Training LLMs with verifiable rewards uses 1bit signal per generated response. This hides why the model failed.

Today, we introduce a simple algorithm that enables the model to learn from any rich feedback!
And then turns it into dense supervision.

(1/n)

Joint work with Jonas Wildberger, Frederik Träuble, Maximilian Mordig, Sergios Gatidis, Bernhard Schölkopf, and @arkrause.bsky.social

Workshop manuscript studying the format shift from structured to unstructured data (openreview.net/pdf?id=Wd05q...)

AdaCVD: Adaptable Cardiovascular Disease Risk Prediction from Heterogeneous Data using LLMs (arxiv.org/pdf/2505.24655)

Clinical notes are messy, inconsistent, and unstructured—yet they hold some of the most valuable signals in real-world clinical practice.

Join us today at ICML at the Foundation Models for Structured Data workshop to see how we can make sense of these notes!
📍 West Ballroom D