VISTA: Architecture-Agnostic Token Attribution Visualization Without Backpropagation at Half the GPU Cost

arXiv·medium signal

VISTA provides token attribution analysis across any transformer architecture without requiring backpropagation — cutting GPU memory usage nearly in half compared to gradient-based methods. Unlike architecture-specific interpretability tools, VISTA works on any model in the Transformer family. Practical for teams who need to debug or explain model behavior in production without doubling inference cost.