Explainable AI Berlin

Heading to the EMNLP BlackboxNLP Workshop this Sunday? Don’t miss @nfel.bsky.social and @lkopf.bsky.social poster on „Interpreting Language Models Through Concept Descriptions: A Survey“
aclanthology.org/2025.blackbo...

#EMNLP #BlackboxNLP #XAI #Interpretapility

Counterfactual explainers for dynamic graphs by Qu et al. by @scadsai.bsky.social
arxiv.org/abs/2403.16846
Explainable Biomedical Claim Verification by Liang et al. from @dfki.bsky.social
arxiv.org/abs/2502.21014

We were happy to see other explainability-themed posters:
A study of monosemanticity of SAE features in VLMs by Pach et al. from @munichcenterml.bsky.social
arxiv.org/abs/2504.02821
User-centered research for data attribution by Nguyen et al. from @tuebingen-ai.bsky.social
arxiv.org/abs/2409.16978

Manuel Welte presented ongoing work on intrinsic interpretability of transformer models through a novel approach for restructuring internal representations.

@lkopf.bsky.social and @eberleoliver.bsky.social presented the PRISM framework for multi-concept feature descriptions in LLMs.
arxiv.org/abs/2506.15538

We are grateful for the opportunity to present some of our work at the All Hands Meeting of the German AI Centers, hosted by @dfki.bsky.social in Saarbrücken.

Andreas Lutz @eberleoliver.bsky.social Manuel Welte @lorenzlinhardt.bsky.social @lkopf.bsky.social

#AI #XAI #Interpretability

Happy to share that our PRISM paper has been accepted at #NeurIPS2025 🎉

In this work, we introduce a multi-concept feature description framework that can identify and score polysemantic features.

📄 Paper: arxiv.org/abs/2506.15538

#NeurIPS #MechInterp #XAI

a black background with green text that says `` hello , world '' ALT: a black background with green text that says `` hello , world ''

This is the eXplainable AI research channel of the machine learning group of Prof. Klaus-Robert Müller at Technische Universität Berlin @tuberlin.bsky.social & BIFOLD @bifold.berlin.
Let's connect!
#XAI #ExplainableAI #MechInterp #MachineLearning #Interpretability