I built a benchmark to score the 'architectural intelligence' of neural nets (github.com)

A new open-source Python package, gwo-benchmark, implements the "Generalized Windowed Operation" (GWO) theory to score a neural operation's "architectural intelligence" — i.e., how smart an operation is relative to its size. Rather than only reporting accuracy, the framework computes an Operational Complexity (Ω_proxy) that combines Descriptive Complexity (C_D, how many basic primitives describe the operation) and Parametric Complexity (C_P, extra parameters/auxiliary nets) via Ω_proxy = C_D + α·C_P. Users inherit from GWOModule, declare C_D, list parametric modules, and run standardized evaluations (Evaluator) on datasets like CIFAR-10 to get complexity, latency and performance reports; the repo is pip-installable and includes examples and a live leaderboard. Technically, GWO reduces any op to three components—Path (where to look), Shape (what to look at), and Weight (how to value it)—each mapped to a small primitive vocabulary (e.g., STATIC_SLIDING, DENSE_SQUARE, SHARED_KERNEL, CONTENT_AWARE, DYNAMIC_ATTENTION). C_D is the sum of primitive scores (LLM prompts are provided to help map code to primitives); C_P is computed from listed submodules. The benchmark emits a composite score and tier (S/A/B/C/D) to compare operations (StandardConv ~990, DeformableConv ~771, DepthwiseConv ~681) and aims to push Pareto-efficient operation design, guide research on compact alternatives to large models, and standardize architectural-efficiency evaluation.