Clutch or Cry at SemEval-2026 Task 12: Offline Retrieval-Augmented Generation with Frozen DeBERTa for Abductive Event Reasoning

AbstractWe present our system for SemEval-2026 Task 12 on abductive event reasoning. Initial experiments with direct fine-tuning of large language models suffered from severe overfitting due to limited training data, while smaller models failed under context-length constraints, leading to random guessing under the strict Exact Match evaluation metric. To address these challenges, we propose a two-stage offline Retrieval-Augmented Generation (RAG) pipeline that separates semantic evidence retrieval from multi-label classification. We employ a dense retriever (all-MiniLM-L6-v2) to extract the single most relevant sentence (top-k=1) and feed it into a partially frozen DeBERTa-v3-Large classifier trained with BCEWithLogitsLoss. Freezing the lower 12 layers effectively mitigates overfitting while preserving pre-trained semantic knowledge. Our approach eliminates long-context truncation issues, reduces hallucination, and achieves a final Exact Match accuracy of 0.72 on the official test set.