MIST cuts misleading-context flips to 16.3%

MIST tests selective trust with 4 matched context conditions per reasoning item—clean, misleading, correct, and irrelevant—rather than rewarding models that simply ignore outside evidence. The benchmark paper reports that misleading context could flip every tested model from a correct answer to a wrong one; its SCOPE training method reduced those flips while preserving accuracy under the three control conditions. Builders should pair adversarial context tests with benign controls, extending production-grade agent evaluation beyond a single pass/fail score.