iiinova — where your digital self keeps evolving

I’m honored to present CBG at #icml! Come stop by at the morning 7/7 posters (Hall A #2701). Paper and code dgeyfman.github.io/calibrated-g.... Thanks to my amazing collaborators: Daniel Geyfman, Jan Groeneveld, Hyunsoo Lee, @tkaraletsos.bsky.social, and Stephan Mandt.

dgeyfman.github.io

000

On a Bayesian inference benchmark, CBG converges to the correct calibrated distribution while other methods clearly remain biased with more compute. CBG also sets a SOTA on the InverseBench black-hole imaging task, and scales up to natural images.

CBG takes diffusion steps using samples from a fast generative model, avoiding approximations existing methods use which introduce unavoidable bias. It comes in gradient-free and gradient-based variants that both become more calibrated with more compute.

100

AI for science needs calibrated models and uncertainty, but we show standard test-time diffusion guidance methods miss the Bayesian posterior. Our Calibrated Bayesian Guidance (CBG) fixes this, enabling Bayesian inference and setting a new SOTA for black-hole imaging.

Our CBG fits a posterior right, baselines (DPS, NDTM) are biased.

Huge thanks to amazing collaborators: Justus, Farrin, Theo, Sameer and Stephan! Meet us at #ICLR: Apr 23, morning poster P3-#608

000

GitHub - mandt-lab/ptp: Parallel Token Prediction for Language Models (ICLR 2026)

Parallelize the model you care about with our code: github.com/mandt-lab/ptp

github.com

Parallel Token Prediction for Language Models (ICLR 2026) - mandt-lab/ptp

We verify that this works in a speculative decoding experiment: We distill Vicuna-7B on conversations to predict the same output, but at greatly reduced latency. We achieve a 2.4x speedup over AR on diverse text tasks on one GPU, with 3.2x possible with optimized implementation.

Autoregressive models produce text by predicting the histogram of the next token. Random auxiliary variables determine the token to choose. Parallel Token Prediction directly learns what external randomness maps to which token. This allows to predict many tokens at once.