Artem Zholus

I am excited to share that our BindGPT paper won the best poster award at #AAAI2025! Congratulations to the team! Work led by @artemzholus.bsky.social!

Thanks!

At the same time, it outperforms all domain-specific (diffusion) models while remaining virtually free of inductive bias. We demonstrate that the key to this success is simple: it's data and scale. 1/2

Happy to share that one of my latest works has been accepted to AAAI25! BindGPT is a new foundational model for generative chemistry. It tackles various generative tasks, 3D molecule generation and 3D conformer generation, with a single model—something previously impossible. 1/2
bindgpt.github.io

Both papers compare with the most extensively tested baselines (DreamerV3 and TD-MPC2, both have 100+ envs in their papers). The second paper is extremely rigorous in the depth of their study (they had to prove a negative result). That’s why I think it’s more complicated than this.

DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning The ability to predict future outcomes given control actions is fundamental for physical reasoning. However, such predictive models, often called world models, have proven challenging to learn and are...

Today I found two world model papers that make opposite claims in their central message:
DINO-WM arxiv.org/abs/2411.04983 - frozen vision models are good for MBRL
arxiv.org/abs/2411.10175 - frozen vision models are bad for MBRL (accepted at NeurIPS btw)
What's funny is that they are one week apart

Hi Adhiraj, can you please add me to the pack? I work on LLMs and vision-based world models in my PhD.

A lot more exciting things are coming! For example, we’ll soon share the results from my internship at Google DeepMind. Stay tuned! 🧵5/5

I also built a new foundational LLM for drug discovery: BindGPT bindgpt.github.io, which solves 3D generative molecule tasks using an LLM. We developed the full-stack pipeline: pretraining, SFT, RL, and prompting. It’s exciting for both LLM and AI4Science communities! 🧵4/5

A year later, I began my PhD at Mila, and my first PhD paper was accepted for an oral presentation at ICLR 2024! 🎉 We developed a world model-based agent that excels at RL memory tasks, leveraging a State-Space Model (SSM) as its primary memory backbone. Check it out: recall2imagine.github.io 🧵3/5

My past research also covered instruction following with embodied agents - I co-organized two NeurIPS competitions in 2021 and 2022. 2/5

Hi everyone, my name is Artem and I am a PhD student Mila. Follow me if you are interested in AI agents, world models, LLMs and beyond! I am interning at Meta AI where I scale world models to real world. I also interned at Google DeepMind where we built new foundation models for video! More in 🧵1/5