Antonios Tragoudaras

Now: PhD in Physics-Grounded Foundational Models @Inria
Previously: 🤖🎓 MSc AI @UvA_Amsterdam - @Ellis_Amsterdam MSc Honors Student - @VISLab_UvA

[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)

[Andrii Zadaianchuk 🇺🇦 @ZadaianchukML]

In MORPHEUS, we also showed that any prerequisites for quantifying physical plausibility in VGMs is objects permanence, which most model were also strangling with. Great to see approaches that are significantly improving it.

https://t.co/988ExVIVjI

[Andrii Zadaianchuk 🇺🇦 @ZadaianchukML]

A robot looks at a cluttered table. The mug hides behind the box. The bowl is half out of frame.3D perception returns what the sensor saw: a shell full of holes.

You don't perceive the table that way. Why should a robot?

TL;DR: RecGen @ #ECCV2026 Friday 10:30, ExHall#375!
🧵

[Chenyu Zhang @AngelicaCZhang]

Regretfully unable to attend #ICML this year due to unexpected visa processing delay. Still, I’m very keen to discuss Morpheus and explore physical consistency in video generation with fellow researchers. If you're interested, I am happy to share my thoughts!

[Andrii Zadaianchuk 🇺🇦 @ZadaianchukML]

Science is all about being agnostic, open-minded.

When I seen huge increase in visual quality of the VGMs like Veo3 and claims that soon they would be basic tool of world modelling, we decided to test properly how well scaled VGMs output follow structure of the world: physics 🧵

[Antonios Tragoudaras @TragosML]

With a wonderful team of:
- @Cherniavskii_D @AngelicaCZhang @AntoniosVozikis Thijmen Nijdam Derck Prinzhorn Nicu Sebe
- With the one and only @ZadaianchukML and the best advisor ever: @egavves
🏛️ @UvA_Amsterdam & @UniTrento - Presentation funded by @Ellis_Amsterdam

[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)

[Antonios Tragoudaras @TragosML]

📄 Paper: https://t.co/cyR880F22d
💻 Code: https://t.co/0o04QZj8FQ
🤗 Data: https://t.co/9HmVTunBm6
🌐 Site: https://t.co/h6iseKv5W4

[Antonios Tragoudaras @TragosML]

The takeaway: physical plausibility isn't a free bonus of scale. It has to be measured — and, eventually, built in. That's physics-grounded AI.

Come talk it through at #ICML2026: 🗓️ Thu Jul 9, 2:30–4:15 PM KST 📌 Hall A, Poster #4611
🧵(6/6)

[Antonios Tragoudaras @TragosML]

What we learned 🔎

• Temporal context wins: conditioning on first+last frames is the most physically consistent; single images are hardest. • You need two scores — equation-fit + conserved quantities catch different failures. • Bigger prompts ≠ better physics.
🧵(5/6)

[Antonios Tragoudaras @TragosML]

Even the best models — with advanced prompting & video conditioning — make beautiful clips that quietly break Newton:

⏸️ objects freeze · 👯 duplicate · 👻 vanish · 🎯 pendulums gain energy from nowhere

And open source often keeps pace with the big closed models. 🧵(4/6)

[Antonios Tragoudaras @TragosML]

How Morpheus works — no human judges, no single "correct" future:

Track every object (SAM-2) → fit its motion to the governing equation (a PINN) → check the invariants: is energy & momentum conserved?

Two complementary scores, one verdict ⚖️ 🧵(3/6)

[Antonios Tragoudaras @TragosML]

Our answer: not really.

We test generated videos against something that can't be faked — conservation laws. Energy & momentum don't negotiate.

Real footage scores 0.96–0.99. The best model tops out at 0.66.
🧵(2/6)

[Antonios Tragoudaras @TragosML]

@sedielem @icmlconf We need Diffusion Circle to return at ICML

[Antonios Tragoudaras @TragosML]

@AmitLeViAI I think you forgot to give credit to making the architecture figures with Nano Banana 🍌

[Antonios Tragoudaras @TragosML]

@AmitLeViAI Hahahah. Can you share the prompts please?🙏

[Antonios Tragoudaras @TragosML]

@iclr_conf @iclr_conf still waiting for list of accepted workshops

[Antonios Tragoudaras @TragosML]

Lots of cool research coming soon. Stay tuned! 🔭 For anyone interested in my thesis work, feel free to follow our project website for upcoming updates
https://t.co/fSaGM2zwJr

[Antonios Tragoudaras @TragosML]

Life Update:
A month ago I have officially graduated from the Master’s in AI at UvA!

Next Chapter: I’m already 1 month deep into my PhD at Inria (Paris Center) with https://t.co/uAUe9jhtYp

Excited to keep working on Video Foundation Models & Physical World Understanding.

[Antonios Tragoudaras @TragosML]

A massive thank you to @Ellis_Amsterdam for this research opportunity, for funding the research visit, and for hosting an amazing ceremony to celebrate the completion of the program alongside other talented students.

https://t.co/HiFetJfNBt

[Antonios Tragoudaras @TragosML]

Deeply grateful to all my supervisors:
@ZadaianchukML
@Cherniavskii_D
@egavves
Your expertise was instrumental to the success of the research project and for my academic growth.

Special thanks to @FrancescoLocat8 , for hosting my research visit and for his co-supervision.