ResearchBenchmarksReasoning Question Answering on ARC ChallengeFollow74.55AccuracyLoPT-GRPO74.040474.172774.30574.4373May 6, 2026Evaluation ResultsMethodMethodLinksAccuracyLoPT-GRPOBackbone=Qwen2.5-32BBackbone=Qwen2.5-32B2026.0574.55E2E-GRPOBackbone=Qwen2.5-32BBackbone=Qwen2.5-32B2026.0574.23BaseBackbone=Qwen2.5-32BBackbone=Qwen2.5-32B2026.0574.06