TRACE: Learning to Compute on Circuit Graphs

TRACE introduces a new paradigm for learning to compute on circuit graphs, utilizing a Hierarchical Transformer architecture that aligns with the step-by-step flow of computation, addressing limitations in traditional message passing neural networks and Transformer-based models. It features a novel function shift learning objective that focuses on predicting discrepancies between true global functions and local approximations, rather than the global function directly. TRACE demonstrates superior performance across various circuit modalities, including Register Transfer Level graphs and And-Inverter Graphs, highlighting its potential for advancing graph representation learning in practical applications.