Explainable AI: Context-Aware Layer-Wise Integrated Gradients for Explaining Transformer Models

Transformers are "black boxes" that need better explanation tools: The paper identifies that while Transformer models perform exceptionally well across many tasks, their deeply layered structure makes it nearly impossible to understand how they reach their decisions—a critical problem for responsible AI deployment.

Existing explainability methods are fragmented and incomplete: Current approaches either only look at the final layer, examine individual tokens in isolation, or focus solely on attention patterns, missing the full picture of how a model's decision actually forms across its internal layers.

CA-LIG provides a unified framework that traces decision-making through all layers: The proposed method tracks how information flows through each layer of a Transformer, showing both supporting and opposing evidence for predictions—enabling more complete explanations than previous approaches.

The framework works across diverse real-world applications: The method was successfully tested on multiple practical tasks including sentiment analysis, hate speech detection in low-resource languages (using XLM-R and AfroLM), and image classification, demonstrating broad applicability rather than narrow laboratory results.

Better explainability enables more trustworthy AI systems: By making Transformer decision-making more transparent and interpretable, this work supports the responsible deployment of AI systems in high-stakes domains where understanding why a model made a decision is as important as the accuracy of that decision.