If I had to bet on LLM infrastructure impact, this might become the most important paper from my lab. This ICLR paper intrinsically parallelizes language models, building on ideas from normalizing flows. 🚀

Felix Draxler@drrelax.bsky.social · 4mo ago

LLMs are autoregressive and slow? No! Parallel Token Prediction decodes multiple consistent tokens in one model call. PTP allows arbitrary dependencies in one call, unlike discrete diffusion. Practical: 2.4x speedup github.com/mandt-lab/ptp ICLR: Apr 23, morning poster P3-#608

Really excited to see this. I’ve been involved since the early AABI days, and it’s been incredible to see the trajectory — from NeurIPS/ICML workshops, to a symposium, and now to a full conference as ProbML. Huge thanks to the program/general chairs and the whole organizing team!

Symposium on Probabilistic Machine Learning@probml.bsky.social · 6mo ago

ProbML 2026 (formerly AABI) invites submissions on probabilistic ML (both Bayesian and otherwise!), July 5 in Seoul (co-located with ICML). Website: probml.cc. Tracks: proceedings (PMLR), workshop, fast track. New focus includes applications in healthcare and climate! Submit by: 20 March 2026.

Recently gave a @leapstc.bsky.social lecture at Columbia and at UCLA on a question I’m excited about: How can we design diffusion models for scientific inference—uncertainty-aware, calibrated, steerable, and heavy-tailed? www.youtube.com/watch?v=QeLZ...

Scientific Inference with Diffusion Generative Models

YouTube video by LEAP

youtube.com

Congratulations to Metod and the team on winning the ICLR QUESTION workshop's Best Paper Award! 🎓

Metod Jazbec@metodjazbec.bsky.social · last yr.

Great too see that our Generative Uncertainty won the best paper award at the ICLR QUESTION workshop (uncertainty-foundation-models.github.io). If you're interested in what Bayesian/ensembling methods can bring to the world of diffusion models, check out the paper 👇 arxiv.org/abs/2502.20946

Just gave a talk on Scientific Inference with Diffusion Models at ETH AI Center, sharing our recent work—from test-time control and distributional matching to uncertainty calibration. Great crowd, thoughtful questions, nice view. Thanks, Julia Vogt, for hosting!

I am so excited to share our latest research paper "Transferring climate change physical knowledge" published in @pnas.org! 🧵 1/4 www.pnas.org/doi/10.1073/...

Thrilled to share that my student Justus Will and former student Yibo Yang had their work selected as an ICLR 2025 Oral (top 2%)! Presenting the first runtime-efficient progressive coding method using diffusion models. 👇

@justuswill.bsky.social · last yr.

Excited to present some recent work on developing "Progressive Compression with Universally Quantized Diffusion Models", accepted as an Oral at ICLR'25. 🧵1/4

Excited to start my sabbatical at ETH Zurich! The CS department treated me to a beautiful apartment. In exchange, I brought the California sun with me. 😎

Happy to announce three #ICLR2025 papers: 1️⃣ Heavy-tailed diffusion models (Kushagra Pandey + #Nvidia collaborators) 2️⃣ Progressive Compression w/ Universally Quantized Diffusion (Yibo Yang + Justus Will) 3️⃣ AstroCompress Benchmark (with UC Berkeley Physicists) More details soon! 🚀

Excited about 🎓 Generative Modeling Summer School / Statlearn 🗓️ Mar 31 - Apr 4, 2025, France ℹ️ gemss.ai 🎟️ Deadline: Jan 27, 2025 🧑‍🏫 Yingzhen Li, @stephanmandt.bsky.social, Aude Sportisse, Anna Korba, @glouppe.bsky.social, @jesfrellsen.bsky.social, @pamattei.bsky.social, @jmtomczak.bsky.social, TBA

Congratulations to newly minted PhD, Gavin Kerrigan, @gavinkerrigan.bsky.social . Not too hard to spot who Gavin is in the photo alongside his thesis committee.