When LRP Diverges from Leave-One-Out in Transformers
This work addresses the reliability of efficient feature attribution methods for researchers and practitioners using Transformers, but it is incremental as it builds on existing LRP diagnostics.
The paper tackled the problem of Layer-Wise Relevance Propagation (LRP) not aligning well with Leave-One-Out (LOO) feature importance in Transformers, finding that bilinear propagation rules violate axiomatic soundness and bypassing softmax layers improves alignment by up to 30% in later layers.
Leave-One-Out (LOO) provides an intuitive measure of feature importance but is computationally prohibitive. While Layer-Wise Relevance Propagation (LRP) offers a potentially efficient alternative, its axiomatic soundness in modern Transformers remains largely under-examined. In this work, we first show that the bilinear propagation rules used in recent advances of AttnLRP violate the implementation invariance axiom. We prove this analytically and confirm it empirically in linear attention layers. Second, we also revisit CP-LRP as a diagnostic baseline and find that bypassing relevance propagation through the softmax layer -- backpropagating relevance only through the value matrices -- significantly improves alignment with LOO, particularly in middle-to-late Transformer layers. Overall, our results suggest that (i) bilinear factorization sensitivity and (ii) softmax propagation error potentially jointly undermine LRP's ability to approximate LOO in Transformers.