Mure framework accelerates DNN mutation testing by 44% with zero accuracy loss

Researchers at the University of [affiliation inferred] led by Ali Ghanbari have released Mure, a novel framework that speeds up deep neural network (DNN) mutation testing without any trade-off in accuracy. Mutation testing is a rigorous method for evaluating test dataset quality by introducing small changes (mutants) into a model and checking if tests detect them. However, it is computationally expensive because each mutant requires a full forward pass. Mure exploits the observation that mutants and the original model share large portions of computation. By memoizing (caching) the results of the common prefix and only executing the mutated suffix for each mutant, Mure achieves provably identical outputs to exhaustive mutation testing. The paper, accepted at ISSTA 2026, provides a formal proof of soundness and identifies conditions under which speed-up is guaranteed.

Empirical evaluation on 15 diverse DNN models—ranging from small networks with a few thousand parameters to large ones with millions—shows Mure yields an average 44.54% reduction in computational cost. For comparison, state-of-the-art lossy techniques can achieve up to 88.97% acceleration, but at the cost of measurable errors in mutation score. Mure’s lossless approach means developers can trust the results completely. Even under aggressive mutation rates (5% of neurons mutated), Mure maintains over 20% speed-up. The framework’s effectiveness depends on the mutation generation selection ratio—as the fraction of mutated neurons increases, memoization opportunities shrink, but the reduction is predictable. This makes Mure immediately useful for CI/CD pipelines where accurate test adequacy metrics are critical.