ResearchBenchmarksMathematical Reasoning on GSM8K Multi-stepMathematical Reasoning on GSM8K Multi-stepFollow54CR (%)Entropy Gate51.352.655455.35Jun 2, 2026Evaluation resultsMethodMethodLinksCR (%)SETask PreservedEntropy GateModel=Claude Opus 4.8,...Model=Claude Opus 4.8, Tokens=702026.06540.904—