news
| Nov 04, 2026 | Members of the ELLIS Units Berlin and Warsaw met on 4–5 November 2026 at the Polish Embassy in Berlin and the Google AI Center Berlin to exchange ideas and explore opportunities for collaboration between the two units. I gave a talk representing the ELLIS Unit Berlin on “Moving up and down explanation levels: Model Understanding and Scientific Discovery” and our work on AI interpretability and model explanations. Link: Warsaw and Berlin researchers strengthen connections across ELLIS |
|---|---|
| Oct 01, 2026 | Gave a talk at the University of Oslo in the Language Technology Group research seminar. I presented my research on explainable AI and mechanistic interpretability, exploring how we can understand the inner workings of modern language models from input features and internal representations to feature interactions and algorithmic mechanisms. Many thanks to the LTG group for having me. Slides. |
| Jul 08, 2026 | Our paper on discovering algorithmic primitives was presented at ICML 2026. AlgoTrace introduces a framework for tracing and steering algorithmic primitives in LLM latent space, showing primitives can be extracted, injected, and composed to steer multi-step reasoning. Sources: ICML poster page · OpenReview |
| Jul 08, 2026 | Our new preprint, coauthored by Maximilian S. Ernst, Lorenz Linhardt, and Aaron Peikert, introduces Distributed Sparse Interventions (DSI), a method for identifying neuron sets whose targeted interventions elicit task-relevant behaviour in instruction‑tuned LMs. Sources: arXiv abstract · PDF |
| Oct 24, 2025 | Oliver Eberle has been awarded the 2025 Heinz‑Billing Prize for the Advancement of Scientific Computing. The award was announced 24 Oct 2025 (ceremony 23–24 Oct 2025) in recognition of contributions to digital humanities and scientific computing. Sources: MPIWG announcement · BIFOLD announcement |