ResearchBenchmarksCode Generation and Reasoning on Code Domain AverageFollow73.7AccuracyDASD69.74870.77471.872.826May 21, 2026Evaluation ResultsMethodMethodLinksAccuracyDASDBackbone=Qwen3-8BBackbone=Qwen3-8B2026.0573.7GRPOBackbone=Qwen3-8BBackbone=Qwen3-8B2026.0572.3BaseBackbone=Qwen3-8BBackbone=Qwen3-8B2026.0572.1OPSDBackbone=Qwen3-8BBackbone=Qwen3-8B2026.0569.9