Investigating Noise and Silence in Large Language Models (C05)

• Subject Area: Methods in Artificial Intelligence and Machine Learning ; Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
• Term: since 2026
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 551281144

This project investigates how large language models (LLMs) internally represent and respond to the phenomena of noise and silence. We treat noise as a structural disruption within model representations and silence as a learnable discourse feature that conveys pragmatic functions like hesitation and turn-taking. Using fine-tuned models and specially annotated corpora, we explore these phenomena across tasks using Visual Analytics tools that explain how silence and noise propagate through model layers. The project contributes both technical insight and linguistic theory by revealing latent mechanisms in LLM behavior.

• DFG Programme: Collaborative Research Centres
• Subproject of: SFB 1760: SiNoSi: Silence, Noise and Signal in Language
• Applicant Institution: Universität Konstanz
• Project Head: Professor Dr. Daniel Keim