Targeted Visualization of the Backbone of Encoder LLMs

Encoder models like BERT pose explainability challenges: The paper identifies that while encoder-based LLMs are successful, they carry risks including bias and vulnerability to adversarial attacks, highlighting the need for better tools to understand how these models make decisions.

Current explainability methods have significant gaps: Existing approaches in NLP focus mainly on explaining individual predictions, but lack global visualization methods that can inspect overall model behavior—a capability that has emerged in other AI domains but remains underdeveloped for language models.

DeepView offers a new approach to model transparency: The researchers adapted DeepView, a visualization technique previously used for image classifiers, to work with BERT-based text classifiers, enabling two-dimensional visualization of how models make decisions across datasets rather than just for single inputs.

The method works across different model configurations: The visualization approach was tested on pre-trained models, fine-tuned models, and multi-task learning setups, suggesting it could be a practical tool for inspecting various real-world NLP systems used in practice.

Adversarial robustness can now be visually inspected: By applying DeepView to adversarially perturbed samples, the method enables researchers and practitioners to visually detect when models are vulnerable to adversarial attacks—a key responsible AI concern—rather than only discovering vulnerabilities through testing.