3 papers presented at PerCom2026

21 March 2026, by Alhassan Abdelhalim

Photo: Janick Edinger

From March 16 to 20, our research group presented three papers at the IEEE PerCom 2026 conference in Pisa, Italy. The papers were accepted in the workshops PerFail, PerComAI, and TrustSense.

LLMs Explain't: A Post-Mortem on Semantic Interpretability in Transformer Models (PerFail)

Authors: Alhassan Abdelhalim, Janick Edinger, Sören Laue, Michaela Regneri

This paper examines widely used methods for explaining large language models and shows that some popular interpretation approaches can produce convincing results without actually revealing how the models work. The findings highlight the need for more rigorous and reliable explainability methods, especially when LLMs are deployed in pervasive and distributed systems.

Context-Aware Inference in Segmentation with Block-skipping Early Exit Networks (PeRConAI)

Authors: Lennart Bengtson, Paula Gómez Matos, Philipp Kisters , Janick Edinger

This work presents a new approach for making semantic segmentation models faster on resource-constrained devices. By combining early exits with dynamic block-skipping, the paper shows how potential computational savings can be turned into actual runtime improvements for real-time AI applications.

Beyond the Beaten Path: Using Gamification to Expand Urban Crowdsensing Coverage (TrustSense)

Authors: Inga Lange; Juliane Namneck; Lennart Bengtson; Philipp Kisters; Janick Edinger

This paper investigates whether gamification can motivate people to take less frequently used routes in order to improve urban crowdsensing coverage. The results suggest that simple elements such as badges and leaderboards alone may not be enough, and that more personalized incentives are needed to encourage broader participation.

We are especially proud that the TrustSense paper received the Best Paper Award.