Clutch or Cry at SemEval-2026 Task 12: Offline Retrieval-Augmented Generation with Frozen DeBERTa for Abductive Event Reasoning
Year: 2026
Type: conference-paper
Abstract: This paper presents our system for SemEval-2026 Task 12: Abductive Event Reasoning.We initially approached this task using direct fine-tuning of standard language models; however, the limited volume of training data caused severe overfitting, yielding up to 95% accuracy on training splits but failing to generalize on unseen test data.Attempts to mitigate this by deploying smaller models resulted in token-limit truncations and random guessing.Given the strict all-or-nothing evaluation metric, these guesses were heavily penalized.To overcome these compounding issues, we pivoted to a two-stage offline Retrieval-Augmented Generation (RAG) pipeline.By decoupling semantic evidence retrieval from classification and fine-tuning a partially frozen deberta-v3-large model, we constrained the model's reasoning space to explicitly retrieved clues.This pipeline successfully curbed overfitting and random guessing, achieving a robust Exact Match accuracy of 0.72.
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166868553