 15 optimizers:
https://t.co/e4uBFuqVhB

Details 👇

[Yandex Research @YandexResearch]

GraphPFN got the best paper award at the Graph Foundation Models workshop at ICML!

[Nikita Kachaev @ ICML 🇰🇷 @judokach]

22-DoF HUMANOID + SIX-LINK PENDULUM attached to the torso SOLVED with RL!

no sweep. first seed success.

@yacineMTB's six-link cartpole was cool.

BUT this is the harder version
not a cart. not 2D. not fixed-base.
the only actuator is the robot body!

[Dmitry Eremeev @eremeev_d42]

GraphPFN is at ICML 2026 – see you tomorrow at Poster Session 1 poster 2411

Btw, we’ve recently released GraphPFN-1.3 with faster pretraining, better results and PyPI package!

[Ivan Rubachev @puhsuuu]

@kolya_kartashev is the guy here :)
Come by the poster (are you in Seoul?)

[Ivan Rubachev @puhsuuu]

https://t.co/edKjk7jyUM for some reason this got me super hyped for the ICML trip a few month ago

[Ivan Rubachev @puhsuuu]

I’ll be at ICML this year

Happy to ch̶i̶l̶l̶ chat

Seoul is very nice btw! 🇰🇷 🍎

[Yandex Research @YandexResearch]

We've got hype... and 8 papers at ICML this year! Come talk to us! The thread with links to the papers below 🧵

[sankalp @dejavucoder]

last month i was learning RL in a bottom up manner and struggled to come up with good ideas for experiments. soon, i also realised that if i want to get better at research, i gotta get better at ideas. in this post i make my problem everyone's problem.

https://t.co/123stDbxe8

[Khurram Javed @kjaved_]

The Physical Atari work is finally in a good state to share. I procrastinate by making nice UIs and visualizations, and I am quite happy with how the website for Physical Atari turned out (link in the reply). 



There are many ways to use this platform for research. For me, its appeal is that it can be used to bring some rigour to research on RL applied to robotics.

It is common to use RL in robotics to build demos using sim-to-real or off-policy learning from human-behaviour datasets, but these methods are rarely clearly compared with one another in ways that increase our understanding of the area.



We tried sim-to-real on the Physical Atari system and found the performance to be abysmal. This is perhaps not surprising. Sim-to-real is not a paradigm that scales with compute and data. A human has to constantly be involved to reduce the sim-to-real gap by repeatedly redeploying policies.

I'm sure that, with enough iterations, this gap could be reduced on Physical Atari and sim-to-real could be made to work, but good sim-to-real transfer does not emerge simply from more data and compute.

Surprisingly, we found that even real-to-real transfer---in which policies learned on one robot body are evaluated on a different body---significantly degraded performance. It’s possible that due to differences in robot bodies (arising from manufacturing variation and wear over time), the best policies are learned by doing RL directly on the robot. Learning directly on a robot bakes the exact dynamics into the policies, which can be superior to training for much longer under incorrect dynamics.

The work will be presented at RLC 2026 in Montreal, and the associated paper is on the website.

[Ivan Rubachev @puhsuuu]

@francoisfleuret @mike64_t https://t.co/kiuhtJG7X9

[antirez @antirez]

Distributing LLM inference in DwarfStar: https://t.co/Pgb548OZ3w

[Jeremy Howard @jeremyphoward]

@tokumin I feel that the trend towards training models to autonomously go off and try to do everything themselves is anti-human.

We should, IMO, be training LLMs to support humans in their learning, creativity, and iterative experimentation.

[Machine Learning Street Talk @MLStreetTalk]

Michael I. Jordan on the new MLST.

Four things:

> AGI is a PR term. It confuses young people.

> Discourse is bipolar, either alarmist or exuberant, this is in his words "so demoralizing" for 20- and 25-year-old researchers.

> ML's methods came from statistics and operations research, NOT the AI tradition.

> Data markets are Stackelberg games, not optimisation problems. A lot of ML researchers have never computed an equilibrium.

Michael I. Jordan is a no-nonsense original gangster of the field and was described by Science magazine, back in 2016 as the most influential living computer scientist.

[Fernando 🌺🌌 @zetalyrae]

New day, new post. Today is about why having a superhuman AI available at all times might not change your life much.

[Ivan Rubachev @puhsuuu]

nice discussion of "The Emacsification of Software" on hn:
https://t.co/wt3PEwTeRM

[Ivan Rubachev @puhsuuu]

@mempirate Not tigerbeetle, but https://t.co/DBgPkA4GzR

[Ivan Rubachev @puhsuuu]

@antirez True! Also https://t.co/HCKW02QHQp for emacs users is a hidden gem

[Yandex Research @YandexResearch]

@iclr_conf 2026 starts tomorrow — see you there. This year, 6 papers by Yandex Research and our collaborators were accepted.
Starting a thread 🧵

[Ivan Rubachev @puhsuuu]

@xidulu In our domain (tabular NNs) generalization is more important than reaching a target loss at a compute budget and Muon is helpful there https://t.co/EwCoPBXl0I plus the @industriaalist's qlabs slowrun is also pointing into this direction (muon being better at genearlization)

[Ivan Rubachev @puhsuuu]

@kepano You may like https://t.co/RNTgvzv2Ej

[llm_enjoyer @LLMenjoyer]

what talking to any team from a frontier lab is like
(yes this video is proof @cursor_ai is not a lab)