Yam at SemEval-2026 Task 4: Failure-Driven Prompt Evolution for Narrative Comparison

AbstractWe present a structured, parameter-free system for SemEval-2026 Task 4 on Narrative Story Similarity. Instead of treating similarity as scalar embedding proximity, we align model reasoning with the task ontology by decomposing each story into abstract theme, course of action, and outcome, and performing contrastive comparison over these dimensions. Our primary contribution is a closed-loop, failure-driven prompt optimization procedure that iteratively refines concise guideline documents while keeping model parameters fixed and reverting updates that degrade performance, thereby isolating improvements attributable to structured reasoning rather than representation learning. Ontology-aligned decomposition alone achieves 70% accuracy on both train and test sets; with controlled guideline evolution, performance improves to 76% on train and 73% on test without additional supervision or fine-tuning. These results demonstrate that structured prompt optimization can meaningfully enhance contrastive narrative reasoning in a fully parameter-free setting.