Deutsch
English
Three Papers Accepted at EMNLP 2026 Main Conference
Three papers co-authored by Jannik Brinkmann, a member of Prof. Dr. Christian Bartelt’s research group, have been accepted to the main conference of EMNLP 2026. EMNLP is a leading conference for natural language processing. It holds an A* rating—the highest category—in the ICORE 2026 conference ranking.
All three papers study how large language models work internally.
1. “Mood Matters: How Syntactic Sensitivity Undermines Safety Alignment” shows that a model’s safety response can depend on sentence structure, not only on meaning. Tests across 16 models found that changing how a harmful request is phrased can weaken refusal behavior.
The paper was written by Alina Klerings, Jannik Brinkmann, Heiner Stuckenschmidt, and Simone Paolo Ponzetto. Read the preprint on arXiv.
2. “The Interlingua Hypothesis: LLMs Translate via a Latent Task-agnostic Feature Space” studies whether language models translate through a shared internal representation space. Its results provide several forms of evidence consistent with this idea.
The paper was written by Jacob Brinton, Jannik Brinkmann, Mark Crovella, and Aaron Mueller. A public preprint is not yet available.
3. “In-Context Learning Beyond Copying: A Training-Time Analysis of Abstractive ICL” asks whether models must first learn to copy patterns before they can learn new tasks from examples. The results show that more abstract forms of learning can develop even when copying is strongly reduced.
The authors are Kerem Şahin, Sheridan Feucht, Adam Belfki, Jannik Brinkmann, Aaron Mueller, David Bau, and Chris Wendler. Read the preprint on arXiv, published under the earlier title “In-Context Learning Without Copying.”
Together, the papers offer new insights into model safety, learning from context, and machine translation. All three were accepted in the EMNLP Main Conference track “Interpretability and Analysis of Models for NLP.”
