Model checking has achieved remarkable success in the hardware domain, largely due to the accumulation of intricate optimizations and finely tuned implementation details. As tools evolve, diagnosing performance differences to better understand the interplay of these factors has become increasingly important. Yet existing problems that reveal such differences are often too large for meaningful inspection, limiting their diagnostic value.

To address the problem, this experience paper proposes AIGROW, a framework for generating hardware model checking problems, and introduces our experience on diagnosing performance differences in model checkers with the generated problems. AIGROW uses a feedback-guided process that evolves problems based on runtime information, selectively retaining those that become more difficult for a target checker. Performance differences are then revealed by evaluating these problems across hardware model checkers that have similar algorithms.

Our evaluation demonstrates that AIGROW generates problems that are more than 100 times smaller than those produced by existing generators, while still revealing substantial performance differences. Diagnosing the performance differences has led to concrete improvements in CAR-based checkers: (1) uncovering structural inefficiencies in their exploration strategies, (2) solving 18 previously unsolvable HWMCC’24 problems, and (3) reducing runtime from hours to minutes in several cases.