Benchmarks and Generalization

The survey does not introduce a new benchmark, run new experiments, or provide new statistical tests. Evidence is constrained by proprietary netlists, undisclosed commercial tool internals, and uneven public benchmarks. Generalization across PDKs, libraries, tools, and design types is a central deployment question. Future benchmarks should report confidence intervals, paired non-parametric tests, and rare-event metrics instead of only MAE or R2. The paper does not find public evidence for a universal graph foundation model across major circuit modalities.