ETH AI Center

We’re now on Bluesky!🦋 Follow us at @ eth-ai-center.bsky.social (bsky.app/profile/eth-ai…) for updates on AI research, events, and more. See you there! #AI #ETHZurich #BLUESKY

Our recent finding on Diffusion Alignment: a reward model in pixel space can be easily transferred to score noisy diffusion latents directly — at small finetuning cost, via stitching. This makes Faster & Better for both Training & Inference Alignment. Meet StitchVM👇 1/

Long time no post! 👋 Excited to share that our 3rd AI for Math Workshop has been officially accepted at ICML 2026! 🎉 📍July 10, Seoul 🧠 Toward Self-Evolving Scientific Agents 📝 Paper submission open now! 🌐 Check out ai4math2026.github.io #ICML2026 #AI4Math

🥳The 3rd ICML AI for Math Workshop. Welcome to join us! 📄Call for paper: openreview.net/group?id=ICML.… 🏆Challenge submission: #challenge" target="_blank" rel="nofollow noopener">ai4math2026.github.io/#challenge 🙋Reviewer nomination: forms.gle/8hfysRAKcSBmSa…

🤝 AI builders at @ETH_en unite! Great to have ETHAIClub.org on peer learning & side projects launched with AI builders Thomas Bollenbach, Alessio Rimoldi, & Sven Jonscher! 🚀 → first Demo Night: luma.com/jc3yzdc8?utm_s… Supported by @AV_ETH @ETH_AI_Center

will be in Zürich on May 21!

LLM Agents? Post-Training? Training to Train? On Thu 21 May 2026 at 15:00, it's my pleasure to host @maksym_andr (@ELLISInst_Tue) for an academic talk at @ETH_AI_Center on PostTrainBench.com (@icmlconf 2026). Info & registration (only in-person): ai.ethz.ch/research/event…

🪞Symmetry helps shape RL policies. 🧠Can it also improve memory and task understanding? ✅We introduce Symmetry-Guided Memory Augmentation, combining structured experience augmentation with memory-based context inference to boost learning efficiency. 🔗sites.google.com/view/eth-sgma/…

🚀 Two new papers from our team are now available on ArXiv, both tackling core bottlenecks in RL post-training 1. Annotating human preference datasets without spending a fortune 2. Quantifying uncertainty for reward models 🔗lasgroup.github.io/rlhf

Robotic World Model is a black-box, end-to-end learned dynamics model for general robotics. While primarily tested on rigid robots, we see its greater potential in systems with more complex dynamics. (s-a-s') Data in -> Policy out 🥳Now on Colab colab.research.google.com/drive/1SRL0ss5…

💬The upcoming AI for Good Webinar Series - with @mariotelfig (@istecnico, @itnewspt) and @ArnoutDevos (@ELLISforEurope, @ETH_AI_Center) on "The why of AI: Uncovering cause and effect in observational data" 📅 Wed, Feb 18th 🕓 16:00-17:00 CET 🔗Register: aiforgood.itu.int/event/the-why-…

Deployed LLMs and users generate millions of conversations every day. These are full of useful learning signals, yet we don't use them for training. We introduce self-distillation for learning directly from user conversations – no rewards, no labels, no extra models.

The submissions to our @CVPR workshop OpenSUN3D on Open-World 3D Scene Understanding and Representations are open! 🏔️🤖☀️ ➡️ The deadline for the proceedings track is on the 5th of March!🏃‍♀️ 🌍: opensun3d.github.io 📝: openreview.net/group?id=thecv…

🌍 World models need to know what they don't know. 📢 Our Doctoral Fellow Chenhao Li @breadli428 incorporates uncertainty into his Robotic World Model to enable real robot control. 🚀 Project: sites.google.com/view/uncertain…

Today @mimicrobotics and friends are excited to share mimic-video, a new class of Video-Action Model that elevates video model backbones as first class citizens for robot learning!

Shoutout to @jjonpai, @Liam862373, @oier_mees and the rest of the team for making this happen! @mimicrobotics @Microsoft @ETH @ETH_AI_Center @UCBerkeley @NVIDIARobotics You can find the paper website here: mimic-video.github.io

Tired of conditioning 3D generative models only through images and text? ➡️ We just released code of SpaceControl, our method to control the generation directly in 3D ✨👩‍🎨 👾: github.com/spacecontrol3d… 🌍: spacecontrol3d.github.io @FrancisEngelman @orlitany @IanHuang3D @mapo1

Are images and text enough to control 3D generative models? SpaceControl shows there’s more: we turn 3D drafts into detailed, perfectly aligned 3D assets ✨🚀 📜: arxiv.org/pdf/2512.05343 🌍: spacecontrol3d.github.io

ZurichRobotics#2 is on December 15th at the @ETH_AI_Center! Antonio Arbues @arbwes (@loki_robotics) will talk about building useful robots in the age of AI and Giacomo Manzoni (Hexagon) about imitation learning for dexterous manipulation. RSVP below, spots limited.

🌍World models are compelling to robot intelligence, yet deployment remains fragile. 📢Our Doctoral Fellow Chenhao Li (@breadli428) dives deep with his Robotic World Model and isolates the missing components blocking reliable real-world control. 🚀Project: sites.google.com/view/roboticwo…

Getting a dynamics model that is compatible with PPO's long rollouts is hard. It took us a long time to find effective yet simple tricks to make imagination rollouts less prone to hallucination. In the end, autoregressive training with backpropagation through time worked out.