MIST reframes a common robustness problem for language models. A model that ignores misleading context may look safer, but if it ignores useful context too, it fails at the reason users provided documents or signals in the first place.
The benchmark presents reasoning items under four matched conditions: clean, misleading, correct-context, and irrelevant-context. It also introduces SC2W, a paired metric that counts how often a misleading signal turns a clean-correct answer into a wrong one.
The paper’s useful distinction is selective trust. Future systems need to learn when outside context deserves weight, not simply become more stubborn or more obedient. That is especially relevant for retrieval systems and agents that mix model knowledge with live evidence.