PLEX: Perturbation-free Local Explanations for LLM-Based Text Classification

PLEX offers a faster alternative to existing explanation methods: Traditional explainability approaches like LIME and SHAP require generating thousands of perturbed sentences to explain a single prediction, creating significant computational burden. PLEX eliminates this need through one-time training, dramatically reducing the time and computational resources required to explain LLM predictions.

The method maintains accuracy while improving efficiency: PLEX shows over 92% agreement with established explanation methods (LIME and SHAP) across four different text classification tasks (sentiment analysis, fake news detection, COVID-19 misinformation, and depression detection), demonstrating it produces reliable explanations without the computational overhead.

Explanations remain trustworthy under rigorous testing: When researchers removed words that PLEX identified as influential, the classification accuracy declined in similar ways as it did when using LIME and SHAP explanations, validating that PLEX correctly identifies which words actually matter for the model's decisions.

This advancement supports more practical AI deployment: By reducing computational costs, PLEX makes it feasible to provide real-time explanations for LLM-based text classification systems in real-world applications, helping organizations better understand and audit their AI systems' decision-making processes.