Dimitris Tsirmpas, Katerina Korre, John Pavlopoulos — arXiv preprint (2026). 📄 Full text (arXiv)
Synthesis
This study asks when (not just how) LLMs should facilitate online discussions, creating PEFK, a corpus standardizing and aggregating facilitation datasets, and running the first survey on facilitation timing with expert facilitators and LLM-as-a-judge models.
Key asymmetry: humans are more cautious while LLMs are excessively eager to facilitate, although both are more certain when judging that facilitation is not needed.
Corrective attempts found trained ModernBert classifiers more reliable than alternative LLM setups, though existing datasets impose a relatively low performance ceiling — a benchmark-quality finding for automated discussion facilitation.
For online learning, the work informs when AI should intervene in discussion forums (MOOC-style and classroom), connecting facilitation timing to engagement and moderation research.
Related Pages
Citation
APA: Tsirmpas, D., Korre, K., & Pavlopoulos, J. (2026). To facilitate or not to facilitate: Human and LLM facilitator tendencies in online discussions. arXiv:2607.28643.