TRACE: Learning to Compute on Circuit Graphs

A new machine-learning framework called TRACE has been introduced to model the functional behavior of circuit graphs, addressing what its authors describe as a fundamental architectural mismatch in current graph-learning methods [1]. The paradigm, detailed in a paper by Ziyang Zheng and colleagues, departs from mainstream message passing neural networks (MPNNs) and conventional Transformer-based models, which the authors argue rely on a flawed assumption that prevents them from capturing the position-aware, hierarchical nature of computation [1]. Circuit graphs, often represented as directed acyclic graphs (DAGs) where edges flow in one direction without forming closed loops, are foundational in computer science and scheduling applications [3]. TRACE employs a Hierarchical Transformer designed to mirror the step-by-step flow of computation, replacing the permutation-invariant aggregation used in prior architectures [1]. The model is trained under a novel objective called function shift learning. Rather than predicting a complex global function directly, TRACE learns to predict the discrepancy between the true global function and a simple local approximation that assumes input independence [1]. The work was validated across multiple circuit modalities, including Register Transfer Level graphs, And-Inverter Graphs, and post-mapping netlists. On a comprehensive suite of benchmarks, TRACE substantially outperformed all prior architectures, according to the paper [1]. The research was submitted to arXiv on September 26, 2025, and revised through June 2026 [1]. Graph representation learning has broad applications in artificial intelligence, a field encompassing tasks such as reasoning, problem-solving, and decision-making [9]. The Laplacian matrix, a core tool in spectral graph theory, is often used to construct low-dimensional embeddings for machine learning applications by relating graph properties to eigenvalues and eigenvectors [4]. The TRACE framework represents a shift toward architecturally aligned models for circuit computation, a domain where the structure of the graph directly encodes functional relationships.