Post by Sebastian Steindl

Happy to announce that two of our new papers were accepted to #EMNLP2025 Findings: 1. On the Effectiveness of Prompt-Moderated LLMs for Math Tutoring at the Tertiary Level: We found that both groups (moderated and unmoderated) which received LLM assistance during an exercise phase outperform the control group during an un-assisted exam. Interestingly, we did not find signs of shallow learning, nor did we see a positive effect of the prompt moderation. Huge thanks to all co-authors: Fabian Brunner, Nada Sissuono, Dominik Schwagerl, Florian Schöler-Niewiera, Ulrich Schäfer 2. An Improved, Strong Baseline for Pre-Trained Large Language Models as Task-Oriented Dialogue Systems We show that no sophisticated mechanisms are required to enable current LLMs to perform Dialogue State Tracking. We provide an updated baseline for the performance of LLMs on MultiWOZ. A simple Self-Check mechanism can often greatly improve the performance. Huge thanks to all co-authors: André Kestler, Ulrich Schäfer, Bernd Ludwig