Chandar Lab

Extremely excited to share our recent work on diffusion world models. We ask a simple question - what space supports diffusion world modeling the most and how do we evaluate that?Turns out representation is the answer with JEPA space yielding the strongest diffusion world models!

Diffusion world models can help test and improve robot policies before running them on real robots. But can the choice of latent space make the WM more faithful? We show that semantic spaces beat reconstruction spaces on task relevant metrics. hskalin.github.io/semantic-wm

Checkout the recent work on diffusion world models from our lab! The work studies a simple but important design choice: should the world model think in pixels reconstruction space or in semantic feature space?

Still at #ICLR2026 and interested in training dynamics, simplicity bias, training under data distribution shift, and model merging? Come to our workshop posters to see what we are working on! 🧵👇

Want to know the expressivity of Mamba 3? Come by our ICLR poster! Sat, Apr 25 • 3:15 PM – 5:45 PM Pavilion 4 P4-#4409 The Expressive Limits of Diagonal SSMs for State-Tracking Joint work with Behnoush Khavari, Siamak Ravanbakhsh, and @apsarathchandar.

I'm in Rio de Janeiro for ICLR 2026! 🇧🇷 I will be presenting our paper "Is Depth Heterogeneity a Barrier to Model Merging?" (spoiler: not really) at the Workshop for Test-time Updates on Monday April 27th! Please dm if you're there and you'd like to geek out about... (1/2)

Sugar-Shack with our lab!! 🍁🪵 Saying bye bye to a long, harsh winter and hello to sunny days!!☀️🤩

📣 Announcing the CoLLAs Seminars A year-long exploration of one of the central challenges in AI: building systems that can learn continually, adapt in real time, and improve over their lifetime. Join us on May 13th at 11 am ET as we kick off the series with Pulkit Agrawal speaking on “Rethinking Post training”. ℹ️ Learn more: lnkd.in/erEdDxgP ✉️ Join our mailing list: lnkd.in/eEGwH-3E 🔗 Zoom link for the talk: lnkd.in/ekkHE5nX

If you work in continual learning, adaptation, online learning, updatable ML, unlearning, model editing and any other form of non-stationary/non-iid learning settings, consider joining the @CoLLAs_Conf mailing list here: groups.google.com/g/collas-list!

Hierarchical planning unlocks long-horizon, non-greedy behavior in JEPA world models. Paper: arxiv.org/pdf/2604.03208 Website: kevinghst.github.io/HWM/ Code: github.com/kevinghst/HWM_…

The abstract deadline for @CoLLAs_Conf is this Friday! If you work in any non I.i.d learning settings, consider submitting to CoLLAs!

We're thrilled to see the Workshop on Weight-Space Symmetries coming to #ICML2026! Huge shoutout to our postdoc @KateLobacheva for co-organizing it. We're excited for the ideas and discussions this workshop will bring to the community!

5/ This changes what engineering work looks like. Less: - writing every line manually More: - structuring problems - defining constraints - validating outputs The model accelerates execution, but only within the system you design.

4/ The loop is where reliability comes from You don’t rely on a single generation. You: - generate - execute - observe failures - feed that back in The system converges through this iteration, observed every once in a while by you!

Quoting this from Aidan, who presented this as a talk to the lab last Friday. If you think LLMs are unreliable for coding, this is the missing piece: reliability comes from how you structure the work around the model. The article lays this out very concretely. 🧵↓

I wrote an article on agentic coding for beginners after my talk at @apsarathchandar @ChandarLab group. We cover history of AI coding tools, the importance of model harnesses, and general principles in simple research workflows. Feedback is very welcome! aidanli.dev/writing/articl…

🗣️ Shoutout to the authors: @pranshumalviya8, @ravi_iitm and @apsarathchandar!!! (published at CoLLAs 2022). 🔗 Learn more at: arxiv.org/abs/2105.05155

🔥 This was the first work to show that you can successfully use adaptive gradient optimizers for lifelong learning and still beat SGD!!! (i.e. RMSProp < SGD < TAG-RMSProp) 🔥 Across benchmarks, TAG also improved final accuracy over baselines like ER and A-GEM.

In lifelong learning, continuously learning new tasks can cause ML models to forget previously acquired knowledge. For this, our lab introduced TAG (Task-based Accumulated Gradients): a general wrapper on top of adaptive gradient optimizers. 📈