relAI work at ICML 2026

mc

We are happy to announce that relAI will be represented in Seoul, South Korea, from July 6 to 11 at the Forty-Third International Conference on Machine Learning (ICML2026)!. ICML is widely recognized as one of the top three most influential conferences in machine learning and artificial intelligence research.

📖 About relAI Publications at ICML

  • relAI contributes fourteen publications to the conference, including nine in the main track.
  • relAI PhD student Emre Kavak will present a Spotlight paper (ranked in the top 2.2%).
  • Five relAI papers will be presented at ICML Workshops, with contributions from three relAI MSc students.

🤝Meet relAI Students

If you are attending ICML we encourage you to engage in discussions about relAI research with the following relAI students present at the conference:

You can find their research papers in the list below.

Full list of relAI publications at ICML 2026:

    Spotlight Presentation - Main Track


  1. DISCO: Mitigating Bias in Deep Learning with Conditional Distance Correlation
    Emre Kavak, Tom Nuno Wolf, Christian Wachinger

    2. Posters - Main Track


  2. Certifying Graph Neural Networks Against Label and Structure Poisoning
    Lukas Gosch, Xichuan Chen, Yan Scholten, Stephan Günnemann
  3. Rank-Learner: Orthogonal Ranking of Treatment Effects
    Henri Arno, Dennis Frauen, Emil Javurek, Thomas Demeester, Stefan Feuerriegel
  4. Interpretable Self-Supervised Learning via Representer Landmarks and Nyström Approximation
    Maedeh Zarvandi, Michael Timothy, Theresa Wasserer, Debarghya Ghoshdastidar
  5. SAD-Flower: Flow Matching for Safe, Admissible, and Dynamically Consistent Planning
    Tzu-Yuan Huang, Armin Lederer, Dai-Jie Wu, Xiaobing Dai, Sihua Zhang, Hsiu-Chin Lin, Shao-Hua Sun, Stefan Sosnowski, Sandra Hirche
  6. Frequentist Consistency of Prior-Data Fitted Networks for Causal Inference
    Valentyn Melnychuk, Vahid Balazadeh, Stefan Feuerriegel, Rahul G. Krishnan
  7. Reading Between the Tokens: Improving Preference Predictions through Mechanistic Forecasting
    Sarah Ball, Simeon Allmendinger, Niklas Kühl, Frauke Kreuter
  8. Don't Walk the Line: Boundary Guidance for Filtered Generation
    Sarah Ball, Andreas Haupt
  9. ExPLAIND: Unifying Model, Data, and Training Attribution to Study Model Behavior
    Florian Eichin, Yupei Du, Philipp Mondorf, Maria Matveev, Barbara Plank, Michael A. Hedderich
  10. Conflicting Biases at the Edge of Stability: Norm versus Sharpness Regularization
    Vit Fojtik, Maria Matveev, Hung-Hsu Chou, Gitta Kutyniok, Johannes Maly

    Workshops

  1. Proxy Scoring Enables Benchmarking LLM Forecasters Without Waiting for Outcomes
    Julius Hege, Gitta Kutyniok
    ICML 2026, Forecasting as a New Frontier of Intelligence Workshop
  2. What Intermediate Layers Know: Detecting Jailbreaks from Entropy Dynamics
    Sofiia Nikolenko, Michele Papucci, Mina Rezaei, Shireen Kudukkil Manchingal
    ICML 2026, The 2nd Workshop on Epistemic Intelligence in Machine Learning
  3. Bigger Is Not Better: Inverse Scaling and Arbitration Failure in Counterfactual Visual Grounding
    Fabian Grob, Sanghwan Kim, Cordelia Schmid, Zeynep Akata
    ICML 2026, Mechanistic Interpretability Workshop
  4. On the Uncertainty in Prior-Data Fitted Network Pretraining
    Manuel Hülskamp, Julius Kobialka, Emanuel Sommer, David Rügamer
    ICML 2026, 2nd Workshop on Foundation Models for Structured Data (FMSD)
  5. ExPLAIND: Unifying Model, Data, and Training Attribution to Study Model Behavior
    Florian Eichin, Yupei Du, Philipp Mondorf, Maria Matveev, Barbara Plank, Michael A. Hedderich
    ICML 2026, Mechanistic Interpretability Workshop

We are pleased to share relAI’s contribution to the European Embodied Robotics Week, whichtook place last week in Munich. The event was organizedby RoboTUM together with START Munich and the ESRA (European Student Robotics Association) network.

The Robotics Festival convened a diverse range of stakeholders from the European robotics and physical artificial intelligence ecosystem. Specialists, students, and robotics enthusiasts participated in city-wide events, including open houses at robotics laboratories, makerspaces, and studios throughout Munich, as well as a hackathon.

The event ended with a summit featuring talks and panels from leaders in embodied intelligence. relAI was represented in the panel discussion “Dexterity and Manipulation for robotics”  by relAI Fellow Prof. Khadiv.

Machine learning (ML) - based techniques enable robots to perform a range of tasks in complex environments. Research in this area has increased significantly in recent years, with various ML methods being explored to allow robots to make decisions based on task requirements and changes in their surroundings. However, some of the tested approaches, such as reinforcement and imitation learning, often struggle with training stability and capturing the multimodal nature of behavior. In this blog post, Tzu-Yuan Huang introduces a promising ML approach for robotic decision-making: diffusion models.

👉 If you want to learn more about it, check it out: https://zuseschoolrelai.de/blog/robotic-decision-making-diffusion-models/

relAI is proud to announce an outstanding achievement – a first author publication of Moritz Knolle, one of relAI’s PhD students, in Nature journal.

Medical AI models are increasingly utilized in applications such as diagnosing and remotely treating patients. While these models have proven valuable to both practitioners and patients, concerns remain about the privacy of patients whose information is used to train these AI systems. This issue has been examined by relAI PhD student Moritz Knolle, in a study published this week in Nature. The analysis revealed privacy vulnerabilities in medical AI models, emphasizing the importance of reliable AI research in medicine and healthcare 🩺.

🔍 Check the summary below and the article to learn more about the study!.

Summary of the article

Individuals whose data are used to train medical AI models may be at risk of being identified in cyber-attacks, according to a Nature paper published this week. Underrepresented groups may face disproportionately higher risks of having their data compromised, the study indicates. The researchers find these individuals are not accounted for in current risk assessments and call for further mitigation and strict access control. 

Medical AI models may improve global health outcomes, especially in areas in which specialized expertise is not available. Yet, the sensitive data used to train these models may be exposed to privacy attacks. Membership inference attacks (MIAs) are used by attackers to determine whether an individual’s data were used to train a model. From these attacks, a patient’s medical data and private information can be determined. Previous research on data risk has been determined by whole datasets, and does not take an individual’s risk into account.

Moritz Knolle and colleagues conduct a privacy audit to focus on individual privacy risk, finding that medical AI models may pose a privacy risk to individual data contributors. Using seven large datasets made up of real-world clinical data, including medical images, electrocardiograms and electronic health records, the authors determine the most vulnerable among data-contributing patients. They find that at an individual level, those targeted by the MIAs were successfully done so with almost no error. At a group level, those identified as underrepresented in datasets include people with rare diseases, people from a minority racial group or, socioeconomic status, or those having the less-common gender. With more distinctive data that are encoded by AI models, these groups and individuals are found to be more vulnerable and disproportionately exposed to privacy attacks. The authors find the instances of successful MIAs attacks rise with model capacity and size. 

These findings show privacy attacks, such as MIA, are more effective at successfully targeting on an individual level than currently thought. The authors conclude that privacy risk assessment must now take individual risk into account, and vulnerable models be further protected."

👉 Link to article: https://www.nature.com/articles/s41586-026-10688-0

🎉 Congratulations to relAI PhD student Moritz Knolle, relAI Fellow Daniel Rückert, former relAI Fellow Georgios Kaissis, and co-authors for the fantastic work!

We are proud to announce that relAI director Gitta Kutyniok has been appointed Vice President of the Berlin-Brandenburg Academy of Sciences and Humanities (BBAW). Her term will begin on October 1, 2026.

The Berlin-Brandenburg Academy of Sciences and Humanities has a remarkable 325-year history of bringing together brilliant scholars and scientists from various backgrounds and disciplines, including 82 Nobel Prize winners. This Academy plays a vital role in our society by not only advancing research in the humanities but also addressing critical scientific and social issues through collaboration. We deeply value its mission to foster understanding and dialogue between the scientific community and the broader society, as this connection is essential for tackling the challenges we face together.

🎉 Congratulations!

Last week, the relAI family gathered for our annual retreat in the beautiful and serene Bad Kohlgrub, from June 17th to 19th. During this time, relAI students, fellows, directors, and management came together to share ideas and visions for the future of our school, the role of AI in education, and the future of science in the era of agentic AI.

The program included keynote talks from relAI fellows Enkelejda Kasneci, Sandra Hirche, Niki Kilbertus, and Christoph Kern. It also featured lightning talks by relAI students, group discussions, and fellow/student one-on-one sessions. Additionally, there was plenty of time for networking and social activities.

Keynote by Enkelejda Kasneci
Keynote by Sandra Hirche
Keynote by Niki Kilbertus
Keynote by Christoph Kern

In the one-on-one sessions, relAI fellows and students connected directly to discuss their research experiences, career paths, and future aspirations. These conversations provided a unique mentorship opportunity, often sparking new perspectives for both mentors and mentees.

This year's group discussions focused on questions about the two strategic pillars that are currently reshaping the landscape of reliable AI:

  • Future of Science in Times of Agentic AI
    • How do agentic workflows change the way we conduct and validate our research?
    • What are the theoretical and practical requirements for genuinely dependable autonomous systems?
  • AI for Next Generation Education
    • What roles will AI and human teachers play moving forward? How can we lead the adoption and ethical use of AI technologies?
    • How can we establish standards for reliability in pedagogical AI tools?

A highlight of this year’s retreat was the presentation of the final relAI certificates to MSc and PhD students who completed the program. This moment provided an opportunity to reconnect with relAI alumni and learn about their career paths after leaving relAI.

This inspiring retreat strengthened our community and paved the way for our efforts to shape the future of reliable AI.

A big thank you to everyone who participated and contributed to making this event a success!

Frontier AI systems have recently solved IMO problems, discovered new mathematical constructions, and resolved open Erdős problems. Yet a more mundane question remains: 🤔how should a working mathematician actually use these tools day to day?

In this talk, titled “The Agentic Researcher: Turning AI Coding Agents into Research Assistants” Emil Partow, PhD Student of our relAI Director Prof. Gitta Kutyniok presented a recent paper from Prof. Pokutta (Zuse-Institut Berlin) that offers a concrete answer. The authors propose a five-level taxonomy of AI integration into research, ranging from classical work without AI to fully autonomous research loops. They have implemented this idea within an open-source framework.

Following a detailed presentation by Emil Partow, members of relAI and Prof. Kutyniok's research group gathered to discuss this forward-looking topic. The presentation explained the open-source tool that implements core research "commandments" (such as preventing the falsification of experimental data) via a practical, actionable loop. Participants then discussed how AI agents are already shaping research methodologies, what is required to implement these workflows successfully, and how to ensure human oversight remains at the center of the process. Emil also shared a practical case study demonstrating the tool in action, sparking a broader reflection on the evolving role of AI in modern research.

Machine learning models, such as ChatGPT and those used in autonomous driving, are becoming essential tools in our daily lives. However, the existence of Adversarial Examples demonstrates that these systems are not free from vulnerabilities. To ensure their reliability, it is crucial to proactively address the potential risks associated with their use in critical safety applications.

In a recent blog post, relAI PhD student  Lukas Gosch introduces the concept of Adversarial Examples and discusses Certifiable Robustness, a methodology designed to combat 🛡️ them.

What is an Adversarial Example?

As Lukas Gosch outlines, Adversarial Examples are deliberately crafted inputs that cause machine learning models to misclassify data. For example, the strategic placement of stickers on traffic signs can lead to incorrect identification of road signs by machine learning systems used in autonomous vehicles. Additionally, if an adversary manipulates the training data upon which these models are built, this too qualifies as an Adversarial Example.

How to combat Adversarial Examples?

In his post, Lukas describes Certifiable Robustness, a methodology for verifying the resilience of machine learning systems against adversarial examples, and explores the challenges associated with it.

👉 Check it out https://zuseschoolrelai.de/blog/a-beginners-guide-to-certifiable-robustness/

🎤 Michael Lachner, CEO of Aqarios and one of the first relAI alumni, is developing solutions that combine quantum computing, AI, and advanced optimization algorithms to solve complex challenges in industry and business.  In his interview with DAAD, he talks about his vision for the future of quantum computing in industry and Germany’s role as a leader in this transformation.

The interview in a 📸 snapshot:

🔹 Companies should start preparing today to leverage quantum computing and avoid missing the next major technological leap.

🔹 Germany has strong potential to play a leading role in the global AI and quantum landscape – driven by top-tier research, growing investment, and innovative start-ups.

🔹 Lachner values the academic excellence and networking opportunities provided by relAI and Zuse Schools across Germany.

👉 Read the complete interview: https://www.daad.de/de/der-daad/daad-journal/themen/2026/zuse-school-alumnus-michael-lachner-im-portraet/