Defining facilitation as speech produced by facilitators imposes a low performance ceiling on predictive facilitative models¶

Claim · All claims
Evidence · 1 study · associational r1 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? In the facilitator detection task, models score unreliably across diverse datasets, indicating a structural issue with the datasets rather than confusion with the moderator role. → Dimitris Tsirmpas 2026

Evidence¶

Dimitris Tsirmpas 2026¶

Dimitris Tsirmpas, Katerina Korre, John Pavlopoulos. (2026). To Facilitate or not to Facilitate: Human and LLM Facilitator Tendencies in Online Discussions. https://arxiv.org/abs/2607.28643 doi:10.48550/arxiv.2607.28643

q2 · i? · associational · r1

Facilitator detection (NLU) experiments with the same six LLMs and ModernBert across PEFK datasets (Table 5). Despite isolated strong results such as Qwen32B on CeRI (F1p=0.904, F1n=0.971), models "do not score reliably across diverse datasets" (e.g., LLaMa70B F1p=0.500 on WT).

"such as Qwen32B achieving high scores on CeRI (F1p=0.904, F1n=0.971)–the models generally do not score reliably across diverse datasets (e.g., LLaMa70B in WT yields F1p=0.500)"

Discussion¶

Related Claims¶

• ModernBert classifiers are more consistent than out-of-the-box LLMs at facilitation timing prediction, outperforming them in three of five datasets — related
• Unreliability in an outcome measure reduces the statistical power of treatment-control comparisons — a broader claim this one bears on