ResearchBenchmarksMathematical Reasoning on MGSM Japanese (Pass@1)Follow89.6Pass@1 AccuracyBaseline85.1287.3689.691.84Jun 5, 2026Evaluation ResultsMethodMethodLinksPass@1 AccuracyAverage Output TokensBaselineModel=Qwen3-4B-Thinkin...Model=Qwen3-4B-Thinking-25072026.0689.62,482DyConModel=Qwen3-4B-Thinkin...Model=Qwen3-4B-Thinking-25072026.0689.61,823