Fredrik Johansson

I'm hiring a post-doc for the Healthy AI Lab 📣. Come join us to work on machine learning to improve decision-making, e.g., in health applications. If you can't stop thinking about research problems until you make progress, you may be the right fit! chalmers.se/en/about-chalm…

Key to this is to decouple environment interaction from language generation while maintaining the reasoning capabilities of pre-trained models. Project page: expa-rl.github.io Pre-print: arxiv.org/abs/2510.07581 PS. Nick is on the job market!

🚨 Preprint on LLMs in external environments: Zhongqi (Nick) Yue, a great post-doc in my lab, has led the development of EARL—a new reinforcement learning framework for LLMs to interact with external environments, greatly improving over text-only interaction in reasoning tasks.

Read Newton's PhD thesis here: research.chalmers.se/publication/54…

Thank you Branislav Kveton, Sandeep Juneja, Slawomir Nowaczyk and Yevgeny Seldin for serving as Newton's grading committee and opponent!

Last Friday, Newton Mwai Kinyanjui defended his PhD thesis "Leveraging Structural Priors and Historical Data for Practical Treatment Personalization with Multi-Armed Bandits". It's been a pleasure having you in the lab, Newton! Looking forward to see the next chapter!

Anton's thesis: research.chalmers.se/publication/54… Lena's thesis: research.chalmers.se/en/publication…

This week has been an absolute joy for me as the leader of the Healthy AI Lab! Two of my students, Anton Matsson (3rd from right) and Lena Stempfle (2nd from right), defended their theses and became the first PhD graduates under my supervision 🎉 You will both be sorely missed!❤️

Going to ICML next week? Find Anton and talk to him about our work on Prediction models that learn to avoid missing values!

@StempfleLena @antmats94 @Mwai_Newton In a tree, features are used depending on the values and availability of other features. By penalizing reliance on missing values in computing predictions, we can learn trees that are unlikely to ask for a feature unless it is available (and predictive)

@StempfleLena @antmats94 @Mwai_Newton In this work, we show that you can fit interpretable models that are unlikely to even *ask for* features that are missing at test time by learning to avoid using them in the first place. For trees, this is contextual...

I'm thrilled to share that @StempfleLena's, @antmats94's, and @Mwai_Newton's paper "Prediction models that learn to avoid missing values" was accepted to ICML and awarded a spotlight! Arxiv here: arxiv.org/abs/2505.03393 More below👇

10 days to go to apply for this PhD student position in my lab!

Last but not least: The student will be co-advised by the fantastic Uri Shalit!

Deadline May 15!

Our CS department is listing 5 positions. Make sure to select "F. Johansson - Reducing waste in tabular machine learning by generalizing out-of-table" to indicate which project you would like to work in.

We have an *opening for a PhD student* in my group on machine learning generalization "out-of-table". Help build methods that learn from large volumes of tabular data to generate models for new tasks! Apply here: chalmers.se/en/about-chalm…

@StempfleLena In the "Expert study" paper, w. Arthur James, Julie Josse and Tobias Gauss, @StempfleLena investigates clinician familiarity and preferences for interpretable models in prediction with missing values using a large survey of French trauma physicians.

@StempfleLena In the first, w @StempfleLena, Yaochen Rao and the CorEvitas team, Anton explores how to represent history in modeling of clinical policy using interpretable machine learning. Most IML models are designed for tables, not sequences, but can be applied with the right representation

In case you missed it at ML4H, Anton Matsson presented "How Should We Represent History in Interpretable Models of Clinical Policies?" arxiv.org/pdf/2412.07895 and @StempfleLena 's "Expert Study on Interpretable Machine Learning Models with Missing Data" arxiv.org/abs/2411.09591