Daniil Cherniavskii

PhD student @ University of Amsterdam

[Daniil Cherniavskii @Cherniavskii_D]

This work was done during my research internship at Meta FAIR with Daniel Severo and Karen Ullrich. I’m grateful to everyone who supported the project and shared feedback.

We’d be very happy to hear your thoughts!

#FlowMatching #DiffusionModels

[Daniil Cherniavskii @Cherniavskii_D]

New preprint: Context-weighted Discrete Flow Matching.

Paper: https://t.co/iHokRakAtG

We study how local context relates to token predictability in discrete flow and diffusion models -- and how it can be used to improve sampling.

🧵

[Daniil Cherniavskii @Cherniavskii_D]

The idea is not specific to text. On QM9, context weighting produces up to **2.8× more valid** and **1.9× more novel molecules** than the corresponding vanilla DFM, showing that the same local signal can also help structured generation.

[Antonios Tragoudaras @TragosML]

🎬Video models now make stunningly realistic clips — and we already call them "world models." #ICML2026.🇰🇷

But looking real ≠ being real.

Before trusting them for robotics, self-driving & science: do they actually obey physics?

Morpheus puts them to the test 🧵(1/6)

[Daniil Cherniavskii @Cherniavskii_D]

@BoWang87 Looks awfully similar to TTRL (https://t.co/qAipvpgq2g). Just SFT instead of GRPO. This has been known for a while now, e.g., maximum likelihood sharpening objective (https://t.co/A6XlSTgs1o).

[Kwang Moo Yi @kwangmoo_yi]

Preprint of today: Zhang, Cherniavskii, Zadaianchuk, Tragoudaras, et al., "MORPHEUS: Benchmarking Physical Reasoning of Video Generative Models with Real Physical Experiments" -- https://t.co/i2N9KY9OoT

Can we use off-the-shelf video models to simulate physics? Perhaps not yet.

[Daniil Cherniavskii @Cherniavskii_D]

@nikitadurasov Same.

[Petar Veličković @PetarV_93]

Unveiling the next chapter in our quest towards the elusive theory of *all* neural networks: Categorical Deep Learning! 🐈🤖

https://t.co/EMFRpsA6NX

See thread 🪡 for a bucket list 🪣 of concepts our framework can handle!

w/ @bgavran3 P. Lessard @andrewdudzik @tlvg @_joaogui1

[Efstratios Gavves @egavves]

📢📢Excited to announce the 1st Causal &Object-Centric Representations for Robotics Full-day Workshop at @CVPRCall
Abstracts deadline 4/28
W. amazing @FrancescoLocat8 @RoozbehMottaghi @ZadaianchukML @phillip_lippe @Priyam8Parashar @PatkiSiddharth R.Desai
https://t.co/gy24Lu5SOo

[Phillip Lippe @phillip_lippe]

Along with our new tutorials (https://t.co/7AtaB7j8l9), we organize a lecture on the topic! Feel free to join, also if you’re new to JAX

[David M. Knigge @davidmknigge]

The @UvA_Amsterdam and @ELLISforEurope are starting a masterclass series where experts explain all about their favourite topics in and around Machine Learning.

Join us for the first on March 6th; "Differential Geometry for Deep Learning" masterclass by @gabel_alex. ✨

👇

[Erik Bekkers @erikjbekkers]

✨PhD vacancy alert✨ Daan Pelt and I are looking for someone that wants to work with us on generative modelling of shape and geometry!

* Position at @AmlabUva & @LIACS.
* Apply before 3 Nov!
* Apply here: https://t.co/Z10uCkCXaA
* Part of the HI project: https://t.co/Yp8T1OtHso