AttnLRP: Attention-Aware Layer-Wise Relevance Propagation for Transformers
This addresses the problem of biased predictions and hallucinations in Large Language Models for researchers and practitioners needing interpretability, though it is incremental as it builds on existing attribution methods.
The paper tackles the challenge of achieving faithful and computationally efficient attributions for transformer models by extending Layer-wise Relevance Propagation to handle attention layers, demonstrating that their method surpasses existing approaches in faithfulness and enables understanding of latent representations across models like LLaMa 2 and vision transformers.
Large Language Models are prone to biased predictions and hallucinations, underlining the paramount importance of understanding their model-internal reasoning process. However, achieving faithful attributions for the entirety of a black-box transformer model and maintaining computational efficiency is an unsolved challenge. By extending the Layer-wise Relevance Propagation attribution method to handle attention layers, we address these challenges effectively. While partial solutions exist, our method is the first to faithfully and holistically attribute not only input but also latent representations of transformer models with the computational efficiency similar to a single backward pass. Through extensive evaluations against existing methods on LLaMa 2, Mixtral 8x7b, Flan-T5 and vision transformer architectures, we demonstrate that our proposed approach surpasses alternative methods in terms of faithfulness and enables the understanding of latent representations, opening up the door for concept-based explanations. We provide an LRP library at https://github.com/rachtibat/LRP-eXplains-Transformers.