@inproceedings{zou-etal-2026-sutta,
    title = "sutta at {S}em{E}val-2026 Task 12: A Multi-Perspective Retrieve-Verify-Aggregate Framework for Abductive Event Reasoning",
    author = "Zou, Junliu  and
      Yang, Liang  and
      Zeng, Jingjie",
    editor = "Kochmar, Ekaterina  and
      Ghosh, Debanjan  and
      North, Kai  and
      Komachi, Mamoru",
    booktitle = "Proceedings of the 20th {I}nternational {W}orkshop on {S}emantic {E}valuation (2026)",
    month = jul,
    year = "2026",
    address = "San Diego, California, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/ingest-acl-workshops/2026.semeval-1.323/",
    pages = "2564--2574",
    ISBN = "979-8-89176-414-9",
    abstract = "We present our system for SemEval-2026 Task 12: Abductive Event Reasoning (AER). The task asks models to identify the direct causes of real-world events from multiple-choice options using retrieved documents. Rather than fine-tuning on the training data, we built a zero-shot ``Retrieve-Verify-Aggregate'' pipeline around Qwen3-8B. We first isolate relevant evidence using BM25 and cross-encoder reranking. To evaluate causal links, we prompt the model with several distinct ``personas'' and aggregate their independent decisions through majority voting. Our system scored 0.7614 on the official test set. This performance suggests that strict retrieval combined with diverse reasoning prompts can help compact open-source models ignore irrelevant context and perform complex causal inference, entirely without task-specific training."
}