Neural-MCTS Test Prioritization for Smart Contract Mutation Testing

Abstract

The immutability of deployed smart contracts increases the impact of undetected faults. Mutation testing evaluates test suite effectiveness by injecting controlled faults (mutants) into contract code and observing whether existing tests detect them. However, the number of generated mutants and the required test executions make mutation testing costly, limiting its scalability in realistic settings. We present ASCENT , an online test prioritization technique based on a Neural Monte Carlo Tree Search (Neural-MCTS) algorithm that dynamically learns and adapts test execution strategies during mutation analysis. The approach prioritizes tests in real time without requiring prior knowledge of the system under test, reducing execution cost while preserving fault detection effectiveness. This paper provides a comprehensive evaluation across five real-world Solidity projects. We analyze behavior under a wide range of hyperparameter configurations, examine how different state representations influence prioritization performance, and evaluate an asynchronous execution model that decouples search, inference, and training to reduce wall-clock execution time. The results show consistent reductions across all projects, ranging from 28–61% depending on project features and hyperparameter configurations, alongside a parallelized variant that substantially reduces wall-clock execution time with controlled trade-offs in prioritization performance.