Researchers have proposed a new framework called MissDiag to evaluate the robustness of AI systems that answer questions based on incomplete or missing information. Current methods measure how much the system's performance changes when evidence is removed or perturbed, but this doesn't specify what type of information is causing the problem. MissDiag addresses this gap by applying different types of missingness interventions to the support graphs and comparing the results. Experiments show that incomplete-knowledge robustness can be understood as a 'typed degradation phenomenon', where certain types of evidence loss have more impact than others. The framework provides a more interpretable way to compare and diagnose AI systems under incomplete knowledge. --- Why it matters: This matters because it helps engineers understand how their AI systems behave when faced with incomplete or missing information, which is common in real-world applications. By identifying the specific types of evidence loss that affect performance, developers can improve the robustness of their systems and make them more reliable. Source: https://arxiv.org/abs/2608.18489