Efstratios Gavves

Great honor to be giving a keynote in the next DAGM!

Stay tuned for my next grand vision on 'Cyberphysical AI' towards Robot General Intelligence.

Robot learning algorithms that are literally aware of and can interact with the physical world, its physics, and dynamics, critical for System II AI.

Language Agents Meet Causality -- Bridging LLMs and Causal World... Large Language Models (LLMs) have recently shown great promise in planning and reasoning applications. These tasks demand robust systems, which arguably require a causal understanding of the...

🔥 4/4 #ICLR2025 Language Agents Meet Causality – Bridging LLMs and Causal World Models, integrating causal representation learning with LLMs to enable causally-aware reasoning and planning
openreview.net/forum?id=y9A...

with John Gkountouras @phillip-lippe.bsky.social Matthias Lindemann Ivan Titov

Grounding Continuous Representations in Geometry: Equivariant... Conditional Neural Fields (CNFs) are increasingly being leveraged as continuous signal representations, by associating each data-sample with a latent variable that conditions a shared backbone...

🔥 3/4 #ICLR2025 Grounding Continuous Representations in Geometry: Equivariant Neural Fields (ENF), for geometry-informed continuous signal representations openreview.net/forum?id=A4e...
+@dafidofff.bsky.social, @davidmknigge.bsky.social @erikjbekkers.bsky.social S. Papa, R. Valperga, S. Vadgama

CaPo: Cooperative Plan Optimization for Efficient Embodied... In this work, we address the cooperation problem among large language model (LLM) based embodied agents, where agents must cooperate to achieve a common goal. Previous methods often execute actions...

🔥 2/4 #ICLR2025 CaPo: Cooperative Plan Optimization for Efficient Embodied Multi-Agent Cooperation, a framework enhancing cooperation efficiency among LLM-based embodied agents through strategic planning
openreview.net/forum?id=KRv...
wtih Jie Liu, Yingjun Du, @cgmsnoek.bsky.social et al

Dream to Manipulate: Compositional World Models Empowering Robot... A world model provides an agent with a representation of its environment, enabling it to predict the causal consequences of its actions. Current world models typically cannot directly and...

🔥1/4 #ICLR2025 Dream to Manipulate: Compositional World Models Empowering Robot Imitation Learning with Imagination, integrating real-time photorealism &physics simulation for enhanced robot learning openreview.net/forum?id=3RS... +L. Barcelona, @zadaianchuk.bsky.social,@samuelepapa.bsky.socialetal

🎉 I am thrilled to share that our team has achieved a tremendous milestone with FOUR papers accepted to #ICLR2025! Huge congratulations to all the amazing students and collaborators. Details per papers in separate messages. Looking forward to an exciting conference in Singapore! 🛫🇸🇬 #AI #ML

Something like that 🤣

Thanks!

I very much agree so. Do you have references or are these anecdotal ad hoc conversations?

That said, I think there exist also technical reasons why there is this disconnect (and where embodiment also helps). But I will say my opinion a bit later to not bias.

Ok, but I think right now they work really well, to the extent that is a 'near-solved' problem, academically speaking.

Sure, but I think the same one would say about object recognition, a few years ago.

Why our computational models cannot be generally enough?

(I know I'm being annoying 🤣)

But why? I will also put my answers in the end, I'm not taking the easy way out here 🤣

One doesn't exclude the other. Either way, regarding system I, I think we are close. Or good enough to seriously start thinking about system II

I ve been thinking questions on
- world models
- physics & causality
- System I & System II reasoning

1. Why NNs can't model physics and causal mechanisms easily?
2. Why world models can't model physics and causality easily?
3. Why System II is not working yet?

I got some ideas, but your thoughts?

Job alert! 🚨 (a bit special but anyway)

If you are (or know someone who is) a Phd about to graduate or just graduated, and have to skip some in between time, I currently have a PostDoc position that could run for 6 months.

Needed: Experience with publishing at Top A conferences.
Just ping me. 👍

Fantastic, thanks! This feels really nice, I hope I have enough time though

go.bsky.app/5KXsVXb

Created Starter Pack for VISLab, if you see more people whom I should add suggest here

Thanks!

Maybe you can take the initiative to make one for VIS :P

What would you mean by an UvA starter pack? My research team that I can reference?

😂 Good point

yeap!

No, ideally I don't want to maintain two platforms, I barely have time for one. I just want to see if there is any easy way to replicate the network that I have made in Twitter, now in Bluesky.

So, how do I use this thing :P

Do I lose all my followers on Twitter?

And also, I am just curious, how often do you use Twitter/Bsky? It seems to me that everyone uses it but my whole day is so full that I barely have time to browse every now and then.