Reliability-Oriented Evaluation of Explanation Signals under Real and Synthetic Stress Tests

This paper presents a reliability-oriented empirical study of global importance signals in machine learning models, focusing on stability, sensitivity, and cross-model consistency rather than proposing a new explanation algorithm. (i) Across two public real-world datasets and controlled synthetic stress tests, we show that explanation reliability can degrade even when predictive performance remains strong (often high); (ii) we provide empirical evidence that stability, sensitivity, and cross-model agreement capture complementary failure modes overlooked by accuracy-centric evaluation; and (iii) we discuss implications for reliability-aware XAI benchmarking under spurious correlations and retraining variability. These findings treat explanation reliability as a primary empirical property in our evaluated setting and offer practical guidance under the considered regimes for building and using future XAI benchmarks.