ResearchBenchmarksErasure Dependence Evaluation on Aggregate (AIME24, AIME25, ZebraLogic, AutoLogi, GPQA-D)Erasure Dependence Evaluation on Aggregate (AIME24, AIME25, ZebraLogic, AutoLogi, GPQA-D)Follow-1.6ASGCERL full-1.68-1.64-1.6-1.56May 22, 2026Evaluation resultsMethodMethodLinksASGESRMSGCERL fullTraining=SFT+RLTraining=SFT+RL2026.05-1.699.88.8