PushingBoundaries at NakbaVirality Shared Task: Recursive Prompt Improvement for Multimodal Virality Classification

Abstract

This paper describes our participation in the NakbaVirality shared task at the NakbaNLP Workshop (LREC–COLING 2026). We investigate Recursive Prompt Improvement (RPI), an instruction-level optimization strategy for virality classification in high-stakes geopolitical discourse. In this work, we propose a self-supervised approach to iteratively improve the classification prompt without human intervention. We begin with a basic prompt that guides the LLM to perform multi-class classification, incorporating contextual information about the tweets. After obtaining predictions, we identify misclassified tweets and feed them back to the model with an instruction to refine and improve the original classification prompt. This process is repeated over multiple iterations to assess whether performance improves over time. Our results show a remarkable improvement in F1 score from the first iteration to the final one. Although the proposed method does not reach the accuracy of models fine-tuned directly on task-specific data, it demonstrates that iterative, self-supervised prompt refinement can serve as a viable proxy for fine-tuning. By leveraging the model’s own errors as feedback, this approach reduces reliance on computationally expensive training procedures and heavy GPU usage, while preserving much of the adaptability typically associated with fine-tuned models. This paradigm opens promising avenues for resource-efficient model adaptation and suggests new directions for scalable, low-cost performance improvement without traditional fine-tuning. The code has been shared in Github.

Resources

Details