RL Alignment of Reasoning

Tencent's PatternEval clocks a 48.64-point failure gap when hybrid models skip thinking

Accuracy scores stay flat while response behaviour falls apart. The fix Tencent proposes costs under a point of accuracy, but it lives at training time, where the operator routing for cost has no access.