You can now use PDFs as images: Just pass a PDF file to the image function and Typst will handle the rest. PDF images let you use vector and raster graphics from almost any other app. PDF images can be used for every Typst export target, including HTML.

cc @polymathicai.bsky.social @flatironinstitute.org

Sorry I mistyped @sedielem.bsky.social handle!

cc @rubenohana.bsky.social, @michael-mccabe.bsky.social, @glouppe.bsky.social, François Lanusse and Shirley Ho

This work marks the final chapter of my PhD. Next up: writing my thesis and embark on a new adventure! Stay tuned 🚀

This work was conducted as part of my internship with @PolymathicAI at the @FlatironInst in New York 🗽 It was an amazing experience to be part of a talent-dense team in such an outstanding environment. I wholeheartedly recommend it to anyone!

The paper is full of details and insights. We also release the code and model weights. Go check it out!

Thanks to my co-authors ❤️

Paper: arxiv.org/abs/2507.02608
Data: github.com/PolymathicA...
Code: github.com/polymathicA...
Blog: polymathic-ai.org/blog/lostin...

I am beyond proud to finally release this work. We started with a simple question, and we ended up with totally unexpected results. We had to read and do tons of experiments to convince ourselves. S/O @sedielem for the insights on LDMs and V-information!

sander.ai/2025/04/15/...

While these results seemingly violate the data processing inequality, they are well aligned with @xuyilun2's theory of usable information, where a representation can hold more V-information from the point of view of a computationally constrained observer.

arxiv.org/abs/2002.10689

Our experiments also show that latent diffusion models are consistently more accurate than deterministic neural solvers while producing diverse, statistically plausible trajectories.

When we started this project, we expected accuracy to degrade as the compression rate increased. To our surprise, we found that accuracy remained constant or even improved!

Our methodology is quite simple: we train auto-encoders to compress the state of dynamical systems (fluids, stars, ...) and train a latent diffusion model to emulate the dynamics in the compressed space. We then study the impact of the compression rate on the emulation accuracy.

It would be fantastic if EurIPS posted the number of registrations. It could convince more people to register and have a snowball effect. Maybe more people would go to EurIPS than NeurIPS!