A Close Look at Decomposition-based XAI-Methods for Transformer Language Models

2025-02-21

AI Summary

This work addresses the lack of systematic comparative evaluation of decomposition-based XAI methods—specifically ALTI-Logit and Layer-wise Relevance Propagation (LRP)—in Transformer language models. We conduct the first quantitative and qualitative analysis of their attribution performance under a unified benchmark, focusing on subject–verb agreement, a syntax-sensitive linguistic task. Experiments span BERT, GPT-2, and LLaMA-3, and introduce AttnLRP, a novel LRP variant explicitly designed for attention mechanisms. We propose algorithmic and implementation optimizations and construct the first publicly available, manually annotated dataset dedicated to attribution evaluation in language models. Results reveal method-specific attribution biases and consistency disparities in syntactic reasoning, empirically validating AttnLRP’s superior modeling of attention dynamics. All code and data are open-sourced to support standardized, reproducible XAI evaluation for language models.

Technology Category

• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models
• Machine Learning: Large Multimodal Models (LMMs)
• Humans and AI: Explainable AI (XAI) for Human Understanding

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Various XAI attribution methods have been recently proposed for the transformer architecture, allowing for insights into the decision-making process of large language models by assigning importance scores to input tokens and intermediate representations. One class of methods that seems very promising in this direction includes decomposition-based approaches, i.e., XAI-methods that redistribute the model's prediction logit through the network, as this value is directly related to the prediction. In the previous literature we note though that two prominent methods of this category, namely ALTI-Logit and LRP, have not yet been analyzed in juxtaposition and hence we propose to close this gap by conducting a careful quantitative evaluation w.r.t. ground truth annotations on a subject-verb agreement task, as well as various qualitative inspections, using BERT, GPT-2 and LLaMA-3 as a testbed. Along the way we compare and extend the ALTI-Logit and LRP methods, including the recently proposed AttnLRP variant, from an algorithmic and implementation perspective. We further incorporate in our benchmark two widely-used gradient-based attribution techniques. Finally, we make our carefullly constructed benchmark dataset for evaluating attributions on language models, as well as our code, publicly available in order to foster evaluation of XAI-methods on a well-defined common ground.

Problem

• Analyze ALTI-Logit and LRP attribution methods
• Evaluate XAI methods on subject-verb agreement tasks
• Compare and extend ALTI-Logit, LRP, and AttnLRP techniques

Innovation

• Decomposition-based XAI methods
• Comparative analysis ALTI-Logit LRP
• Public benchmark dataset code