Maximizing Scoring Divergence in Automated Essay Assessment with LLaMA-Based Meta-Attention Networks

Automated essay scoring systems often struggle with objectivity and handling diverse writing styles, leading to biased evaluations. This study proposes a Meta-Learning approach that exploits biases within large language models (LLMs) to maximize scoring system divergence for more objective evaluations. We introduce the Dynamically Guided Meta-Attention Network (DGMAN), integrating an ensemble of LLaMA models, dynamic attention mechanisms, adversarial perturbations, and meta-learning optimization. DGMAN generates input texts that increase scoring discrepancies while preserving text quality. Using LLaMA-7B, LLaMA-13B, and LLaMA-30B, the dynamic attention mechanism adjusts model weights based on context, enhancing horizontal variance between scoring systems. Adversarial perturbations fine-tune texts to generate larger divergences while maintaining fluency. Experimental results show that DG-MAN improves model divergence, optimizes fluency, and ensures uniqueness. Future work will focus on optimizing the method for various scoring systems and exploring applications in fairness enhancement and personalized evaluation.