ResearchBenchmarksMathematical Reasoning on GSM8K (test) (Final Eval Reward, Training Time)Follow0.797Final RewardRecompute0.790760.792380.7940.79562Dec 6, 2025Evaluation ResultsMethodMethodLinksFinal RewardTraining Time (hours)RecomputeBackbone=Qwen2.5-1.5B-...Backbone=Qwen2.5-1.5B-Instruct2025.120.7971.82SyncBackbone=Qwen2.5-1.5B-...Backbone=Qwen2.5-1.5B-Instruct2025.120.7932.36LoglinearBackbone=Qwen2.5-1.5B-...Backbone=Qwen2.5-1.5B-Instruct2025.120.7911.53