MATH4AI at ICML 2026 in Seoul

MATH4AI at ICML 2026 in Seoul

Our PhD students Julius Hege and Maria Matveev travelled to Seoul to represent MATH4AI at the International Conference on Machine Learning (ICML) 2026. The group was represented with two main conference papers and one workshop paper.

Conflicting Biases at the Edge of Stability: Norm versus Sharpness Regularization
Main Conference Paper

Led by Maria Matveev and Vít Fojtík, this paper investigates how different implicit regularization effects compete with each other when training neural networks with large learning rates. The work demonstrates that neither effect alone is sufficient to explain in general why neural networks generalize well.

Authors: Maria Matveev, Vít Fojtík, Hung-Hsu Chou, Gitta Kutyniok, and Johannes Maly

Proxy Scoring Enables Benchmarking LLM Forecasters Without Waiting for Outcomes
Workshop Paper — Forecasting as a New Frontier of Intelligence

This work introduces proxy scoring methods to evaluate LLM-based forecasters without needing to wait for real-world outcomes to resolve. The scores obtained correlate strongly with resolved-outcome metrics such as the Brier score on existing LLM forecasting data.

Authors: Julius Hege and Gitta Kutyniok

ExPLAIND: Unifying Model, Data, and Training Attribution to Study Model Behavior
Main Conference Paper

This paper proposes a unified framework to attribute model behavior to the model, the data, and the training process, based on the exact path kernel. The framework provides a versatile interpretability approach, successfully applied to algorithmic and general language modeling tasks.

Authors: Florian Eichin, Yupei Du, Philipp Mondorf, Maria Matveev, Barbara Plank, and Michael A. Hedderich

Congratulations to all authors, and thank you to the ICML organizers for an insightful and successful conference!