Constraint Geometry Benchmarks
A benchmark framework for evaluating mechanism attribution under partial observability and interacting constraints.
Constraint Geometry Benchmarks evaluate whether AI systems can distinguish among competing governing structures that produce overlapping or identical outcomes. Unlike traditional benchmarks that measure state recognition or outcome prediction, this framework assesses mechanism attribution—identifying the structural configuration most consistent with observed trajectories. As the paper states, the benchmark focuses on “interacting constraints, compensation, partial observability, signal degradation, dominance relationships, intervention structure, and mechanism attribution.”
The Constraint Geometry Benchmark Suite (CGBS) introduces synthetic oncology‑inspired environments designed to reduce shortcut learning and label leakage. Models are evaluated not only on outcome prediction but also on Structural Alignment Accuracy, which measures whether the governing mechanism is correctly identified. The paper’s core contribution is a benchmark‑generation methodology for structural reasoning under controlled conditions, distinguishing mechanism attribution from both outcome prediction and full causal discovery.
Download paper as pdf
Download paper as pdf