mdok-style at SemEval-2026 Task 10: Finetuning LLMs for Conspiracy Detection
TLDR
This paper finetunes Qwen3-32B with data augmentation and self-training for conspiracy detection, achieving 8th place in SemEval-2026 Task 10.
Key contributions
- Developed `mdok-style` system for SemEval-2026 Task 10: conspiracy detection in Reddit comments.
- Finetuned Qwen3-32B using data augmentation and self-training to overcome small training datasets.
- Ranked 8th out of 52 submissions (85th percentile) in the SemEval-2026 Task 10 competition.
- Showcased the effectiveness of a machine-generated text detection approach for conspiracy detection.
Why it matters
This paper introduces a highly effective method for detecting conspiracy beliefs in social media, a critical task for combating misinformation. By adapting techniques from machine-generated text detection, it demonstrates strong performance even with limited data. This work offers valuable insights for future research in text classification and misinformation detection.
Original Abstract
SemEval-2026 Task 10 is focused on conspiracy detection. Specifically, the goal is to detect whether a Reddit comment expresses a conspiracy belief. Our submitted mdok-style system utilizes data augmentation and self-training (to cope with a rather small amount of training data) to finetune the Qwen3-32B model for a binary text-classification task. The submitted system is very competitive, ranking in the 85th percentile (8th out of 52 submissions). The results shown that our approach, which originated in machine-generated text detection, can be used for conspiracy detection as well.
📬 Weekly AI Paper Digest
Get the top 10 AI/ML arXiv papers from the week — summarized, scored, and delivered to your inbox every Monday.